OK, so hear me out:-
\n\nI like image competitions. I had a lot of fun in my first one (Peking University). But seriously, so many image competitions consecutively? Why?
\n\nIs it because you want us to use more of your GPU resources or is it something else? I really don't want to be participating in all-image comps.
\n\nBefore Kaggle had good comps - tabular data that you could hop onto your laptop and start working on. Now it is handing us PhD-level problems on CV and not giving us enough GPU resources to work with it. (excluding TPUs because they normally give me pitiful accuracy)
\n\nWhy would Kaggle want to toss so many comps at us with CV? It is a huge loss for R users too.
\n\nJust take my question into account, and I await your reply.
", "messages": [ { "id": "865544", "postDate": "05/28/2020 17:32:55", "content": "OK, so hear me out:-
\n\nI like image competitions. I had a lot of fun in my first one (Peking University). But seriously, so many image competitions consecutively? Why?
\n\nIs it because you want us to use more of your GPU resources or is it something else? I really don't want to be participating in all-image comps.
\n\nBefore Kaggle had good comps - tabular data that you could hop onto your laptop and start working on. Now it is handing us PhD-level problems on CV and not giving us enough GPU resources to work with it. (excluding TPUs because they normally give me pitiful accuracy)
\n\nWhy would Kaggle want to toss so many comps at us with CV? It is a huge loss for R users too.
\n\nJust take my question into account, and I await your reply.
", "rawMarkdown": "OK, so hear me out:-\n\nI like image competitions. I had a lot of fun in my first one (Peking University). But seriously, so many image competitions consecutively? Why?\n\nIs it because you want us to use more of your GPU resources or is it something else? I really don't want to be participating in all-image comps. \n\nBefore Kaggle had good comps - tabular data that you could hop onto your laptop and start working on. Now it is handing us PhD-level problems on CV and not giving us enough GPU resources to work with it. (excluding TPUs because they normally give me pitiful accuracy)\n\nWhy would Kaggle want to toss so many comps at us with CV? It is a huge loss for R users too. \n\nJust take my question into account, and I await your reply.", "votes": null }, { "id": "865583", "postDate": "05/28/2020 18:01:50", "content": "I am assuming this has to do with google acquiring Kaggle and their interest in CV. I am with you on this, I do believe a good balance of competition types is needed. I for once want to see a good Recommender Systems competition. The last good one I know of (Netflix) yielded game-changing models such as Matrix Factorization but none since then.
", "rawMarkdown": "I am assuming this has to do with google acquiring Kaggle and their interest in CV. I am with you on this, I do believe a good balance of competition types is needed. I for once want to see a good Recommender Systems competition. The last good one I know of (Netflix) yielded game-changing models such as Matrix Factorization but none since then.", "votes": null }, { "id": "865593", "postDate": "05/28/2020 18:14:54", "content": "Believe me, if we could source more launch-viable non-image competitions, we would absolutely welcome them! We have nothing to gain from launching any particular kinds of competitions and, in fact, have more to lose by not having a more diverse portfolio of live competitions. We recognize that monotony in competition type is not ideal, but are really working with what we have in our pipeline at any given time.
\n\nTo help you appreciate what it takes to launch a competition -- our team thoroughly vets several leads every week. The vast majority of them are not viable for a variety of reasons (not enough data, not an actual problem, no budget from the host, leakage-vulnerable, public data, etc). Each lead can take several months to come to fruition as a launch because they undergo such extensive scrutiny. Our small team is frankly so focused on doing our best to get clean (leakage-free) competitions out to our community that being picky about diversity of data format has to take a back seat.
\n\nI know that, on the receiving end, these words amount to a lot of excuses, but the summary is that we hear you, we're doing our best, and there is no great CV conspiracy.
