Toward the close of this competition, I reflect on the importance of generalizability in machine learning. Humans have an innate ability to generalize; do machines hold the same potential?
\n\nWhile it's valuable to solve specific problems using ML, it will be interesting to see how well it can generalize in coming years. I imagine a day where machines regularly use knowledge from a given domain to solve problems in an entirely new domain. This could include advancements in finding cures for diseases, creating new styles of music, or interacting with humans through natural language.
\n\nMany of us were surprised by the new image classes in stage 2 even after being warned. However, of all the images that could have possibly shown up in stage 2, they were actually quite similar to the images in stage 1. After all, both sets of images contained nuclei (for the most part). Can you imagine everyone's reaction if the stage 2 images were pictures of cats, dogs, or Donald Trump? Humans can recognize these things, so why not machines?
\n\nThis has been a wonderful competition! Many thanks to the competition organizers for stimulating thought and conversation about generalizability.
", "messages": [ { "id": "314085", "postDate": "04/14/2018 15:15:08", "content": "Toward the close of this competition, I reflect on the importance of generalizability in machine learning. Humans have an innate ability to generalize; do machines hold the same potential?
\n\nWhile it's valuable to solve specific problems using ML, it will be interesting to see how well it can generalize in coming years. I imagine a day where machines regularly use knowledge from a given domain to solve problems in an entirely new domain. This could include advancements in finding cures for diseases, creating new styles of music, or interacting with humans through natural language.
\n\nMany of us were surprised by the new image classes in stage 2 even after being warned. However, of all the images that could have possibly shown up in stage 2, they were actually quite similar to the images in stage 1. After all, both sets of images contained nuclei (for the most part). Can you imagine everyone's reaction if the stage 2 images were pictures of cats, dogs, or Donald Trump? Humans can recognize these things, so why not machines?
\n\nThis has been a wonderful competition! Many thanks to the competition organizers for stimulating thought and conversation about generalizability.
", "rawMarkdown": "Toward the close of this competition, I reflect on the importance of generalizability in machine learning. Humans have an innate ability to generalize; do machines hold the same potential? \n\nWhile it's valuable to solve specific problems using ML, it will be interesting to see how well it can generalize in coming years. I imagine a day where machines regularly use knowledge from a given domain to solve problems in an entirely new domain. This could include advancements in finding cures for diseases, creating new styles of music, or interacting with humans through natural language.\n\nMany of us were surprised by the new image classes in stage 2 even after being warned. However, of all the images that could have possibly shown up in stage 2, they were actually quite similar to the images in stage 1. After all, both sets of images contained nuclei (for the most part). Can you imagine everyone's reaction if the stage 2 images were pictures of cats, dogs, or Donald Trump? Humans can recognize these things, so why not machines? \n\nThis has been a wonderful competition! Many thanks to the competition organizers for stimulating thought and conversation about generalizability.", "votes": null }, { "id": "314205", "postDate": "04/14/2018 21:10:54", "content": "Those are some important points you made, it is very nice that a Kaggle competition provoked such a high level discussion about intelligence.
\n\nAs for my own view, I am a bit skeptical that \"humans have an innate ability to generalize\". I still remember the \"proof\" of it made by Gary Marcus in his post https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1 - check point 9. The essence of the \"proof\" is that if we, as humans, have the same opinion about something, then we must all be correct, because hey, it's us who define the \"ground truth\".
", "rawMarkdown": "Those are some important points you made, it is very nice that a Kaggle competition provoked such a high level discussion about intelligence. \n\nAs for my own view, I am a bit skeptical that \"humans have an innate ability to generalize\". I still remember the \"proof\" of it made by Gary Marcus in his post https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1 - check point 9. The essence of the \"proof\" is that if we, as humans, have the same opinion about something, then we must all be correct, because hey, it's us who define the \"ground truth\".", "votes": null }, { "id": "314209", "postDate": "04/14/2018 21:31:40", "content": "@bobutis, those are great thoughts. Keep them coming!
\n\nWhile humans cannot generalize in every situation, there are several examples of where they do generalize. For example, most people that burn their hand on a hot stove will learn to avoid touching anything that is extremely hot (e.g., fire). What are your thoughts?
", "rawMarkdown": "bobutis, those are great thoughts. Keep them coming!\n\nWhile humans cannot generalize in *every* situation, there are several examples of where they do generalize. For example, most people that burn their hand on a hot stove will learn to avoid touching anything that is extremely hot (e.g., fire). What are your thoughts?", "votes": null }, { "id": "314212", "postDate": "04/14/2018 21:43:55", "content": "What do we \"learn\" exactly in that example? We do avoid contact with fire and pain in general, even if we did not burn our hands on a stove.
