When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. As I respect each of these Grandmasters individually, I'm disappointed in this kind of anti-competitive behavior. One could counter-argue that all you have to do is come up with a better solution, but access to powerful hardware is a clear edge in certain competitions where cycles are important for a winning solution.
", "messages": [ { "id": 1309688, "postDate": "2021-05-16T07:58:13.043Z", "content": "You're probably 4 places higher because of that.
", "rawMarkdown": "You're probably 4 places higher because of that.", "votes": 19, "replies": [ { "id": 1311197, "postDate": "2021-05-17T08:17:12.433Z", "content": "You're on a roll mate
", "rawMarkdown": "You're on a roll mate", "votes": 1 }, { "id": 1311310, "postDate": "2021-05-17T10:07:38.737Z", "content": "Hi there! :D
", "rawMarkdown": "Hi there! :D" }, { "id": 1316346, "postDate": "2021-05-20T13:30:50.847Z", "content": "While I admire @mconway's concern for the small-time players, I absolutely agree with @nofreewill's observation that a merger at the top moves me a place or two up the LB. If five GMs are playing as individuals, maybe the best I can hope for is sixth. I they merge into a superteam, I can target second place. In fact, mergers push the pursuit of Kaggle medals further away from being a zero sum game.
", "rawMarkdown": "While I admire @mconway's concern for the small-time players, I absolutely agree with @nofreewill's observation that a merger at the top moves me a place or two up the LB. If five GMs are playing as individuals, maybe the best I can hope for is sixth. I they merge into a superteam, I can target second place. In fact, mergers push the pursuit of Kaggle medals further away from being a zero sum game. ", "votes": 1 }, { "id": 1316385, "postDate": "2021-05-20T13:57:32.553Z", "content": "That's assuming there's no synergy in teaming up. I think the way people are describing GMs here (natural talent a la Kipchoge) is a bit silly especially when you look at many GMs' solo track record or lack thereof.
\nKaggle is a platform with a bunch of arbitrary rules surrounding competitions - a lot of which are geared towards gamifying data science, not necessarily identifying the best data scientists in the world.
", "rawMarkdown": "That's assuming there's no synergy in teaming up. I think the way people are describing GMs here (natural talent a la Kipchoge) is a bit silly especially when you look at many GMs' solo track record or lack thereof. \n\nKaggle is a platform with a bunch of arbitrary rules surrounding competitions - a lot of which are geared towards gamifying data science, not necessarily identifying the best data scientists in the world. \n", "votes": 3 }, { "id": 1316472, "postDate": "2021-05-20T15:30:25.610Z", "content": "Bronze is definitely easier to get because of mergers, silver most probably is, too. But they make the very top places harder and harder to reach, because merging definitely makes scores higher compared to the members' individual scores.
", "rawMarkdown": "Bronze is definitely easier to get because of mergers, silver most probably is, too. But they make the very top places harder and harder to reach, because merging definitely makes scores higher compared to the members' individual scores.", "votes": 1 }, { "id": 1316491, "postDate": "2021-05-20T15:36:47.663Z", "content": "\n\nespecially when you look at many GMs' solo track record or lack thereof.
\n
@leecming I agree on the assessment but maybe not on the cause. Requirement for competition GM is one solo gold. There is no upside in getting more solo gold medals. Why would GMs take risks in pursuing solo gold medals?
\nI personally think that competition GM should have at least 2 gold medals. It would filter the lucky winners. But Kaggle is not designed to evaluate individuals as you wrote. It is designed to provide value to competition sponsors. Encouraging strong teaming makes more sense for that.
", "rawMarkdown": "> especially when you look at many GMs' solo track record or lack thereof. \n\n@leecming I agree on the assessment but maybe not on the cause. Requirement for competition GM is one solo gold. There is no upside in getting more solo gold medals. Why would GMs take risks in pursuing solo gold medals?\n\nI personally think that competition GM should have at least 2 gold medals. It would filter the lucky winners. But Kaggle is not designed to evaluate individuals as you wrote. It is designed to provide value to competition sponsors. Encouraging strong teaming makes more sense for that.\n\n\n", "votes": 6 }, { "id": 1316821, "postDate": "2021-05-20T23:04:47.120Z", "content": "Kaggle should introduce another tier above GM with even more requirements.
\nWhat would it be called? Start using royal ranks - I wouldn't mind being a Duke or Baron of Kaggle ;)
More seriously, Kaggle could consider a martial arts or Go (boardgame) style \"dan ranking\" which has numerical groupings. Maybe for each solo gold you get, you go up one \"dan\".
", "rawMarkdown": "Kaggle should introduce another tier above GM with even more requirements.\nWhat would it be called? Start using royal ranks - I wouldn't mind being a Duke or Baron of Kaggle ;)\n\nMore seriously, Kaggle could consider a martial arts or Go (boardgame) style \"dan ranking\" which has numerical groupings. Maybe for each solo gold you get, you go up one \"dan\". \n\n", "votes": 1 }, { "id": 1316832, "postDate": "2021-05-20T23:26:40.427Z", "content": "\n\nKaggle should introduce another tier above GM
\n
There is the GGGGMaster ;)
\nThe points which decay are for the fine-grained ranking, including the fact that you see your standing among all Kagglers.
", "rawMarkdown": "> Kaggle should introduce another tier above GM\n\nThere is the [GGGGMaster](https://web.archive.org/web/20190505233019/http://blog.kaggle.com/2019/04/01/ggggmaster) ;)\n\nThe points which decay are for the fine-grained ranking, including the fact that you see your standing among all Kagglers.", "votes": 1 } ] }, { "id": 1309637, "postDate": "2021-05-16T06:58:06.650Z", "content": "Kaggle was never a fair game and won't be. It is a crowdsourcing platform for getting the best solution for its customers. Unfortunately it is not chess.
", "rawMarkdown": "Kaggle was never a fair game and won't be. It is a crowdsourcing platform for getting the best solution for its customers. Unfortunately it is not chess.", "votes": 13, "replies": [ { "id": 1309660, "postDate": "2021-05-16T07:33:10.403Z", "content": "Is it though? Limiting team sizes to a max of five indicates that there are other considerations than just sourcing for the \"best solution\".
\nI've always been a skeptic of the argument that Kaggle is mainly about delivering models to sponsors. It seems more a pedagogical platform with sponsors getting PR or \"building awareness\" more than anything else.
", "rawMarkdown": "Is it though? Limiting team sizes to a max of five indicates that there are other considerations than just sourcing for the \"best solution\". \n\nI've always been a skeptic of the argument that Kaggle is mainly about delivering models to sponsors. It seems more a pedagogical platform with sponsors getting PR or \"building awareness\" more than anything else.", "votes": 7 }, { "id": 1309670, "postDate": "2021-05-16T07:45:14.517Z", "content": "You are also right. My point is that there is some gamification but it is not a pure skill game. Hardware, luck, teaming strategies, using leaks are as important as skills. You can make it to Kaggle Top 10 with limited Machine Learning skills. The same is impossible for chess.
", "rawMarkdown": "You are also right. My point is that there is some gamification but it is not a pure skill game. Hardware, luck, teaming strategies, using leaks are as important as skills. You can make it to Kaggle Top 10 with limited Machine Learning skills. The same is impossible for chess.", "votes": 7 }, { "id": 1310536, "postDate": "2021-05-16T18:09:35.390Z", "content": "In chess, it is generally impossible for a person to get to the TOP … alpha zero will still win. And here it is possible for everyone (theoretically)
", "rawMarkdown": "In chess, it is generally impossible for a person to get to the TOP ... alpha zero will still win. And here it is possible for everyone (theoretically)", "votes": 1 }, { "id": 1310656, "postDate": "2021-05-16T19:44:37.637Z", "content": "\n\nIn chess, it is generally impossible for a person to get to the TOP … alpha zero will still win. And here it is possible for everyone (theoretically)
\n
DeepMind: \"hold my beer\"
\nBut yeah, true AutoML seems to be a bit far away, but who knows.
", "rawMarkdown": ">In chess, it is generally impossible for a person to get to the TOP … alpha zero will still win. And here it is possible for everyone (theoretically)\n\nDeepMind: \"hold my beer\"\n\nBut yeah, true AutoML seems to be a bit far away, but who knows.", "votes": 1 } ] }, { "id": 1310758, "postDate": "2021-05-16T23:25:50.090Z", "content": "Yes your are right, it is the case here, partly, but that is not the end of the story.
\nI also had this thought in my Kaggle beginning, those with more/powerful HW had an advantage but with now 50+ competitions, I believe it’s more a solution game before HW game. Yes one can train the same idea and model quicker, but in the end, the idea itself behind is the key. And if one want advantage in calculation power, one can start the competition early, use free Kaggle GPU/TPU, Colab GPU/TPU, it will give many hours.
\nAnd with many techniques one can squeeze more out of the code, like using Mixed Precision with Volta devices, using GPU instead of CPU in the whole code, all making it possible to train bigger model and faster training. And in HW power terms the Kaggle TPU version have power.
\nFor example the Google Landmark Retrieval 2020 competition with 5M images in the training set, that competition was also a HW competition but by analyzing the winners write-up, the person only used Colab/Kaggle TPU, ~400 hours, what a dedicated work, and it was also a brilliant solution, respect.
\nBy reading the top solutions writeup from every competition, one can see that these smart people often came up with a new level of SOTA AI / Data Science solution that help both solving the problem to the host and also takes the AI area to the next level. When I read them, I don’t think of what GPU/TPU version they used, I more see the bright idea and innovative code, and how they solved the labyrinth step by step.
\nIn terms of HW power advantages I think cloud Auto AI solution that bring a solution in minutes through benchmarking different models and tuning parameters can really help, but those are often licensed and not often seen used here, can believe that a quantum computer under the hood to a full current-market-solutions-loaded Auto AI would give some hint where to start 😉.