", "rawMarkdown": "Believe me, if we could source more launch-viable non-image competitions, we would absolutely welcome them! We have nothing to gain from launching any particular kinds of competitions and, in fact, have more to lose by not having a more diverse portfolio of live competitions. We recognize that monotony in competition type is not ideal, but are really working with what we have in our pipeline at any given time. \n\nTo help you appreciate what it takes to launch a competition -- our team thoroughly vets several leads every week. The vast majority of them are not viable for a variety of reasons (not enough data, not an actual problem, no budget from the host, leakage-vulnerable, public data, etc). Each lead can take several months to come to fruition as a launch because they undergo such extensive scrutiny. Our small team is frankly so focused on doing our best to get clean (leakage-free) competitions out to our community that being picky about diversity of data format has to take a back seat. \n\nI know that, on the receiving end, these words amount to a lot of excuses, but the summary is that we hear you, we're doing our best, and there is no great CV conspiracy.", "votes": null }, { "id": "865645", "postDate": "05/28/2020 19:07:58", "content": "Yes, you are right, there are more CV competitions nowadays, but there are lots of other types, as well:\n- Jigsaw multilingual toxic comments; [Text]\n- Tweet sentiment extraction; [Text]\n- M5 accuracy; [Time series, forecasting]\n- M5 uncertainty; [Time series, forecasting]\n- TREC Covid information retrieval; [Text, kudos]
\n\nRecently completed:\n- Liverpool ION Switching, [Signal data]\n- Abstraction and Reasoning Challenge; [AI]
\n\nI think Kaggle is an excellent platform!
", "rawMarkdown": "Yes, you are right, there are more CV competitions nowadays, but there are lots of other types, as well:\n- [Jigsaw multilingual toxic comments; [Text]](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification)\n- [Tweet sentiment extraction; [Text]](https://www.kaggle.com/c/tweet-sentiment-extraction)\n- [M5 accuracy; [Time series, forecasting]](https://www.kaggle.com/c/m5-forecasting-accuracy)\n- [M5 uncertainty; [Time series, forecasting]](https://www.kaggle.com/c/m5-forecasting-uncertainty)\n- [TREC Covid information retrieval; [Text, kudos]](https://www.kaggle.com/c/trec-covid-information-retrieval)\n\nRecently completed:\n- [Liverpool ION Switching, [Signal data]](https://www.kaggle.com/c/liverpool-ion-switching)\n- [Abstraction and Reasoning Challenge; [AI]](https://www.kaggle.com/c/abstraction-and-reasoning-challenge)\n\nI think Kaggle is an excellent platform!", "votes": null }, { "id": "865702", "postDate": "05/28/2020 20:22:33", "content": "Dear @nxrprime,
\n\nI wrote this post \"Medals for 'traditional' machine learning?\" a week ago along similar lines. I personally look forward to a good old-fashioned structured or tabular data competition!
\n\nAll the best,\ncarl
", "rawMarkdown": "Dear @nxrprime, \n\nI wrote this post [\"Medals for 'traditional' machine learning?\"](https://www.kaggle.com/general/152552) a week ago along similar lines. I personally look forward to a good old-fashioned structured or tabular data competition!\n\nAll the best,\ncarl", "votes": null }, { "id": "865797", "postDate": "05/28/2020 22:24:28", "content": "Thank you for the honest and thoughtful explanation. Much appreciated.
", "rawMarkdown": "Thank you for the honest and thoughtful explanation. Much appreciated.", "votes": null }, { "id": "865812", "postDate": "05/28/2020 22:47:48", "content": "There is some diversity inside CV competitions ^^
\n\nFor someone not familiar with them, Classifications like this one are easier to handle than Segmentations which, in turn, are easier to handle than Objects Detections. For instance , this one would be more challenging if you had to predict the bounding boxes containing skin lesions after implicitly classifying malignant
\n\nTPU is may be not mature yet, but it's great to handle big and very large dataset like the ongoing Jigsaw III competition or big resolution like here.
", "rawMarkdown": "There is some diversity inside CV competitions ^^\n\n For someone not familiar with them, Classifications like this one are easier to handle than Segmentations which, in turn, are easier to handle than Objects Detections. For instance , this one would be more challenging if you had to predict the bounding boxes containing skin lesions after implicitly classifying malignant\n\nTPU is may be not mature yet, but it's great to handle big and very large dataset like the ongoing Jigsaw III competition or big resolution like here.", "votes": null }, { "id": "867828", "postDate": "05/30/2020 16:49:23", "content": "AutoMLs are doing better on pure iid tabular data. I think it must be really hard to find a host who would be interested to increase the accuracy of the order of 10^-2.
