Very good thing http://www.ccom.ucsd.edu/~cdeotte/programs/neuralnetwork.html.\nGet more interesting findings about neural network training.
", "messages": [ { "id": "944712", "postDate": "07/25/2020 09:30:42", "content": "Very good thing http://www.ccom.ucsd.edu/~cdeotte/programs/neuralnetwork.html.\nGet more interesting findings about neural network training.
", "rawMarkdown": "Very good thing http://www.ccom.ucsd.edu/~cdeotte/programs/neuralnetwork.html.\nGet more interesting findings about neural network training.", "votes": null }, { "id": "945378", "postDate": "07/25/2020 19:09:11", "content": "Thanks for sharing! I don't get why so many downvotes. Here's two other playgrouds that you might enjoy:\n- https://playground.tensorflow.org/\n- https://www.cs.ryerson.ca/~aharley/vis/conv/
", "rawMarkdown": "Thanks for sharing! I don't get why so many downvotes. Here's two other playgrouds that you might enjoy:\n- https://playground.tensorflow.org/\n- https://www.cs.ryerson.ca/~aharley/vis/conv/", "votes": null }, { "id": "946469", "postDate": "07/26/2020 15:46:07", "content": "Here is my own playground:\nhttps://rcijov.github.io/tf_playground/
\n\nNo downvotes :) share the love, not the hate!
", "rawMarkdown": "Here is my own playground:\n[https://rcijov.github.io/tf_playground/](https://rcijov.github.io/tf_playground/)\n\nNo downvotes :) share the love, not the hate!", "votes": null }, { "id": "946507", "postDate": "07/26/2020 16:07:19", "content": "Awesome. Yours is nice.
", "rawMarkdown": "Awesome. Yours is nice.", "votes": null }, { "id": "946512", "postDate": "07/26/2020 16:12:48", "content": "Hi Zavod. Thanks for posting a link to my neural network playground. Watching neural networks train has increased my intuition and made me a better NN architect.
\n\nThat website is my simple one. I have a private one that is more complex. It allows me to build different architectures (more layers etc) with different losses and activations. Also, you can interact with the learning process to encourage or discourage the patterns that it learns. Playing with the complex one has really helped my understanding.
", "rawMarkdown": "Hi Zavod. Thanks for posting a link to my neural network playground. Watching neural networks train has increased my intuition and made me a better NN architect.\n\nThat website is my simple one. I have a private one that is more complex. It allows me to build different architectures (more layers etc) with different losses and activations. Also, you can interact with the learning process to encourage or discourage the patterns that it learns. Playing with the complex one has really helped my understanding.", "votes": null }, { "id": "946518", "postDate": "07/26/2020 16:18:00", "content": "I published a playground for Naive Bayes, Logistic Regression, Tree, kNN, and SVM here. And i published a webpage where you can draw a digit (0,1,2,3 etc) and have a NN classify it here
\n\nBefore joining Kaggle, I spent months writing every algorithm from scratch in C and JavaScript. And studying all the mathematics. This has helped my intuition for all ML models.
", "rawMarkdown": "I published a playground for Naive Bayes, Logistic Regression, Tree, kNN, and SVM [here][1]. And i published a webpage where you can draw a digit (0,1,2,3 etc) and have a NN classify it [here][2]\n\nBefore joining Kaggle, I spent months writing every algorithm from scratch in C and JavaScript. And studying all the mathematics. This has helped my intuition for all ML models.\n\n[1]: http://www.ccom.ucsd.edu/~cdeotte/programs/classify.html\n[2]: http://www.ccom.ucsd.edu/~cdeotte/programs/MNIST.html", "votes": null }, { "id": "946519", "postDate": "07/26/2020 16:18:27", "content": "I make about different 15 insights with this model, and the most important is that resulting network is highly dependent on initial phase of learning
", "rawMarkdown": "I make about different 15 insights with this model, and the most important is that resulting network is highly dependent on initial phase of learning", "votes": null }, { "id": "946535", "postDate": "07/26/2020 16:27:28", "content": "Maybe i can publish the 2 hidden layer model soon. My public website is currently only 1 hidden layer. When you add a second hidden layer, it significantly changes the learning process. You would be surprised how much more \"fluid\" and \"adaptable\" the network becomes. It bends and flexs and find the solution much faster.