In Kaggle competition the HW really give an edge, oh yes, but the Data Science input and digging the path in new/unknow territory is crucial to grab the first place.
", "rawMarkdown": "Yes your are right, it is the case here, partly, but that is not the end of the story.\n\nI also had this thought in my Kaggle beginning, those with more/powerful HW had an advantage but with now 50+ competitions, I believe it’s more a solution game before HW game. Yes one can train the same idea and model quicker, but in the end, the idea itself behind is the key. And if one want advantage in calculation power, one can start the competition early, use free Kaggle GPU/TPU, Colab GPU/TPU, it will give many hours.\n\nAnd with many techniques one can squeeze more out of the code, like using Mixed Precision with Volta devices, using GPU instead of CPU in the whole code, all making it possible to train bigger model and faster training. And in HW power terms the Kaggle TPU version have power.\n\nFor example the Google Landmark Retrieval 2020 competition with 5M images in the training set, that competition was also a HW competition but by analyzing the winners write-up, the person only used Colab/Kaggle TPU, ~400 hours, what a dedicated work, and it was also a brilliant solution, respect.\n\nBy reading the top solutions writeup from every competition, one can see that these smart people often came up with a new level of SOTA AI / Data Science solution that help both solving the problem to the host and also takes the AI area to the next level. When I read them, I don’t think of what GPU/TPU version they used, I more see the bright idea and innovative code, and how they solved the labyrinth step by step. \nIn terms of HW power advantages I think cloud Auto AI solution that bring a solution in minutes through benchmarking different models and tuning parameters can really help, but those are often licensed and not often seen used here, can believe that a quantum computer under the hood to a full current-market-solutions-loaded Auto AI would give some hint where to start 😉. \n\nIn Kaggle competition the HW really give an edge, oh yes, but the Data Science input and digging the path in new/unknow territory is crucial to grab the first place.", "votes": 11 }, { "id": 1309680, "postDate": "2021-05-16T07:52:42.413Z", "content": "Access to powerful hardware is indeed an advantage in this competition as it seems. But I do not believe this has necessarily anything to do with the number of grandmasters in a team. There are some rather \"unknown\" kagglers with much more hardware than you can imagine, and then there are some top kagglers with much less hardware than you would think.
\nIn this competition it was clear from the start that it will be at least partly a HW game. There have been quite a few competitions lately, and I wonder why you decided to join this competition and not others if that bothers you? I don't think this will be ever a fully fair game, unless Kaggle decides to do kernel-only fitting competitions again, which doesn't seem to be their plan.
", "rawMarkdown": "Access to powerful hardware is indeed an advantage in this competition as it seems. But I do not believe this has necessarily anything to do with the number of grandmasters in a team. There are some rather \"unknown\" kagglers with much more hardware than you can imagine, and then there are some top kagglers with much less hardware than you would think. \n\nIn this competition it was clear from the start that it will be at least partly a HW game. There have been quite a few competitions lately, and I wonder why you decided to join this competition and not others if that bothers you? I don't think this will be ever a fully fair game, unless Kaggle decides to do kernel-only fitting competitions again, which doesn't seem to be their plan.", "votes": 12, "replies": [ { "id": 1309707, "postDate": "2021-05-16T08:13:33.190Z", "content": "Also, not all competitions are hardware hungry and rather need clever thinking and clever solutions. You may find those more enjoyable to try.
\nLike in all sports or competitions: Equipment (Hardware) can give you an advantage, but it's often times less than you would expect. Kipchoge would still outrun most people with only sandals. Tennis pros would still dominante amateurs with a beginner racket. It usually only gives you that last percentage that makes the difference between the very very best.
", "rawMarkdown": "Also, not all competitions are hardware hungry and rather need clever thinking and clever solutions. You may find those more enjoyable to try.\n\nLike in all sports or competitions: Equipment (Hardware) can give you an advantage, but it's often times less than you would expect. Kipchoge would still outrun most people with only sandals. Tennis pros would still dominante amateurs with a beginner racket. It usually only gives you that last percentage that makes the difference between the very very best.", "votes": 7 } ] }, { "id": 1316559, "postDate": "2021-05-20T16:41:42.823Z", "content": "I think Kaggle should promote more solo competitions, beside teams ones. If I'm not mistaken, the last one is Instacart held 4 years ago. And that could reduce most of the gamifying part.
", "rawMarkdown": "I think Kaggle should promote more solo competitions, beside teams ones. If I'm not mistaken, the last one is [Instacart](https://www.kaggle.com/c/instacart-market-basket-analysis/overview) held 4 years ago. And that could reduce most of the gamifying part. ", "votes": 9, "replies": [ { "id": 1316677, "postDate": "2021-05-20T19:16:21.237Z", "content": "Well, considering that competition hosts have priority in the end, what could happen, is that Kaggle will be looking (more) at studies of team dynamics.
\nSo in the end we could - this is a hypothetical example - have e.g. something like max teams of two where creativity is far more important, as pairs are the most efficient teams.
\nSo the current state might be, ridiculous as it seems, fairer while still satisfying competition hosts.
Interesting other avenue of investigation: What is the correlation between teams and private leaderboard placement? One could answer that with available data. As almost anyone can team up with others, that could help to improve expected outcome for those who value it more than whatever else might speak against being in a team.
", "rawMarkdown": "Well, considering that competition hosts have priority in the end, what could happen, is that Kaggle will be looking (more) at studies of team dynamics.\nSo in the end we could - this is a hypothetical example - have e.g. something like max teams of two where creativity is far more important, as pairs are the most efficient teams. \nSo the current state might be, ridiculous as it seems, fairer while still satisfying competition hosts.\n\nInteresting other avenue of investigation: What is the correlation between teams and private leaderboard placement? One could answer that with available data. As almost anyone can team up with others, that could help to improve expected outcome for those who value it more than whatever else might speak against being in a team." }, { "id": 1321074, "postDate": "2021-05-24T14:01:25.890Z", "content": "Kaggle Marathons. Where Kaggle organises, no hosts. They are medal based competitions hosted by Kaggle for respective tiers. I have proposed how it works in a comment below.
", "rawMarkdown": "Kaggle Marathons. Where Kaggle organises, no hosts. They are medal based competitions hosted by Kaggle for respective tiers. I have proposed how it works in a comment below." } ] }, { "id": 1311240, "postDate": "2021-05-17T08:59:42.427Z", "content": "As my father always said to me when I was a child, \"life is not fair\".
", "rawMarkdown": "As my father always said to me when I was a child, *\"life is not fair\"*. ", "votes": 9 }, { "id": 1309400, "postDate": "2021-05-15T22:50:33.957Z", "content": "When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. As I respect each of these Grandmasters individually, I'm disappointed in this kind of anti-competitive behavior. One could counter-argue that all you have to do is come up with a better solution, but access to powerful hardware is a clear edge in certain competitions where cycles are important for a winning solution.
", "rawMarkdown": "When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. As I respect each of these Grandmasters individually, I'm disappointed in this kind of anti-competitive behavior. One could counter-argue that all you have to do is come up with a better solution, but access to powerful hardware is a clear edge in certain competitions where cycles are important for a winning solution.", "votes": 12 }, { "id": 1313461, "postDate": "2021-05-18T15:15:30.610Z", "content": "The CommonLit Readability competition looks to be really good for those with low amounts of hardware. With the readability score composition being non-disclosed - as far as I can tell - you also can't just aggregate large amounts of extra data.
\nSo I suggest taking part in that competition, if you want one where ingenuity is far more important than computational power.
Of course you could argue that four grandmasters have a higher expected ingenuity. But I say, it is more their experience that might let them make less mistakes and compose ideas better and know more of them. Though you can still be very creative and read paper (abstracts) without that.
\nSo it is certainly fairer than most other scenarios you can hope for on Kaggle.
You raise an interesting point but I am afraid you jump to wrong conclusions. For instance:
\n\n\nWhen you have a team that collectively works for a company with access to unlimited hardware and resources
\n
They work for different companies. It so happen that I work in the same company as two of them. I wished that what you wrote was true, but it isn't, by any mean. Don't take their team name as face value, it is not the reality.
\nI know of another company that let his kagglers use up to 120 V100 GPU. We have access to a small fraction of that at our company.
\nI reacted just to straighten a bit the expectations here. Compute resource matters, no doubt. But it is not the end of the story.
\nAnd about having many GMs in a single team, it is not the first nor the last time it happens. A team of 3 GMS has won a series of competition recently for instance. No one found it unfair AFAIK.
", "rawMarkdown": "You raise an interesting point but I am afraid you jump to wrong conclusions. For instance:\n\n> When you have a team that collectively works for a company with access to unlimited hardware and resources\n\nThey work for different companies. It so happen that I work in the same company as two of them. I wished that what you wrote was true, but it isn't, by any mean. Don't take their team name as face value, it is not the reality.\n\nI know of another company that let his kagglers use up to 120 V100 GPU. We have access to a small fraction of that at our company. \n\nI reacted just to straighten a bit the expectations here. Compute resource matters, no doubt. But it is not the end of the story.\n\nAnd about having many GMs in a single team, it is not the first nor the last time it happens. A team of 3 GMS has won a series of competition recently for instance. No one found it unfair AFAIK.", "votes": 3, "replies": [ { "id": 1311838, "postDate": "2021-05-17T16:33:27.150Z", "content": "We can agree to disagree. Nvidia is using its Grandmasters (two in this case) to market expensive hardware unattainable to many, thus the team name. Maybe I just have a different viewpoint because I like to help the less fortunate. My friend @robertdscott overcame extreme adversity to become a Data Scientist, and I don't see Nvidia doing anything to level the playing field.
", "rawMarkdown": "We can agree to disagree. Nvidia is using its Grandmasters (two in this case) to market expensive hardware unattainable to many, thus the team name. Maybe I just have a different viewpoint because I like to help the less fortunate. My friend @robertdscott overcame extreme adversity to become a Data Scientist, and I don't see Nvidia doing anything to level the playing field.", "votes": 2 }, { "id": 1311883, "postDate": "2021-05-17T17:20:46.263Z", "content": "I am very flattered to be the only one you responded to. Thanks.