\n\nI would really like competitions like
\n\nTrackML, PLAsTiCC.
\n\nAnother thing the data of past competitions are there only and one can participate in them as well. I do it every now and then. That is best thing I like about kaggle
", "rawMarkdown": "AutoMLs are doing better on pure iid tabular data. I think it must be really hard to find a host who would be interested to increase the accuracy of the order of 10^-2.\n\nI would really like competitions like\n\nTrackML, PLAsTiCC.\n\nAnother thing the data of past competitions are there only and one can participate in them as well. I do it every now and then. That is best thing I like about kaggle", "votes": null }, { "id": "867918", "postDate": "05/30/2020 18:27:00", "content": "I would also like more competitions mixing tabular, Text and eventually Images like Avito and Mercari
\n\nThere is also a special images competition I envoyed : QuickDraw. ...LSTM was almost as competitive as Resnet. I think Transfomers would do great job in competition llke this one.
", "rawMarkdown": "I would also like more competitions mixing tabular, Text and eventually Images like Avito and Mercari \n\nThere is also a special images competition I envoyed : QuickDraw. ...LSTM was almost as competitive as Resnet. I think Transfomers would do great job in competition llke this one.", "votes": null }, { "id": "881160", "postDate": "06/10/2020 19:26:50", "content": "Kind of an off topic questions related to images, does anyone use R Studio for image processing? I have a simple questions about taking a picture of dots and turning it into a 3d model.
", "rawMarkdown": "Kind of an off topic questions related to images, does anyone use R Studio for image processing? I have a simple questions about taking a picture of dots and turning it into a 3d model.", "votes": null }, { "id": "881189", "postDate": "06/10/2020 19:54:24", "content": "I would argue that though classification is definitely a less tricky task compared to segmentation/detection, this simplicity attracts a lot of people which eventually makes the task much harder :)
", "rawMarkdown": "I would argue that though classification is definitely a less tricky task compared to segmentation/detection, this simplicity attracts a lot of people which eventually makes the task much harder :)", "votes": null }, { "id": "897097", "postDate": "06/22/2020 16:04:39", "content": "This is true :)
", "rawMarkdown": "This is true :)", "votes": null } ], "comments": [ { "id": 865583, "author_name": "samikh", "author_url": "", "post_date": "05/28/2020 18:01:50", "content": "I am assuming this has to do with google acquiring Kaggle and their interest in CV. I am with you on this, I do believe a good balance of competition types is needed. I for once want to see a good Recommender Systems competition. The last good one I know of (Netflix) yielded game-changing models such as Matrix Factorization but none since then.
", "votes": null, "replies": [] }, { "id": 865593, "author_name": "juliaelliott", "author_url": "", "post_date": "05/28/2020 18:14:54", "content": "Believe me, if we could source more launch-viable non-image competitions, we would absolutely welcome them! We have nothing to gain from launching any particular kinds of competitions and, in fact, have more to lose by not having a more diverse portfolio of live competitions. We recognize that monotony in competition type is not ideal, but are really working with what we have in our pipeline at any given time.
\n\nTo help you appreciate what it takes to launch a competition -- our team thoroughly vets several leads every week. The vast majority of them are not viable for a variety of reasons (not enough data, not an actual problem, no budget from the host, leakage-vulnerable, public data, etc). Each lead can take several months to come to fruition as a launch because they undergo such extensive scrutiny. Our small team is frankly so focused on doing our best to get clean (leakage-free) competitions out to our community that being picky about diversity of data format has to take a back seat.
\n\nI know that, on the receiving end, these words amount to a lot of excuses, but the summary is that we hear you, we're doing our best, and there is no great CV conspiracy.
", "votes": null, "replies": [ { "id": 865797, "author_name": "zaharch", "author_url": "", "post_date": "05/28/2020 22:24:28", "content": "Thank you for the honest and thoughtful explanation. Much appreciated.