", "rawMarkdown": "What do we \"learn\" exactly in that example? We do avoid contact with fire and pain in general, even if we did not burn our hands on a stove.", "votes": null }, { "id": "314218", "postDate": "04/14/2018 22:01:31", "content": "We experience the stove causing pain, and we learn (or generalize) that touching anything extremely hot causes pain. We could have learned this in a variety of ways, and touching the hot stove is just one specific example.
\n\nYou mentioned that we avoid contact with fire and pain in general. This does not hold true for a newborn baby, because they have not learned yet that heat causes pain. The baby needs to learn through experience (e.g., touching something hot or being told by a parent) that extreme heat causes pain.
", "rawMarkdown": "We experience the stove causing pain, and we learn (or generalize) that touching anything extremely hot causes pain. We could have learned this in a variety of ways, and touching the hot stove is just one specific example.\n\nYou mentioned that we avoid contact with fire and pain in general. This does not hold true for a newborn baby, because they have not learned yet that heat causes pain. The baby needs to learn through experience (e.g., touching something hot or being told by a parent) that extreme heat causes pain.", "votes": null }, { "id": "314220", "postDate": "04/14/2018 22:28:17", "content": "But this \"learning\" is a bit intertwined with definitions that we choose ourselves. You could argue that it all boils down to physics which describes reality \"as it is\", but that's an assumption that's rarely ever explicitly mentioned. One related article I liked a lot is https://www.nature.com/news/physics-qbism-puts-the-scientist-back-into-science-1.14912
", "rawMarkdown": "But this \"learning\" is a bit intertwined with definitions that we choose ourselves. You could argue that it all boils down to physics which describes reality \"as it is\", but that's an assumption that's rarely ever explicitly mentioned. One related article I liked a lot is https://www.nature.com/news/physics-qbism-puts-the-scientist-back-into-science-1.14912", "votes": null }, { "id": "314224", "postDate": "04/14/2018 22:43:27", "content": "That's true that the type of learning I describe certainly relies on definitions and axioms we choose ourselves.
\n\nThanks for sharing the article!
", "rawMarkdown": "That's true that the type of learning I describe certainly relies on definitions and axioms we choose ourselves. \n\nThanks for sharing the article!", "votes": null }, { "id": "315208", "postDate": "04/16/2018 21:57:43", "content": "The generalizability in physics I think is highly related to what's the fundamental theorems being discovered. I feel that for this competition it will be interesting if a model can discover something more fundamental of whether an object is a nuclei or not. For example, there are some white background and black nuclei images, only the black nuclei is being labeled. And there are also some blue nuclei images, some of them has smaller black kernels, but only the outer blue object being labeled. For an non-expert like me it is somewhat confused, so I think there may be some hidden rules about recognizing nuclei, and it will be interesting machines can discover that automatically.
", "rawMarkdown": "The generalizability in physics I think is highly related to what's the fundamental theorems being discovered. I feel that for this competition it will be interesting if a model can discover something more fundamental of whether an object is a nuclei or not. For example, there are some white background and black nuclei images, only the black nuclei is being labeled. And there are also some blue nuclei images, some of them has smaller black kernels, but only the outer blue object being labeled. For an non-expert like me it is somewhat confused, so I think there may be some hidden rules about recognizing nuclei, and it will be interesting machines can discover that automatically.", "votes": null } ], "comments": [ { "id": 314205, "author_name": "bobutis", "author_url": "", "post_date": "04/14/2018 21:10:54", "content": "Those are some important points you made, it is very nice that a Kaggle competition provoked such a high level discussion about intelligence.
\n\nAs for my own view, I am a bit skeptical that \"humans have an innate ability to generalize\". I still remember the \"proof\" of it made by Gary Marcus in his post https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1 - check point 9. The essence of the \"proof\" is that if we, as humans, have the same opinion about something, then we must all be correct, because hey, it's us who define the \"ground truth\".
", "votes": null, "replies": [ { "id": 314209, "author_name": "danieljohnson", "author_url": "", "post_date": "04/14/2018 21:31:40", "content": "@bobutis, those are great thoughts. Keep them coming!
\n\nWhile humans cannot generalize in every situation, there are several examples of where they do generalize. For example, most people that burn their hand on a hot stove will learn to avoid touching anything that is extremely hot (e.g., fire). What are your thoughts?