", "rawMarkdown": "I am very flattered to be the only one you responded to. Thanks.", "votes": 6 } ] }, { "id": 1318833, "postDate": "2021-05-22T15:46:43.387Z", "content": "The four GMs are no longer at the top though ^^
", "rawMarkdown": "The four GMs are no longer at the top though ^^", "votes": 3, "replies": [ { "id": 1319421, "postDate": "2021-05-23T07:06:09.947Z", "content": "Kudos to the (literally) green trio for at least temporarily beating a team of 4 GMs
", "rawMarkdown": "Kudos to the (literally) green trio for at least temporarily beating a team of 4 GMs ", "votes": 3 } ] }, { "id": 1317634, "postDate": "2021-05-21T15:03:39.693Z", "content": "Competition is not the right word to describe what we are doing on Kaggle. It should be called a challenge.
\nWe are all here learning together to solve a challenge for a host and to improve our skills. We are not here competing with each other like a sporting event with equal types of equipment.
So enjoy learning and professional networking on Kaggle and see it as a fun challenge, not like a chess-boxing competition
", "rawMarkdown": "Competition is not the right word to describe what we are doing on Kaggle. It should be called a challenge.\nWe are all here learning together to solve a challenge for a host and to improve our skills. We are not here competing with each other like a sporting event with equal types of equipment.\n\nSo enjoy learning and professional networking on Kaggle and see it as a fun challenge, not like a chess-boxing competition", "votes": 4, "replies": [ { "id": 1317933, "postDate": "2021-05-21T20:09:17.193Z", "content": "I agree, in a competition people would not help each other by sharing e.g. code and bugs etc.
", "rawMarkdown": "I agree, in a competition people would not help each other by sharing e.g. code and bugs etc.", "votes": 2 } ] }, { "id": 1309649, "postDate": "2021-05-16T07:09:24.037Z", "content": "I doubt the \"unlimited\".
\nTheretically, everyone can rent as many Cloud Computing Hardware as she or he wants to get an edge. It just depends on the willingness of the kaggler to pay for that.
\nSame here, if a company wants to provide ressources (they do have a cost) for an expected or even unknown return, why not?
In the end, it is business. A company or a group is paying for a solution. Best solution wins. If it needs large investments to get that best solution you need to calculate the risk and reward ratio.
", "rawMarkdown": "I doubt the \"unlimited\".\nTheretically, everyone can rent as many Cloud Computing Hardware as she or he wants to get an edge. It just depends on the willingness of the kaggler to pay for that. \nSame here, if a company wants to provide ressources (they do have a cost) for an expected or even unknown return, why not?\n\nIn the end, it is business. A company or a group is paying for a solution. Best solution wins. If it needs large investments to get that best solution you need to calculate the risk and reward ratio.", "votes": 4 }, { "id": 1310917, "postDate": "2021-05-17T04:06:49.987Z", "content": "Grandmaster title tells you:
\nAnd I believe they are equally important. The second team on the LB are all novice Kagglers.
\nKaggle was never fair, and will never be. The organizers spend a bunch of money to find a solution, not to organize a competition. First place prize is $25k, what hardware can you build with 25k?
", "rawMarkdown": "Grandmaster title tells you:\n- How much effort you have spent on the platform\n- How good you are\n\nAnd I believe they are equally important. The second team on the LB are all novice Kagglers.\n\nKaggle was never fair, and will never be. The organizers spend a bunch of money to find a solution, not to organize a competition. First place prize is $25k, what hardware can you build with 25k?", "votes": 3 }, { "id": 1318067, "postDate": "2021-05-22T01:00:55.010Z", "content": "Could more code competitions help to balance that?
", "rawMarkdown": "Could more code competitions help to balance that?", "votes": 1, "replies": [ { "id": 1321070, "postDate": "2021-05-24T14:00:17.230Z", "content": "Some code competitions allow \"Models\" as an external dataset. So, still, train on your rigs. Upload model here, make it public/private on the last day.
\nBeen there. Its pseudo fair
", "rawMarkdown": "Some code competitions allow \"Models\" as an external dataset. So, still, train on your rigs. Upload model here, make it public/private on the last day. \n\nBeen there. Its pseudo fair" }, { "id": 1321139, "postDate": "2021-05-24T14:15:51.313Z", "content": "I see.. thanks!
", "rawMarkdown": "I see.. thanks!", "votes": 1 } ] }, { "id": 1323310, "postDate": "2021-05-26T06:36:32.477Z", "content": "When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle.
But this is an assumption right? Maybe it is not true. Maybe there are experts and Masters out there whose companies are giving them even better GPU times. And even if this is true, this is perfectly legit. They worked hard, became GMs and maybe now have access to better resources.
\nThe other point about 4 GM on one team..No rule prohibits this in Kaggle. Whether it is right in spirit - I am not sure. Where do we draw the line - 1 GM per team or 2. Why not 3?
\nHaving said that, I have neither upvoted or downvoted this thread. It has an important point. Maybe Kaggle and Kagglers would have more diversity in solutions if they allowed only 1 GM per team? The sheer brilliance and knowledge/expertise each one carries can benefit the community as a whole when they are pitted against each other…So there is a point there….
", "rawMarkdown": "`When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. `\n\nBut this is an assumption right? Maybe it is not true. Maybe there are experts and Masters out there whose companies are giving them even better GPU times. And even if this is true, this is perfectly legit. They worked hard, became GMs and maybe now have access to better resources. \n\nThe other point about 4 GM on one team..No rule prohibits this in Kaggle. Whether it is right in spirit - I am not sure. Where do we draw the line - 1 GM per team or 2. Why not 3? \n\nHaving said that, I have neither upvoted or downvoted this thread. It has an important point. Maybe Kaggle and Kagglers would have more diversity in solutions if they allowed only 1 GM per team? The sheer brilliance and knowledge/expertise each one carries can benefit the community as a whole when they are pitted against each other...So there is a point there....", "votes": 2 }, { "id": 1316785, "postDate": "2021-05-20T21:56:16.423Z", "content": "Me: let's run an experiment.
\nOne month later…
\nGPU: I am really tired and I only finished 3 epochs.
Joined Kaggle as a student, have 1 gold medal. Now I am 2 yoe. Here's my story:
\nI joined Kaggle as a student. I had an 8 GB laptop when I started. I was in situations where I couldn't run EDA scripts forget competitions. I bagged silver and also team gold medals.
\nMy experiences were good and bad.
\nYes, I got that gold in Code Competitions and without them, I would not probably have gold (in team) today
\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford. And with taxes and customs, just forget them in India. What say, students? Do some Indian users relate?
\nCode Competitions level the playing field by a lot (that especially don't allow pre-run models). I have worked in a professional capacity where Kaggle was a part-time job and also where Kaggling wasn't allowed. It clearly makes it more feasible when you do Kaggle in your day job. Again, unfair but it's the game and how it works.
\n\n\nI believe having a paid job to do Kaggle is what leads one to better places + corporate GPUs boost chances further
\n
But, nonetheless. There are few competitions that are really about feature engineering and not much raw power. And I believe with excellent engineering you can beat the x100 raw performances. I have seen it in practice and earned medals as well.
\nHere are some takeaways I have from my Kaggling experiences;
\nIf you want to compete for earning medals;
\nMake competition spaces. Where Kagglers from their tier only to get compete in yearly seasons.
\nEvery top 20% get a chance to compete in competitions above +1 their tier. So top users from Masters winning competition can compete now with GMs. This way it's a huge motivation to learn, grow in rankings have a fair chance to grab golds.
\nAnd yes, the criterion to be Competitions GM of solo gold is bad. Its harder in a cutting edge CV competition today than some Weather prediction competition 6 years back? Apparently, early to Kaggle would have made you GM than today's competition? I wonder what current GMs think of this?
\nAlso, a single gold should be equated to x10 silvers. Or medal to medal trade-in.
\n\n\nIn this pandemic, I realised. Maybe it's more important to be in your home town, share more stories with family and have a better work-life balance than Kaggling all weekend.
\n
I do want to become GM one day. Maybe discussions/notebooks are feasible. But due to less access to hardware and lot of CV competitions, it's definitely on the harder side for me.
\n(Edited x2)
", "rawMarkdown": "Joined Kaggle as a student, have 1 gold medal. Now I am 2 yoe. Here's my story:\n\nI joined Kaggle as a student. I had an 8 GB laptop when I started. I was in situations where I couldn't run EDA scripts forget competitions. I bagged silver and also team gold medals.\n\nMy experiences were good and bad. \n\n**Yes, I got that gold in Code Competitions and without them, I would not probably have gold (in team) today**\n\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford. And with taxes and customs, just forget them in India. What say, students? Do some Indian users relate? \n\nCode Competitions level the playing field by a lot (that especially don't allow pre-run models). I have worked in a professional capacity where Kaggle was a part-time job and also where Kaggling wasn't allowed. It clearly makes it more feasible when you do Kaggle in your day job. Again, unfair but it's the game and how it works. \n\n> I believe having a paid job to do Kaggle is what leads one to better places + corporate GPUs boost chances further\n\nBut, nonetheless. There are few competitions that are really about feature engineering and not much raw power. And I believe with excellent engineering you can beat the x100 raw performances. I have seen it in practice and earned medals as well. \n\nHere are some takeaways I have from my Kaggling experiences;\n1. Kaggle is a learning ed-tech platform. May be unfair. Prioritize learning than earning medals\n2. Don't get too attached to earning medals. They are surely on your way when you have progressive learnings \n3. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success\n4. I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. \n\nIf you want to compete for earning medals;\n1. Target less data-based competitions and that are more creative\n2. Prefer Code Competitions over normal ones\n3. Engage on learnings and takeaways for your personal growth and career success\n\n#### Solution I propose\n\nMake competition spaces. Where Kagglers from their tier only to get compete in yearly seasons. \n\n- Experts vs. Experts\n- Masters vs. Masters\n\nEvery top 20% get a chance to compete in competitions above +1 their tier. So top users from Masters winning competition can compete now with GMs. This way it's a huge motivation to learn, grow in rankings have a fair chance to grab golds.\n\nAnd yes, the criterion to be Competitions GM of solo gold is bad. Its harder in a cutting edge CV competition today than some Weather prediction competition 6 years back? Apparently, early to Kaggle would have made you GM than today's competition? I wonder what current GMs think of this?\n\nAlso, a single gold should be equated to x10 silvers. Or medal to medal trade-in. \n\n> In this pandemic, I realised. Maybe it's more important to be in your home town, share more stories with family and have a better work-life balance than Kaggling all weekend.\n\nI do want to become GM one day. Maybe discussions/notebooks are feasible. But due to less access to hardware and lot of CV competitions, it's definitely on the harder side for me.\n\n(Edited x2)", "votes": -2, "replies": [ { "id": 1321225, "postDate": "2021-05-24T15:02:55.747Z", "content": "Got downvoted. Would be glad to have healthy conversations on why someone would disagree, and what exactly or all of it. Perhaps any constructive feedback would be great. Much appreciated.