", "votes": null, "replies": [] } ] }, { "id": 865645, "author_name": "pestipeti", "author_url": "", "post_date": "05/28/2020 19:07:58", "content": "Yes, you are right, there are more CV competitions nowadays, but there are lots of other types, as well:\n- Jigsaw multilingual toxic comments; [Text]\n- Tweet sentiment extraction; [Text]\n- M5 accuracy; [Time series, forecasting]\n- M5 uncertainty; [Time series, forecasting]\n- TREC Covid information retrieval; [Text, kudos]
\n\nRecently completed:\n- Liverpool ION Switching, [Signal data]\n- Abstraction and Reasoning Challenge; [AI]
\n\nI think Kaggle is an excellent platform!
", "votes": null, "replies": [] }, { "id": 865702, "author_name": "carlmcbrideellis", "author_url": "", "post_date": "05/28/2020 20:22:33", "content": "Dear @nxrprime,
\n\nI wrote this post \"Medals for 'traditional' machine learning?\" a week ago along similar lines. I personally look forward to a good old-fashioned structured or tabular data competition!
\n\nAll the best,\ncarl
", "votes": null, "replies": [] }, { "id": 865812, "author_name": "serigne", "author_url": "", "post_date": "05/28/2020 22:47:48", "content": "There is some diversity inside CV competitions ^^
\n\nFor someone not familiar with them, Classifications like this one are easier to handle than Segmentations which, in turn, are easier to handle than Objects Detections. For instance , this one would be more challenging if you had to predict the bounding boxes containing skin lesions after implicitly classifying malignant
\n\nTPU is may be not mature yet, but it's great to handle big and very large dataset like the ongoing Jigsaw III competition or big resolution like here.
", "votes": null, "replies": [ { "id": 881189, "author_name": "ddanevskyi", "author_url": "", "post_date": "06/10/2020 19:54:24", "content": "I would argue that though classification is definitely a less tricky task compared to segmentation/detection, this simplicity attracts a lot of people which eventually makes the task much harder :)
", "votes": null, "replies": [] }, { "id": 897097, "author_name": "serigne", "author_url": "", "post_date": "06/22/2020 16:04:39", "content": "This is true :)
", "votes": null, "replies": [] } ] }, { "id": 867828, "author_name": "mks2192", "author_url": "", "post_date": "05/30/2020 16:49:23", "content": "AutoMLs are doing better on pure iid tabular data. I think it must be really hard to find a host who would be interested to increase the accuracy of the order of 10^-2.
\n\nI would really like competitions like
\n\nTrackML, PLAsTiCC.
\n\nAnother thing the data of past competitions are there only and one can participate in them as well. I do it every now and then. That is best thing I like about kaggle
", "votes": null, "replies": [ { "id": 867918, "author_name": "serigne", "author_url": "", "post_date": "05/30/2020 18:27:00", "content": "I would also like more competitions mixing tabular, Text and eventually Images like Avito and Mercari
\n\nThere is also a special images competition I envoyed : QuickDraw. ...LSTM was almost as competitive as Resnet. I think Transfomers would do great job in competition llke this one.