\n\nI noticed that 3 or more layers doesn't add too much more. But 1 hidden layer versus 2 hidden layer is big difference.
", "rawMarkdown": "Maybe i can publish the 2 hidden layer model soon. My public website is currently only 1 hidden layer. When you add a second hidden layer, it significantly changes the learning process. You would be surprised how much more \"fluid\" and \"adaptable\" the network becomes. It bends and flexs and find the solution much faster.\n\nI noticed that 3 or more layers doesn't add too much more. But 1 hidden layer versus 2 hidden layer is big difference.", "votes": null }, { "id": "946541", "postDate": "07/26/2020 16:30:50", "content": "And the it's fun to watch the different activations. Each activation has a \"personality\". Relu, Sigmoid, and TanH all behave differently. They learn different patterns differently and their final regions of classification are different.
\n\nIf i watch a NN learn, i can tell you what activation the network has. It is that significant. I need to update my webpage with different activations too. It's very instructive to watch the different activations learn.
", "rawMarkdown": "And the it's fun to watch the different activations. Each activation has a \"personality\". Relu, Sigmoid, and TanH all behave differently. They learn different patterns differently and their final regions of classification are different. \n\nIf i watch a NN learn, i can tell you what activation the network has. It is that significant. I need to update my webpage with different activations too. It's very instructive to watch the different activations learn.", "votes": null }, { "id": "946568", "postDate": "07/26/2020 16:45:42", "content": "Great recommendation!
", "rawMarkdown": "Great recommendation!", "votes": null } ], "comments": [ { "id": 945378, "author_name": "wittmannf", "author_url": "", "post_date": "07/25/2020 19:09:11", "content": "Thanks for sharing! I don't get why so many downvotes. Here's two other playgrouds that you might enjoy:\n- https://playground.tensorflow.org/\n- https://www.cs.ryerson.ca/~aharley/vis/conv/
", "votes": null, "replies": [] }, { "id": 946469, "author_name": "alincijov", "author_url": "", "post_date": "07/26/2020 15:46:07", "content": "Here is my own playground:\nhttps://rcijov.github.io/tf_playground/
\n\nNo downvotes :) share the love, not the hate!
", "votes": null, "replies": [ { "id": 946507, "author_name": "cdeotte", "author_url": "", "post_date": "07/26/2020 16:07:19", "content": "Awesome. Yours is nice.
", "votes": null, "replies": [] } ] }, { "id": 946512, "author_name": "cdeotte", "author_url": "", "post_date": "07/26/2020 16:12:48", "content": "Hi Zavod. Thanks for posting a link to my neural network playground. Watching neural networks train has increased my intuition and made me a better NN architect.
\n\nThat website is my simple one. I have a private one that is more complex. It allows me to build different architectures (more layers etc) with different losses and activations. Also, you can interact with the learning process to encourage or discourage the patterns that it learns. Playing with the complex one has really helped my understanding.
", "votes": null, "replies": [ { "id": 946518, "author_name": "cdeotte", "author_url": "", "post_date": "07/26/2020 16:18:00", "content": "I published a playground for Naive Bayes, Logistic Regression, Tree, kNN, and SVM here. And i published a webpage where you can draw a digit (0,1,2,3 etc) and have a NN classify it here
\n\nBefore joining Kaggle, I spent months writing every algorithm from scratch in C and JavaScript. And studying all the mathematics. This has helped my intuition for all ML models.