", "rawMarkdown": "Got downvoted. Would be glad to have healthy conversations on why someone would disagree, and what exactly or all of it. Perhaps any constructive feedback would be great. Much appreciated.", "votes": 4 }, { "id": 1321346, "postDate": "2021-05-24T16:25:39.997Z", "content": "I didn't read your whole comment, just few parts of it.
\nBut what I've read, I can see where that downvote could come from. As you said, it would have been nice if the downvoter wrote his points.
You write that Yes, I got that gold in Code Competitions and without them, I would not probably have gold today
\nBut after seeing your profile, I guess it was not much about Code Competitions for you, but being in a group of 6. Maybe for the hole team it was more about Code Competitions, but it was more about being in a big team for the individuals that brought the gold. I guess!!
\nBut you've written about that gold as if it was a solo one.
Also you write that Medals are monitor pixels. You aren't getting gold for real. Buy dogecoin if you wanna be rich
\nDogecoins are just as virtual as kaggle medals. Also your money is virtual. Cash has only value because we know that others will accept it, too. But if not, it worth nothing. Kaggle medals are valuable because you can show that you accomplished something. So those pixels are a lot more valuable than some other pixels.
You write that Now I am a young professional with 2 yoe.
\nI have made a face tracking little toy-machine as my BSc thesis and made a better version of it as my MSc thesis. I've put a lot of work into them and learned a lot about deep learning along the way. So much that I got into the field and have a job for 1.5 years now. I think that I work pretty hard to become better and better, but I'm far from being a professional.
\nSo I find it pretty hard to believe that you've become a professional in just two years. Or I may just misunderstand the word \"professional\".
\nI would say that a professional is closer to a senior than a junior, but I may be wrong. Correct me, if I'm wrong.
Hey, thanks for the comment.
\nI wrote
\n\n\nI bagged silver and also team gold medals.
\n
to make it explicit the gold medal was from the team and not solo. Perhaps I can edit and make it more apparent.
\n\n\nMedals are monitor pixels. You aren't getting gold for ….
\n
I agree with this. May be was too sarcastic. I deeply respect the effort behind the medals. But, having too many feelings attached, the rush and the whole emotions to just medals ( IMO of users and their past behaviour)
\n\n\n\"professional\" meant on a full-time job / working professionally for me. No expert, no junior & no senior.
\n
Maybe I will refactor the comment now. Thanks!
\nEdit:
\nDone now. Thanks
", "rawMarkdown": "Hey, thanks for the comment. \n\nI wrote \n> I bagged silver and also team gold medals. \n\n\nto make it explicit the gold medal was from the team and not solo. Perhaps I can edit and make it more apparent. \n\n> Medals are monitor pixels. You aren't getting gold for .... \n\n\nI agree with this. May be was too sarcastic. I deeply respect the effort behind the medals. But, having too many feelings attached, the rush and the whole emotions to just medals ( IMO of users and their past behaviour)\n\n> \"professional\" meant on a full-time job / working professionally for me. No expert, no junior & no senior. \n\nMaybe I will refactor the comment now. Thanks!\n\nEdit:\n\nDone now. Thanks\n\n" }, { "id": 1321458, "postDate": "2021-05-24T17:36:29.813Z", "content": "\nyou said
\nDon't get too attached to earning medals.`\n\nMedals are monitor pixels. You aren't getting gold for real.\n\nI talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill.\njust imagine telling a football player in FIFA World Cup, when he is having the time of his life (who worked so hard to get to this stage in his life)
\n\" Hey, This is just a game, the world cup will not guarantee you a place in the best football clubs, and …….by the way, the Trophy is not real gold\" 😂
", "rawMarkdown": "@shaz13 \n\nyou said \n\n\n```\nDon't get too attached to earning medals.`\n\nMedals are monitor pixels. You aren't getting gold for real.\n\nI talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill.\n\n```\n\n\n\njust imagine telling a football player in FIFA World Cup, when he is having the time of his life (who worked so hard to get to this stage in his life) \n\n\" Hey, This is just a game, the world cup will not guarantee you a place in the best football clubs, and .......by the way, the Trophy is not real gold\" 😂", "votes": 1 }, { "id": 1321608, "postDate": "2021-05-24T19:54:48.683Z", "content": "I didn't say anything like that.
\nEdit: Thanks for the correction! :)
", "rawMarkdown": "I didn't say anything like that.\n\nEdit: Thanks for the correction! :)", "votes": 1 }, { "id": 1321641, "postDate": "2021-05-24T20:18:17.530Z", "content": "AFAIK, a player who is passionate about football and works hard would win the trophy and still cherish the hard work behind and not just the trophy moment.
\nWhat I want to emphasise is \"hard work\" before the trophy. And, not just the trophy is everything philosophy. Because at that moment, the real gold is his/her talent and not that of trophys'
\nAlso, I am not picking up sides. GM or Novice. Medal or not. Just enjoy Kaggling!
", "rawMarkdown": "AFAIK, a player who is passionate about football and works hard would win the trophy and still cherish the hard work behind and not just the trophy moment. \n\nWhat I want to emphasise is \"hard work\" before the trophy. And, not just the trophy is everything philosophy. Because at that moment, the real gold is his/her talent and not that of trophys' \n\nAlso, I am not picking up sides. GM or Novice. Medal or not. Just enjoy Kaggling!", "votes": -2 }, { "id": 1321645, "postDate": "2021-05-24T20:23:10.450Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1321648, "postDate": "2021-05-24T20:27:18.033Z", "content": "Let's Just enjoy Kaggling! :)
", "rawMarkdown": "Let's Just enjoy Kaggling! :)", "votes": 1 }, { "id": 1321738, "postDate": "2021-05-24T22:49:29.653Z", "content": "Segregating competitions based on tiers seems a cure worse than the disease.
\nIf your solution were implemented, the current #1 team, which consists of unranked members, would not be allowed to participate (and potentially win) against the #2 GM team.
\nP.S. I suspect the reason why you're getting downvoted so heavily is that it comes across as arrogant to make strong, definitive statements about being a data science professional when someone is only a couple years out of college.
", "rawMarkdown": "Segregating competitions based on tiers seems a cure worse than the disease.\n\nIf your solution were implemented, the current #1 team, which consists of unranked members, would not be allowed to participate (and potentially win) against the #2 GM team. \n\nP.S. I suspect the reason why you're getting downvoted so heavily is that it comes across as arrogant to make strong, definitive statements about being a data science professional when someone is only a couple years out of college. ", "votes": 8 }, { "id": 1321857, "postDate": "2021-05-25T03:39:01.997Z", "content": "Ah, the idea was to have a separate competition category organised by Kaggle on yearly basis. All the featured competitions would be unaffected. So, the current #1 team can participate and still win.
\nBut I agree. We have questions like;
\nI am gonna keep this thread quiet, as I understand many users have an alternate opinion, and I respect that. Thanks
", "rawMarkdown": "Ah, the idea was to have a separate competition category organised by Kaggle on yearly basis. All the featured competitions would be unaffected. So, the current #1 team can participate and still win.\n\nBut I agree. We have questions like;\n1. How many users would enroll?\n2. Splitting based on tiers make sense? What will be active engagement over such competitions\n3. Will it become easier to get gold than a featured competition? Does current medal progression hold good?\n\nI am gonna keep this thread quiet, as I understand many users have an alternate opinion, and I respect that. Thanks\n" }, { "id": 1322033, "postDate": "2021-05-25T06:45:34.940Z", "content": "You mentioned that “ I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success. “
\nPlease shed some light on what the necessities are for data science career success. Thanks.
", "rawMarkdown": "You mentioned that “ **I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success.** “\n\nPlease shed some light on what the necessities are for data science career success. Thanks. ", "votes": 4 }, { "id": 1322070, "postDate": "2021-05-25T07:10:11.337Z", "content": "\n\nI talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success.
\n
It is the opposite. Being Master/GM doesn't increase your chance to pass CV screening but when it comes to technical interviews, a true (Grand)Master can easily outperform other candidates.
", "rawMarkdown": "> I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success.\n\nIt is the opposite. Being Master/GM doesn't increase your chance to pass CV screening but when it comes to technical interviews, a true (Grand)Master can easily outperform other candidates.", "votes": 7 }, { "id": 1322094, "postDate": "2021-05-25T07:38:24.207Z", "content": "\n\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford.
\n
What kind of machine is that? You can build a good kaggle-worthy machine for much less.