", "votes": null, "replies": [] } ] }, { "id": 881160, "author_name": "larsen0966", "author_url": "", "post_date": "06/10/2020 19:26:50", "content": "Kind of an off topic questions related to images, does anyone use R Studio for image processing? I have a simple questions about taking a picture of dots and turning it into a 3d model.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "865544": "OK, so hear me out:-\n\nI like image competitions. I had a lot of fun in my first one (Peking University). But seriously, so many image competitions consecutively? Why?\n\nIs it because you want us to use more of your GPU resources or is it something else? I really don't want to be participating in all-image comps. \n\nBefore Kaggle had good comps - tabular data that you could hop onto your laptop and start working on. Now it is handing us PhD-level problems on CV and not giving us enough GPU resources to work with it. (excluding TPUs because they normally give me pitiful accuracy)\n\nWhy would Kaggle want to toss so many comps at us with CV? It is a huge loss for R users too. \n\nJust take my question into account, and I await your reply.", "865583": "I am assuming this has to do with google acquiring Kaggle and their interest in CV. I am with you on this, I do believe a good balance of competition types is needed. I for once want to see a good Recommender Systems competition. The last good one I know of (Netflix) yielded game-changing models such as Matrix Factorization but none since then.", "865593": "Believe me, if we could source more launch-viable non-image competitions, we would absolutely welcome them! We have nothing to gain from launching any particular kinds of competitions and, in fact, have more to lose by not having a more diverse portfolio of live competitions. We recognize that monotony in competition type is not ideal, but are really working with what we have in our pipeline at any given time. \n\nTo help you appreciate what it takes to launch a competition -- our team thoroughly vets several leads every week. The vast majority of them are not viable for a variety of reasons (not enough data, not an actual problem, no budget from the host, leakage-vulnerable, public data, etc). Each lead can take several months to come to fruition as a launch because they undergo such extensive scrutiny. Our small team is frankly so focused on doing our best to get clean (leakage-free) competitions out to our community that being picky about diversity of data format has to take a back seat. \n\nI know that, on the receiving end, these words amount to a lot of excuses, but the summary is that we hear you, we're doing our best, and there is no great CV conspiracy.", "865645": "Yes, you are right, there are more CV competitions nowadays, but there are lots of other types, as well:\n- [Jigsaw multilingual toxic comments; [Text]](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification)\n- [Tweet sentiment extraction; [Text]](https://www.kaggle.com/c/tweet-sentiment-extraction)\n- [M5 accuracy; [Time series, forecasting]](https://www.kaggle.com/c/m5-forecasting-accuracy)\n- [M5 uncertainty; [Time series, forecasting]](https://www.kaggle.com/c/m5-forecasting-uncertainty)\n- [TREC Covid information retrieval; [Text, kudos]](https://www.kaggle.com/c/trec-covid-information-retrieval)\n\nRecently completed:\n- [Liverpool ION Switching, [Signal data]](https://www.kaggle.com/c/liverpool-ion-switching)\n- [Abstraction and Reasoning Challenge; [AI]](https://www.kaggle.com/c/abstraction-and-reasoning-challenge)\n\nI think Kaggle is an excellent platform!", "865702": "Dear @nxrprime, \n\nI wrote this post [\"Medals for 'traditional' machine learning?\"](https://www.kaggle.com/general/152552) a week ago along similar lines. I personally look forward to a good old-fashioned structured or tabular data competition!\n\nAll the best,\ncarl", "865797": "Thank you for the honest and thoughtful explanation. Much appreciated.", "865812": "There is some diversity inside CV competitions ^^\n\n For someone not familiar with them, Classifications like this one are easier to handle than Segmentations which, in turn, are easier to handle than Objects Detections. For instance , this one would be more challenging if you had to predict the bounding boxes containing skin lesions after implicitly classifying malignant\n\nTPU is may be not mature yet, but it's great to handle big and very large dataset like the ongoing Jigsaw III competition or big resolution like here.", "867828": "AutoMLs are doing better on pure iid tabular data. I think it must be really hard to find a host who would be interested to increase the accuracy of the order of 10^-2.\n\nI would really like competitions like\n\nTrackML, PLAsTiCC.\n\nAnother thing the data of past competitions are there only and one can participate in them as well. I do it every now and then. That is best thing I like about kaggle", "867918": "I would also like more competitions mixing tabular, Text and eventually Images like Avito and Mercari \n\nThere is also a special images competition I envoyed : QuickDraw. ...LSTM was almost as competitive as Resnet. I think Transfomers would do great job in competition llke this one.", "881160": "Kind of an off topic questions related to images, does anyone use R Studio for image processing? I have a simple questions about taking a picture of dots and turning it into a 3d model.", "881189": "I would argue that though classification is definitely a less tricky task compared to segmentation/detection, this simplicity attracts a lot of people which eventually makes the task much harder :)", "897097": "This is true :)" }, "source": "meta" }