", "votes": null, "replies": [] }, { "id": 946519, "author_name": "zavodrobotov", "author_url": "", "post_date": "07/26/2020 16:18:27", "content": "I make about different 15 insights with this model, and the most important is that resulting network is highly dependent on initial phase of learning
", "votes": null, "replies": [] }, { "id": 946535, "author_name": "cdeotte", "author_url": "", "post_date": "07/26/2020 16:27:28", "content": "Maybe i can publish the 2 hidden layer model soon. My public website is currently only 1 hidden layer. When you add a second hidden layer, it significantly changes the learning process. You would be surprised how much more \"fluid\" and \"adaptable\" the network becomes. It bends and flexs and find the solution much faster.
\n\nI noticed that 3 or more layers doesn't add too much more. But 1 hidden layer versus 2 hidden layer is big difference.
", "votes": null, "replies": [] }, { "id": 946541, "author_name": "cdeotte", "author_url": "", "post_date": "07/26/2020 16:30:50", "content": "And the it's fun to watch the different activations. Each activation has a \"personality\". Relu, Sigmoid, and TanH all behave differently. They learn different patterns differently and their final regions of classification are different.
\n\nIf i watch a NN learn, i can tell you what activation the network has. It is that significant. I need to update my webpage with different activations too. It's very instructive to watch the different activations learn.
", "votes": null, "replies": [] } ] }, { "id": 946568, "author_name": "hugoherrera11", "author_url": "", "post_date": "07/26/2020 16:45:42", "content": "Great recommendation!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "944712": "Very good thing http://www.ccom.ucsd.edu/~cdeotte/programs/neuralnetwork.html.\nGet more interesting findings about neural network training.", "945378": "Thanks for sharing! I don't get why so many downvotes. Here's two other playgrouds that you might enjoy:\n- https://playground.tensorflow.org/\n- https://www.cs.ryerson.ca/~aharley/vis/conv/", "946469": "Here is my own playground:\n[https://rcijov.github.io/tf_playground/](https://rcijov.github.io/tf_playground/)\n\nNo downvotes :) share the love, not the hate!", "946507": "Awesome. Yours is nice.", "946512": "Hi Zavod. Thanks for posting a link to my neural network playground. Watching neural networks train has increased my intuition and made me a better NN architect.\n\nThat website is my simple one. I have a private one that is more complex. It allows me to build different architectures (more layers etc) with different losses and activations. Also, you can interact with the learning process to encourage or discourage the patterns that it learns. Playing with the complex one has really helped my understanding.", "946518": "I published a playground for Naive Bayes, Logistic Regression, Tree, kNN, and SVM [here][1]. And i published a webpage where you can draw a digit (0,1,2,3 etc) and have a NN classify it [here][2]\n\nBefore joining Kaggle, I spent months writing every algorithm from scratch in C and JavaScript. And studying all the mathematics. This has helped my intuition for all ML models.\n\n[1]: http://www.ccom.ucsd.edu/~cdeotte/programs/classify.html\n[2]: http://www.ccom.ucsd.edu/~cdeotte/programs/MNIST.html", "946519": "I make about different 15 insights with this model, and the most important is that resulting network is highly dependent on initial phase of learning", "946535": "Maybe i can publish the 2 hidden layer model soon. My public website is currently only 1 hidden layer. When you add a second hidden layer, it significantly changes the learning process. You would be surprised how much more \"fluid\" and \"adaptable\" the network becomes. It bends and flexs and find the solution much faster.\n\nI noticed that 3 or more layers doesn't add too much more. But 1 hidden layer versus 2 hidden layer is big difference.", "946541": "And the it's fun to watch the different activations. Each activation has a \"personality\". Relu, Sigmoid, and TanH all behave differently. They learn different patterns differently and their final regions of classification are different. \n\nIf i watch a NN learn, i can tell you what activation the network has. It is that significant. I need to update my webpage with different activations too. It's very instructive to watch the different activations learn.", "946568": "Great recommendation!" }, "source": "meta" }