", "rawMarkdown": "> The cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford.\n\nWhat kind of machine is that? You can build a good kaggle-worthy machine for much less. ", "votes": 8 }, { "id": 1322110, "postDate": "2021-05-25T07:53:53.807Z", "content": "Reasons for Kaggling:
\nFirst a Hollywood prologue of it: AI is in the dawn of a new area, like internet and pc once were, what we do now will echo in eternity! 😎
\nAnd some other reasons:
\nThere are many ways one can learn and fill the up-to-date knowledge and skills of Data Science, Python programming and AI and there is a lot to read and test out there 😉 this is defeatedly a great and fun way doing it. And if one needs a benchmark of the present/past AI knowledge(self, recruiters etc), here is one way getting it.
\nI can’t talk for the GM/best-in-the-area, but I can bet them also are learning tons of new techniques and ideas through the competitions, from their own work and by reading others, a way of many ways, for them to also learn and keep up-to-date as the rest of us. I would be scared if the code just magical popped up for them 😉, is it? 😲 No, I believe they are smart people working hard and dedicated many hours per day (and night) for their AI passion, hence the position/GM-title.
\nHere I will add a big-thanks-code-loop(can it get better 😉) to the hosts that support with a competition, Kaggle that support with the HW and to those who always brings insight and great hints to the competition’s problems.\nwell said @kirderf I like your \"Reasons for Kaggling\"
\nI also like that when Bojan Tunguz become GM he shared this.
\n\"in Kaggle I discovered an incredible place where I could learn and develop my skills. The continuous pace of the competitions, and the challenge of working in a hyper-competitive environment, with some smartest people in the world, forces you to push yourself far beyond what you would normally achieve in any other educational setting.\"
\nhttps://www.kaggle.com/c/ncaaw-march-mania-2021/discussion/230725
also when I see Kaggler like K.Amano
\nhttps://www.kaggle.com/amanooo
I start to see Kaggle as a lifestyle :)
", "rawMarkdown": "well said @kirderf I like your \"Reasons for Kaggling\"\n\nI also like that when Bojan Tunguz become GM he shared this. \n\n`\"in Kaggle I discovered an incredible place where I could learn and develop my skills. The continuous pace of the competitions, and the challenge of working in a hyper-competitive environment, with some smartest people in the world, forces you to push yourself far beyond what you would normally achieve in any other educational setting.\" `\nhttps://www.kaggle.com/c/ncaaw-march-mania-2021/discussion/230725\n\nalso when I see Kaggler like K.Amano \nhttps://www.kaggle.com/amanooo\n\nI start to see Kaggle as a lifestyle :)\n\n\n\n ", "votes": 2 } ] }, { "id": 1310614, "postDate": "2021-05-16T19:03:15.360Z", "content": "Disagree. At least for this competition, LB<1.0 is not something that unlimited hardware resource is able to achieve. The interplay between data science/DL and domain knowledge is crucial to bring a giant leap in the LB score, compared to the pure training set-driven model.
", "rawMarkdown": "Disagree. At least for this competition, LB<1.0 is not something that unlimited hardware resource is able to achieve. The interplay between data science/DL and *domain knowledge* is crucial to bring a giant leap in the LB score, compared to the pure training set-driven model. ", "replies": [ { "id": 1310630, "postDate": "2021-05-16T19:13:06.573Z", "content": "\n\nThe interplay between data science/DL and domain knowledge is crucial to bring a giant leap in the LB score.
\n
Adequate hardware is also crucial, I would say.
", "rawMarkdown": "> The interplay between data science/DL and domain knowledge is crucial to bring a giant leap in the LB score.\n\nAdequate hardware is also crucial, I would say.", "votes": 3 } ] } ], "comments": [ { "id": 1309688, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-16T07:58:13.043000", "content": "You're probably 4 places higher because of that.
", "votes": 19, "replies": [ { "id": 1311197, "author_name": "Alexander Soare", "author_url": "", "post_date": "2021-05-17T08:17:12.433000", "content": "You're on a roll mate
", "votes": 1, "replies": [] }, { "id": 1311310, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-17T10:07:38.737000", "content": "Hi there! :D
", "votes": 0, "replies": [] }, { "id": 1316346, "author_name": "John Mitchell", "author_url": "", "post_date": "2021-05-20T13:30:50.847000", "content": "While I admire @mconway's concern for the small-time players, I absolutely agree with @nofreewill's observation that a merger at the top moves me a place or two up the LB. If five GMs are playing as individuals, maybe the best I can hope for is sixth. I they merge into a superteam, I can target second place. In fact, mergers push the pursuit of Kaggle medals further away from being a zero sum game.
", "votes": 1, "replies": [] }, { "id": 1316385, "author_name": "Chun Ming Lee", "author_url": "", "post_date": "2021-05-20T13:57:32.553000", "content": "That's assuming there's no synergy in teaming up. I think the way people are describing GMs here (natural talent a la Kipchoge) is a bit silly especially when you look at many GMs' solo track record or lack thereof.
\nKaggle is a platform with a bunch of arbitrary rules surrounding competitions - a lot of which are geared towards gamifying data science, not necessarily identifying the best data scientists in the world.
", "votes": 3, "replies": [] }, { "id": 1316472, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-20T15:30:25.610000", "content": "Bronze is definitely easier to get because of mergers, silver most probably is, too. But they make the very top places harder and harder to reach, because merging definitely makes scores higher compared to the members' individual scores.
", "votes": 1, "replies": [] }, { "id": 1316491, "author_name": "CPMP", "author_url": "", "post_date": "2021-05-20T15:36:47.663000", "content": "\n\nespecially when you look at many GMs' solo track record or lack thereof.
\n
@leecming I agree on the assessment but maybe not on the cause. Requirement for competition GM is one solo gold. There is no upside in getting more solo gold medals. Why would GMs take risks in pursuing solo gold medals?
\nI personally think that competition GM should have at least 2 gold medals. It would filter the lucky winners. But Kaggle is not designed to evaluate individuals as you wrote. It is designed to provide value to competition sponsors. Encouraging strong teaming makes more sense for that.
", "votes": 6, "replies": [] }, { "id": 1316821, "author_name": "Chun Ming Lee", "author_url": "", "post_date": "2021-05-20T23:04:47.120000", "content": "Kaggle should introduce another tier above GM with even more requirements.
\nWhat would it be called? Start using royal ranks - I wouldn't mind being a Duke or Baron of Kaggle ;)
More seriously, Kaggle could consider a martial arts or Go (boardgame) style \"dan ranking\" which has numerical groupings. Maybe for each solo gold you get, you go up one \"dan\".
", "votes": 1, "replies": [] }, { "id": 1316832, "author_name": "Gabriel Lindenmaier", "author_url": "", "post_date": "2021-05-20T23:26:40.427000", "content": "\n\nKaggle should introduce another tier above GM
\n
There is the GGGGMaster ;)
\nThe points which decay are for the fine-grained ranking, including the fact that you see your standing among all Kagglers.
", "votes": 1, "replies": [] } ] }, { "id": 1309637, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2021-05-16T06:58:06.650000", "content": "Kaggle was never a fair game and won't be. It is a crowdsourcing platform for getting the best solution for its customers. Unfortunately it is not chess.
", "votes": 13, "replies": [ { "id": 1309660, "author_name": "Chun Ming Lee", "author_url": "", "post_date": "2021-05-16T07:33:10.403000", "content": "Is it though? Limiting team sizes to a max of five indicates that there are other considerations than just sourcing for the \"best solution\".
\nI've always been a skeptic of the argument that Kaggle is mainly about delivering models to sponsors. It seems more a pedagogical platform with sponsors getting PR or \"building awareness\" more than anything else.
", "votes": 7, "replies": [] }, { "id": 1309670, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2021-05-16T07:45:14.517000", "content": "You are also right. My point is that there is some gamification but it is not a pure skill game. Hardware, luck, teaming strategies, using leaks are as important as skills. You can make it to Kaggle Top 10 with limited Machine Learning skills. The same is impossible for chess.
", "votes": 7, "replies": [] }, { "id": 1310536, "author_name": "Pavel Orlov", "author_url": "", "post_date": "2021-05-16T18:09:35.390000", "content": "In chess, it is generally impossible for a person to get to the TOP … alpha zero will still win. And here it is possible for everyone (theoretically)
", "votes": 1, "replies": [] }, { "id": 1310656, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2021-05-16T19:44:37.637000", "content": "\n\nIn chess, it is generally impossible for a person to get to the TOP … alpha zero will still win. And here it is possible for everyone (theoretically)
\n
DeepMind: \"hold my beer\"
\nBut yeah, true AutoML seems to be a bit far away, but who knows.
", "votes": 1, "replies": [] } ] }, { "id": 1310758, "author_name": "Kirderf", "author_url": "", "post_date": "2021-05-16T23:25:50.090000", "content": "Yes your are right, it is the case here, partly, but that is not the end of the story.
\nI also had this thought in my Kaggle beginning, those with more/powerful HW had an advantage but with now 50+ competitions, I believe it’s more a solution game before HW game. Yes one can train the same idea and model quicker, but in the end, the idea itself behind is the key. And if one want advantage in calculation power, one can start the competition early, use free Kaggle GPU/TPU, Colab GPU/TPU, it will give many hours.
\nAnd with many techniques one can squeeze more out of the code, like using Mixed Precision with Volta devices, using GPU instead of CPU in the whole code, all making it possible to train bigger model and faster training. And in HW power terms the Kaggle TPU version have power.
\nFor example the Google Landmark Retrieval 2020 competition with 5M images in the training set, that competition was also a HW competition but by analyzing the winners write-up, the person only used Colab/Kaggle TPU, ~400 hours, what a dedicated work, and it was also a brilliant solution, respect.
\nBy reading the top solutions writeup from every competition, one can see that these smart people often came up with a new level of SOTA AI / Data Science solution that help both solving the problem to the host and also takes the AI area to the next level. When I read them, I don’t think of what GPU/TPU version they used, I more see the bright idea and innovative code, and how they solved the labyrinth step by step.
\nIn terms of HW power advantages I think cloud Auto AI solution that bring a solution in minutes through benchmarking different models and tuning parameters can really help, but those are often licensed and not often seen used here, can believe that a quantum computer under the hood to a full current-market-solutions-loaded Auto AI would give some hint where to start 😉.
In Kaggle competition the HW really give an edge, oh yes, but the Data Science input and digging the path in new/unknow territory is crucial to grab the first place.
", "votes": 11, "replies": [] }, { "id": 1309680, "author_name": "Psi", "author_url": "", "post_date": "2021-05-16T07:52:42.413000", "content": "Access to powerful hardware is indeed an advantage in this competition as it seems. But I do not believe this has necessarily anything to do with the number of grandmasters in a team. There are some rather \"unknown\" kagglers with much more hardware than you can imagine, and then there are some top kagglers with much less hardware than you would think.
\nIn this competition it was clear from the start that it will be at least partly a HW game. There have been quite a few competitions lately, and I wonder why you decided to join this competition and not others if that bothers you? I don't think this will be ever a fully fair game, unless Kaggle decides to do kernel-only fitting competitions again, which doesn't seem to be their plan.
", "votes": 12, "replies": [ { "id": 1309707, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2021-05-16T08:13:33.190000", "content": "Also, not all competitions are hardware hungry and rather need clever thinking and clever solutions. You may find those more enjoyable to try.
\nLike in all sports or competitions: Equipment (Hardware) can give you an advantage, but it's often times less than you would expect. Kipchoge would still outrun most people with only sandals. Tennis pros would still dominante amateurs with a beginner racket. It usually only gives you that last percentage that makes the difference between the very very best.
", "votes": 7, "replies": [] } ] }, { "id": 1316559, "author_name": "Serigne ", "author_url": "", "post_date": "2021-05-20T16:41:42.823000", "content": "I think Kaggle should promote more solo competitions, beside teams ones. If I'm not mistaken, the last one is Instacart held 4 years ago. And that could reduce most of the gamifying part.
", "votes": 9, "replies": [ { "id": 1316677, "author_name": "Gabriel Lindenmaier", "author_url": "", "post_date": "2021-05-20T19:16:21.237000", "content": "Well, considering that competition hosts have priority in the end, what could happen, is that Kaggle will be looking (more) at studies of team dynamics.
\nSo in the end we could - this is a hypothetical example - have e.g. something like max teams of two where creativity is far more important, as pairs are the most efficient teams.
\nSo the current state might be, ridiculous as it seems, fairer while still satisfying competition hosts.
Interesting other avenue of investigation: What is the correlation between teams and private leaderboard placement? One could answer that with available data. As almost anyone can team up with others, that could help to improve expected outcome for those who value it more than whatever else might speak against being in a team.
", "votes": 0, "replies": [] }, { "id": 1321074, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-05-24T14:01:25.890000", "content": "Kaggle Marathons. Where Kaggle organises, no hosts. They are medal based competitions hosted by Kaggle for respective tiers. I have proposed how it works in a comment below.
", "votes": 0, "replies": [] } ] }, { "id": 1311240, "author_name": "Charles", "author_url": "", "post_date": "2021-05-17T08:59:42.427000", "content": "As my father always said to me when I was a child, \"life is not fair\".
", "votes": 9, "replies": [] }, { "id": 1313461, "author_name": "Gabriel Lindenmaier", "author_url": "", "post_date": "2021-05-18T15:15:30.610000", "content": "The CommonLit Readability competition looks to be really good for those with low amounts of hardware. With the readability score composition being non-disclosed - as far as I can tell - you also can't just aggregate large amounts of extra data.
\nSo I suggest taking part in that competition, if you want one where ingenuity is far more important than computational power.
Of course you could argue that four grandmasters have a higher expected ingenuity. But I say, it is more their experience that might let them make less mistakes and compose ideas better and know more of them. Though you can still be very creative and read paper (abstracts) without that.
\nSo it is certainly fairer than most other scenarios you can hope for on Kaggle.
You raise an interesting point but I am afraid you jump to wrong conclusions. For instance:
\n\n\nWhen you have a team that collectively works for a company with access to unlimited hardware and resources
\n
They work for different companies. It so happen that I work in the same company as two of them. I wished that what you wrote was true, but it isn't, by any mean. Don't take their team name as face value, it is not the reality.
\nI know of another company that let his kagglers use up to 120 V100 GPU. We have access to a small fraction of that at our company.
\nI reacted just to straighten a bit the expectations here. Compute resource matters, no doubt. But it is not the end of the story.
\nAnd about having many GMs in a single team, it is not the first nor the last time it happens. A team of 3 GMS has won a series of competition recently for instance. No one found it unfair AFAIK.
", "votes": 3, "replies": [ { "id": 1311838, "author_name": "Mark Conway", "author_url": "", "post_date": "2021-05-17T16:33:27.150000", "content": "We can agree to disagree. Nvidia is using its Grandmasters (two in this case) to market expensive hardware unattainable to many, thus the team name. Maybe I just have a different viewpoint because I like to help the less fortunate. My friend @robertdscott overcame extreme adversity to become a Data Scientist, and I don't see Nvidia doing anything to level the playing field.
", "votes": 2, "replies": [] }, { "id": 1311883, "author_name": "CPMP", "author_url": "", "post_date": "2021-05-17T17:20:46.263000", "content": "I am very flattered to be the only one you responded to. Thanks.
", "votes": 6, "replies": [] } ] }, { "id": 1318833, "author_name": "Theo Viel", "author_url": "", "post_date": "2021-05-22T15:46:43.387000", "content": "The four GMs are no longer at the top though ^^
", "votes": 3, "replies": [ { "id": 1319421, "author_name": "Chun Ming Lee", "author_url": "", "post_date": "2021-05-23T07:06:09.947000", "content": "Kudos to the (literally) green trio for at least temporarily beating a team of 4 GMs
", "votes": 3, "replies": [] } ] }, { "id": 1317634, "author_name": "Faisal Alsrheed", "author_url": "", "post_date": "2021-05-21T15:03:39.693000", "content": "Competition is not the right word to describe what we are doing on Kaggle. It should be called a challenge.
\nWe are all here learning together to solve a challenge for a host and to improve our skills. We are not here competing with each other like a sporting event with equal types of equipment.
So enjoy learning and professional networking on Kaggle and see it as a fun challenge, not like a chess-boxing competition
", "votes": 4, "replies": [ { "id": 1317933, "author_name": "Ali Abdin", "author_url": "", "post_date": "2021-05-21T20:09:17.193000", "content": "I agree, in a competition people would not help each other by sharing e.g. code and bugs etc.
", "votes": 2, "replies": [] } ] }, { "id": 1309649, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2021-05-16T07:09:24.037000", "content": "I doubt the \"unlimited\".
\nTheretically, everyone can rent as many Cloud Computing Hardware as she or he wants to get an edge. It just depends on the willingness of the kaggler to pay for that.
\nSame here, if a company wants to provide ressources (they do have a cost) for an expected or even unknown return, why not?
In the end, it is business. A company or a group is paying for a solution. Best solution wins. If it needs large investments to get that best solution you need to calculate the risk and reward ratio.
", "votes": 4, "replies": [] }, { "id": 1310917, "author_name": "Long Luu", "author_url": "", "post_date": "2021-05-17T04:06:49.987000", "content": "Grandmaster title tells you:
\nAnd I believe they are equally important. The second team on the LB are all novice Kagglers.
\nKaggle was never fair, and will never be. The organizers spend a bunch of money to find a solution, not to organize a competition. First place prize is $25k, what hardware can you build with 25k?
", "votes": 3, "replies": [] }, { "id": 1318067, "author_name": "Caneiro", "author_url": "", "post_date": "2021-05-22T01:00:55.010000", "content": "Could more code competitions help to balance that?
", "votes": 1, "replies": [ { "id": 1321070, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-05-24T14:00:17.230000", "content": "Some code competitions allow \"Models\" as an external dataset. So, still, train on your rigs. Upload model here, make it public/private on the last day.
\nBeen there. Its pseudo fair
", "votes": 0, "replies": [] }, { "id": 1321139, "author_name": "Caneiro", "author_url": "", "post_date": "2021-05-24T14:15:51.313000", "content": "I see.. thanks!
", "votes": 1, "replies": [] } ] }, { "id": 1323310, "author_name": "Allohvk", "author_url": "", "post_date": "2021-05-26T06:36:32.477000", "content": "When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle.
But this is an assumption right? Maybe it is not true. Maybe there are experts and Masters out there whose companies are giving them even better GPU times. And even if this is true, this is perfectly legit. They worked hard, became GMs and maybe now have access to better resources.
\nThe other point about 4 GM on one team..No rule prohibits this in Kaggle. Whether it is right in spirit - I am not sure. Where do we draw the line - 1 GM per team or 2. Why not 3?
\nHaving said that, I have neither upvoted or downvoted this thread. It has an important point. Maybe Kaggle and Kagglers would have more diversity in solutions if they allowed only 1 GM per team? The sheer brilliance and knowledge/expertise each one carries can benefit the community as a whole when they are pitted against each other…So there is a point there….
", "votes": 2, "replies": [] }, { "id": 1316785, "author_name": "sin", "author_url": "", "post_date": "2021-05-20T21:56:16.423000", "content": "Me: let's run an experiment.
\nOne month later…
\nGPU: I am really tired and I only finished 3 epochs.
Joined Kaggle as a student, have 1 gold medal. Now I am 2 yoe. Here's my story:
\nI joined Kaggle as a student. I had an 8 GB laptop when I started. I was in situations where I couldn't run EDA scripts forget competitions. I bagged silver and also team gold medals.
\nMy experiences were good and bad.
\nYes, I got that gold in Code Competitions and without them, I would not probably have gold (in team) today
\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford. And with taxes and customs, just forget them in India. What say, students? Do some Indian users relate?
\nCode Competitions level the playing field by a lot (that especially don't allow pre-run models). I have worked in a professional capacity where Kaggle was a part-time job and also where Kaggling wasn't allowed. It clearly makes it more feasible when you do Kaggle in your day job. Again, unfair but it's the game and how it works.
\n\n\nI believe having a paid job to do Kaggle is what leads one to better places + corporate GPUs boost chances further
\n
But, nonetheless. There are few competitions that are really about feature engineering and not much raw power. And I believe with excellent engineering you can beat the x100 raw performances. I have seen it in practice and earned medals as well.
\nHere are some takeaways I have from my Kaggling experiences;
\nIf you want to compete for earning medals;
\nMake competition spaces. Where Kagglers from their tier only to get compete in yearly seasons.
\nEvery top 20% get a chance to compete in competitions above +1 their tier. So top users from Masters winning competition can compete now with GMs. This way it's a huge motivation to learn, grow in rankings have a fair chance to grab golds.
\nAnd yes, the criterion to be Competitions GM of solo gold is bad. Its harder in a cutting edge CV competition today than some Weather prediction competition 6 years back? Apparently, early to Kaggle would have made you GM than today's competition? I wonder what current GMs think of this?
\nAlso, a single gold should be equated to x10 silvers. Or medal to medal trade-in.
\n\n\nIn this pandemic, I realised. Maybe it's more important to be in your home town, share more stories with family and have a better work-life balance than Kaggling all weekend.
\n
I do want to become GM one day. Maybe discussions/notebooks are feasible. But due to less access to hardware and lot of CV competitions, it's definitely on the harder side for me.
\n(Edited x2)
", "votes": -2, "replies": [ { "id": 1321225, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-05-24T15:02:55.747000", "content": "Got downvoted. Would be glad to have healthy conversations on why someone would disagree, and what exactly or all of it. Perhaps any constructive feedback would be great. Much appreciated.
", "votes": 4, "replies": [] }, { "id": 1321346, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-24T16:25:39.997000", "content": "I didn't read your whole comment, just few parts of it.
\nBut what I've read, I can see where that downvote could come from. As you said, it would have been nice if the downvoter wrote his points.
You write that Yes, I got that gold in Code Competitions and without them, I would not probably have gold today
\nBut after seeing your profile, I guess it was not much about Code Competitions for you, but being in a group of 6. Maybe for the hole team it was more about Code Competitions, but it was more about being in a big team for the individuals that brought the gold. I guess!!
\nBut you've written about that gold as if it was a solo one.
Also you write that Medals are monitor pixels. You aren't getting gold for real. Buy dogecoin if you wanna be rich
\nDogecoins are just as virtual as kaggle medals. Also your money is virtual. Cash has only value because we know that others will accept it, too. But if not, it worth nothing. Kaggle medals are valuable because you can show that you accomplished something. So those pixels are a lot more valuable than some other pixels.
You write that Now I am a young professional with 2 yoe.
\nI have made a face tracking little toy-machine as my BSc thesis and made a better version of it as my MSc thesis. I've put a lot of work into them and learned a lot about deep learning along the way. So much that I got into the field and have a job for 1.5 years now. I think that I work pretty hard to become better and better, but I'm far from being a professional.
\nSo I find it pretty hard to believe that you've become a professional in just two years. Or I may just misunderstand the word \"professional\".
\nI would say that a professional is closer to a senior than a junior, but I may be wrong. Correct me, if I'm wrong.
Hey, thanks for the comment.
\nI wrote
\n\n\nI bagged silver and also team gold medals.
\n
to make it explicit the gold medal was from the team and not solo. Perhaps I can edit and make it more apparent.
\n\n\nMedals are monitor pixels. You aren't getting gold for ….
\n
I agree with this. May be was too sarcastic. I deeply respect the effort behind the medals. But, having too many feelings attached, the rush and the whole emotions to just medals ( IMO of users and their past behaviour)
\n\n\n\"professional\" meant on a full-time job / working professionally for me. No expert, no junior & no senior.
\n
Maybe I will refactor the comment now. Thanks!
\nEdit:
\nDone now. Thanks
", "votes": 0, "replies": [] }, { "id": 1321458, "author_name": "Faisal Alsrheed", "author_url": "", "post_date": "2021-05-24T17:36:29.813000", "content": "\nyou said
\nDon't get too attached to earning medals.`\n\nMedals are monitor pixels. You aren't getting gold for real.\n\nI talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill.\njust imagine telling a football player in FIFA World Cup, when he is having the time of his life (who worked so hard to get to this stage in his life)
\n\" Hey, This is just a game, the world cup will not guarantee you a place in the best football clubs, and …….by the way, the Trophy is not real gold\" 😂
", "votes": 1, "replies": [] }, { "id": 1321608, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-24T19:54:48.683000", "content": "I didn't say anything like that.
\nEdit: Thanks for the correction! :)
", "votes": 1, "replies": [] }, { "id": 1321641, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-05-24T20:18:17.530000", "content": "AFAIK, a player who is passionate about football and works hard would win the trophy and still cherish the hard work behind and not just the trophy moment.
\nWhat I want to emphasise is \"hard work\" before the trophy. And, not just the trophy is everything philosophy. Because at that moment, the real gold is his/her talent and not that of trophys'
\nAlso, I am not picking up sides. GM or Novice. Medal or not. Just enjoy Kaggling!
", "votes": -2, "replies": [] }, { "id": 1321645, "author_name": "", "author_url": "", "post_date": "2021-05-24T20:23:10.450000", "content": "", "votes": 0, "replies": [] }, { "id": 1321648, "author_name": "Faisal Alsrheed", "author_url": "", "post_date": "2021-05-24T20:27:18.033000", "content": "Let's Just enjoy Kaggling! :)
", "votes": 1, "replies": [] }, { "id": 1321738, "author_name": "Chun Ming Lee", "author_url": "", "post_date": "2021-05-24T22:49:29.653000", "content": "Segregating competitions based on tiers seems a cure worse than the disease.
\nIf your solution were implemented, the current #1 team, which consists of unranked members, would not be allowed to participate (and potentially win) against the #2 GM team.
\nP.S. I suspect the reason why you're getting downvoted so heavily is that it comes across as arrogant to make strong, definitive statements about being a data science professional when someone is only a couple years out of college.
", "votes": 8, "replies": [] }, { "id": 1321857, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2021-05-25T03:39:01.997000", "content": "Ah, the idea was to have a separate competition category organised by Kaggle on yearly basis. All the featured competitions would be unaffected. So, the current #1 team can participate and still win.
\nBut I agree. We have questions like;
\nI am gonna keep this thread quiet, as I understand many users have an alternate opinion, and I respect that. Thanks
", "votes": 0, "replies": [] }, { "id": 1322033, "author_name": "sin", "author_url": "", "post_date": "2021-05-25T06:45:34.940000", "content": "You mentioned that “ I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success. “
\nPlease shed some light on what the necessities are for data science career success. Thanks.
", "votes": 4, "replies": [] }, { "id": 1322070, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2021-05-25T07:10:11.337000", "content": "\n\nI talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success.
\n
It is the opposite. Being Master/GM doesn't increase your chance to pass CV screening but when it comes to technical interviews, a true (Grand)Master can easily outperform other candidates.
", "votes": 7, "replies": [] }, { "id": 1322094, "author_name": "Psi", "author_url": "", "post_date": "2021-05-25T07:38:24.207000", "content": "\n\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford.
\n
What kind of machine is that? You can build a good kaggle-worthy machine for much less.
", "votes": 8, "replies": [] }, { "id": 1322110, "author_name": "Kirderf", "author_url": "", "post_date": "2021-05-25T07:53:53.807000", "content": "Reasons for Kaggling:
\nFirst a Hollywood prologue of it: AI is in the dawn of a new area, like internet and pc once were, what we do now will echo in eternity! 😎
\nAnd some other reasons:
\nThere are many ways one can learn and fill the up-to-date knowledge and skills of Data Science, Python programming and AI and there is a lot to read and test out there 😉 this is defeatedly a great and fun way doing it. And if one needs a benchmark of the present/past AI knowledge(self, recruiters etc), here is one way getting it.
\nI can’t talk for the GM/best-in-the-area, but I can bet them also are learning tons of new techniques and ideas through the competitions, from their own work and by reading others, a way of many ways, for them to also learn and keep up-to-date as the rest of us. I would be scared if the code just magical popped up for them 😉, is it? 😲 No, I believe they are smart people working hard and dedicated many hours per day (and night) for their AI passion, hence the position/GM-title.
\nHere I will add a big-thanks-code-loop(can it get better 😉) to the hosts that support with a competition, Kaggle that support with the HW and to those who always brings insight and great hints to the competition’s problems.\nwell said @kirderf I like your \"Reasons for Kaggling\"
\nI also like that when Bojan Tunguz become GM he shared this.
\n\"in Kaggle I discovered an incredible place where I could learn and develop my skills. The continuous pace of the competitions, and the challenge of working in a hyper-competitive environment, with some smartest people in the world, forces you to push yourself far beyond what you would normally achieve in any other educational setting.\"
\nhttps://www.kaggle.com/c/ncaaw-march-mania-2021/discussion/230725
also when I see Kaggler like K.Amano
\nhttps://www.kaggle.com/amanooo
I start to see Kaggle as a lifestyle :)
", "votes": 2, "replies": [] } ] }, { "id": 1310614, "author_name": "human intelligence", "author_url": "", "post_date": "2021-05-16T19:03:15.360000", "content": "Disagree. At least for this competition, LB<1.0 is not something that unlimited hardware resource is able to achieve. The interplay between data science/DL and domain knowledge is crucial to bring a giant leap in the LB score, compared to the pure training set-driven model.
", "votes": 0, "replies": [ { "id": 1310630, "author_name": "nofreewill42", "author_url": "", "post_date": "2021-05-16T19:13:06.573000", "content": "\n\nThe interplay between data science/DL and domain knowledge is crucial to bring a giant leap in the LB score.
\n
Adequate hardware is also crucial, I would say.
", "votes": 3, "replies": [] } ] } ], "raw_markdown_by_id": { "1309688": "You're probably 4 places higher because of that.", "1309637": "Kaggle was never a fair game and won't be. It is a crowdsourcing platform for getting the best solution for its customers. Unfortunately it is not chess.", "1310758": "Yes your are right, it is the case here, partly, but that is not the end of the story.\n\nI also had this thought in my Kaggle beginning, those with more/powerful HW had an advantage but with now 50+ competitions, I believe it’s more a solution game before HW game. Yes one can train the same idea and model quicker, but in the end, the idea itself behind is the key. And if one want advantage in calculation power, one can start the competition early, use free Kaggle GPU/TPU, Colab GPU/TPU, it will give many hours.\n\nAnd with many techniques one can squeeze more out of the code, like using Mixed Precision with Volta devices, using GPU instead of CPU in the whole code, all making it possible to train bigger model and faster training. And in HW power terms the Kaggle TPU version have power.\n\nFor example the Google Landmark Retrieval 2020 competition with 5M images in the training set, that competition was also a HW competition but by analyzing the winners write-up, the person only used Colab/Kaggle TPU, ~400 hours, what a dedicated work, and it was also a brilliant solution, respect.\n\nBy reading the top solutions writeup from every competition, one can see that these smart people often came up with a new level of SOTA AI / Data Science solution that help both solving the problem to the host and also takes the AI area to the next level. When I read them, I don’t think of what GPU/TPU version they used, I more see the bright idea and innovative code, and how they solved the labyrinth step by step. \nIn terms of HW power advantages I think cloud Auto AI solution that bring a solution in minutes through benchmarking different models and tuning parameters can really help, but those are often licensed and not often seen used here, can believe that a quantum computer under the hood to a full current-market-solutions-loaded Auto AI would give some hint where to start 😉. \n\nIn Kaggle competition the HW really give an edge, oh yes, but the Data Science input and digging the path in new/unknow territory is crucial to grab the first place.", "1309680": "Access to powerful hardware is indeed an advantage in this competition as it seems. But I do not believe this has necessarily anything to do with the number of grandmasters in a team. There are some rather \"unknown\" kagglers with much more hardware than you can imagine, and then there are some top kagglers with much less hardware than you would think. \n\nIn this competition it was clear from the start that it will be at least partly a HW game. There have been quite a few competitions lately, and I wonder why you decided to join this competition and not others if that bothers you? I don't think this will be ever a fully fair game, unless Kaggle decides to do kernel-only fitting competitions again, which doesn't seem to be their plan.", "1316559": "I think Kaggle should promote more solo competitions, beside teams ones. If I'm not mistaken, the last one is [Instacart](https://www.kaggle.com/c/instacart-market-basket-analysis/overview) held 4 years ago. And that could reduce most of the gamifying part. ", "1311240": "As my father always said to me when I was a child, *\"life is not fair\"*. ", "1309400": "When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. As I respect each of these Grandmasters individually, I'm disappointed in this kind of anti-competitive behavior. One could counter-argue that all you have to do is come up with a better solution, but access to powerful hardware is a clear edge in certain competitions where cycles are important for a winning solution.", "1313461": "The [CommonLit Readability competition ](https://www.kaggle.com/c/commonlitreadabilityprize/overview) looks to be really good for those with low amounts of hardware. With the readability score composition being non-disclosed - as far as I can tell - you also can't just aggregate large amounts of extra data. \nSo I suggest taking part in that competition, if you want one where ingenuity is far more important than computational power. \n\nOf course you could argue that four grandmasters have a higher expected ingenuity. But I say, it is more their experience that might let them make less mistakes and compose ideas better and know more of them. Though you can still be very creative and read paper (abstracts) without that. \nSo it is certainly fairer than most other scenarios you can hope for on Kaggle.", "1311458": "You raise an interesting point but I am afraid you jump to wrong conclusions. For instance:\n\n> When you have a team that collectively works for a company with access to unlimited hardware and resources\n\nThey work for different companies. It so happen that I work in the same company as two of them. I wished that what you wrote was true, but it isn't, by any mean. Don't take their team name as face value, it is not the reality.\n\nI know of another company that let his kagglers use up to 120 V100 GPU. We have access to a small fraction of that at our company. \n\nI reacted just to straighten a bit the expectations here. Compute resource matters, no doubt. But it is not the end of the story.\n\nAnd about having many GMs in a single team, it is not the first nor the last time it happens. A team of 3 GMS has won a series of competition recently for instance. No one found it unfair AFAIK.", "1318833": "The four GMs are no longer at the top though ^^", "1317634": "Competition is not the right word to describe what we are doing on Kaggle. It should be called a challenge.\nWe are all here learning together to solve a challenge for a host and to improve our skills. We are not here competing with each other like a sporting event with equal types of equipment.\n\nSo enjoy learning and professional networking on Kaggle and see it as a fun challenge, not like a chess-boxing competition", "1309649": "I doubt the \"unlimited\".\nTheretically, everyone can rent as many Cloud Computing Hardware as she or he wants to get an edge. It just depends on the willingness of the kaggler to pay for that. \nSame here, if a company wants to provide ressources (they do have a cost) for an expected or even unknown return, why not?\n\nIn the end, it is business. A company or a group is paying for a solution. Best solution wins. If it needs large investments to get that best solution you need to calculate the risk and reward ratio.", "1310917": "Grandmaster title tells you:\n- How much effort you have spent on the platform\n- How good you are\n\nAnd I believe they are equally important. The second team on the LB are all novice Kagglers.\n\nKaggle was never fair, and will never be. The organizers spend a bunch of money to find a solution, not to organize a competition. First place prize is $25k, what hardware can you build with 25k?", "1318067": "Could more code competitions help to balance that?", "1323310": "`When you have a team that collectively works for a company with access to unlimited hardware and resources, this kind of arrangement violates the spirit of Kaggle. `\n\nBut this is an assumption right? Maybe it is not true. Maybe there are experts and Masters out there whose companies are giving them even better GPU times. And even if this is true, this is perfectly legit. They worked hard, became GMs and maybe now have access to better resources. \n\nThe other point about 4 GM on one team..No rule prohibits this in Kaggle. Whether it is right in spirit - I am not sure. Where do we draw the line - 1 GM per team or 2. Why not 3? \n\nHaving said that, I have neither upvoted or downvoted this thread. It has an important point. Maybe Kaggle and Kagglers would have more diversity in solutions if they allowed only 1 GM per team? The sheer brilliance and knowledge/expertise each one carries can benefit the community as a whole when they are pitted against each other...So there is a point there....", "1316785": "Me: let's run an experiment. \nOne month later... \nGPU: I am really tired and I only finished 3 epochs.", "1321032": "Joined Kaggle as a student, have 1 gold medal. Now I am 2 yoe. Here's my story:\n\nI joined Kaggle as a student. I had an 8 GB laptop when I started. I was in situations where I couldn't run EDA scripts forget competitions. I bagged silver and also team gold medals.\n\nMy experiences were good and bad. \n\n**Yes, I got that gold in Code Competitions and without them, I would not probably have gold (in team) today**\n\nThe cost of building rig a Kaggle rig is $35,000 which is an extremly higher side and something many students, young professionals cannot afford. And with taxes and customs, just forget them in India. What say, students? Do some Indian users relate? \n\nCode Competitions level the playing field by a lot (that especially don't allow pre-run models). I have worked in a professional capacity where Kaggle was a part-time job and also where Kaggling wasn't allowed. It clearly makes it more feasible when you do Kaggle in your day job. Again, unfair but it's the game and how it works. \n\n> I believe having a paid job to do Kaggle is what leads one to better places + corporate GPUs boost chances further\n\nBut, nonetheless. There are few competitions that are really about feature engineering and not much raw power. And I believe with excellent engineering you can beat the x100 raw performances. I have seen it in practice and earned medals as well. \n\nHere are some takeaways I have from my Kaggling experiences;\n1. Kaggle is a learning ed-tech platform. May be unfair. Prioritize learning than earning medals\n2. Don't get too attached to earning medals. They are surely on your way when you have progressive learnings \n3. Having Kaggle Master/GM can get you a chance for an interview. But, not a necessity for professional career success\n4. I talked to recruiters from FAANG. Interviewed Kaggle Master. Kaggle is not a true representation of data science skill. \n\nIf you want to compete for earning medals;\n1. Target less data-based competitions and that are more creative\n2. Prefer Code Competitions over normal ones\n3. Engage on learnings and takeaways for your personal growth and career success\n\n#### Solution I propose\n\nMake competition spaces. Where Kagglers from their tier only to get compete in yearly seasons. \n\n- Experts vs. Experts\n- Masters vs. Masters\n\nEvery top 20% get a chance to compete in competitions above +1 their tier. So top users from Masters winning competition can compete now with GMs. This way it's a huge motivation to learn, grow in rankings have a fair chance to grab golds.\n\nAnd yes, the criterion to be Competitions GM of solo gold is bad. Its harder in a cutting edge CV competition today than some Weather prediction competition 6 years back? Apparently, early to Kaggle would have made you GM than today's competition? I wonder what current GMs think of this?\n\nAlso, a single gold should be equated to x10 silvers. Or medal to medal trade-in. \n\n> In this pandemic, I realised. Maybe it's more important to be in your home town, share more stories with family and have a better work-life balance than Kaggling all weekend.\n\nI do want to become GM one day. Maybe discussions/notebooks are feasible. But due to less access to hardware and lot of CV competitions, it's definitely on the harder side for me.\n\n(Edited x2)", "1310614": "Disagree. At least for this competition, LB<1.0 is not something that unlimited hardware resource is able to achieve. The interplay between data science/DL and *domain knowledge* is crucial to bring a giant leap in the LB score, compared to the pure training set-driven model. " } }