", "votes": null, "replies": [] }, { "id": 314212, "author_name": "bobutis", "author_url": "", "post_date": "04/14/2018 21:43:55", "content": "What do we \"learn\" exactly in that example? We do avoid contact with fire and pain in general, even if we did not burn our hands on a stove.
", "votes": null, "replies": [] }, { "id": 314218, "author_name": "danieljohnson", "author_url": "", "post_date": "04/14/2018 22:01:31", "content": "We experience the stove causing pain, and we learn (or generalize) that touching anything extremely hot causes pain. We could have learned this in a variety of ways, and touching the hot stove is just one specific example.
\n\nYou mentioned that we avoid contact with fire and pain in general. This does not hold true for a newborn baby, because they have not learned yet that heat causes pain. The baby needs to learn through experience (e.g., touching something hot or being told by a parent) that extreme heat causes pain.
", "votes": null, "replies": [] }, { "id": 314220, "author_name": "bobutis", "author_url": "", "post_date": "04/14/2018 22:28:17", "content": "But this \"learning\" is a bit intertwined with definitions that we choose ourselves. You could argue that it all boils down to physics which describes reality \"as it is\", but that's an assumption that's rarely ever explicitly mentioned. One related article I liked a lot is https://www.nature.com/news/physics-qbism-puts-the-scientist-back-into-science-1.14912
", "votes": null, "replies": [] }, { "id": 314224, "author_name": "danieljohnson", "author_url": "", "post_date": "04/14/2018 22:43:27", "content": "That's true that the type of learning I describe certainly relies on definitions and axioms we choose ourselves.
\n\nThanks for sharing the article!
", "votes": null, "replies": [] } ] }, { "id": 315208, "author_name": "liyang2034", "author_url": "", "post_date": "04/16/2018 21:57:43", "content": "The generalizability in physics I think is highly related to what's the fundamental theorems being discovered. I feel that for this competition it will be interesting if a model can discover something more fundamental of whether an object is a nuclei or not. For example, there are some white background and black nuclei images, only the black nuclei is being labeled. And there are also some blue nuclei images, some of them has smaller black kernels, but only the outer blue object being labeled. For an non-expert like me it is somewhat confused, so I think there may be some hidden rules about recognizing nuclei, and it will be interesting machines can discover that automatically.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "314085": "Toward the close of this competition, I reflect on the importance of generalizability in machine learning. Humans have an innate ability to generalize; do machines hold the same potential? \n\nWhile it's valuable to solve specific problems using ML, it will be interesting to see how well it can generalize in coming years. I imagine a day where machines regularly use knowledge from a given domain to solve problems in an entirely new domain. This could include advancements in finding cures for diseases, creating new styles of music, or interacting with humans through natural language.\n\nMany of us were surprised by the new image classes in stage 2 even after being warned. However, of all the images that could have possibly shown up in stage 2, they were actually quite similar to the images in stage 1. After all, both sets of images contained nuclei (for the most part). Can you imagine everyone's reaction if the stage 2 images were pictures of cats, dogs, or Donald Trump? Humans can recognize these things, so why not machines? \n\nThis has been a wonderful competition! Many thanks to the competition organizers for stimulating thought and conversation about generalizability.", "314205": "Those are some important points you made, it is very nice that a Kaggle competition provoked such a high level discussion about intelligence. \n\nAs for my own view, I am a bit skeptical that \"humans have an innate ability to generalize\". I still remember the \"proof\" of it made by Gary Marcus in his post https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning-6e8bfd5ae0f1 - check point 9. The essence of the \"proof\" is that if we, as humans, have the same opinion about something, then we must all be correct, because hey, it's us who define the \"ground truth\".", "314209": "bobutis, those are great thoughts. Keep them coming!\n\nWhile humans cannot generalize in *every* situation, there are several examples of where they do generalize. For example, most people that burn their hand on a hot stove will learn to avoid touching anything that is extremely hot (e.g., fire). What are your thoughts?", "314212": "What do we \"learn\" exactly in that example? We do avoid contact with fire and pain in general, even if we did not burn our hands on a stove.", "314218": "We experience the stove causing pain, and we learn (or generalize) that touching anything extremely hot causes pain. We could have learned this in a variety of ways, and touching the hot stove is just one specific example.\n\nYou mentioned that we avoid contact with fire and pain in general. This does not hold true for a newborn baby, because they have not learned yet that heat causes pain. The baby needs to learn through experience (e.g., touching something hot or being told by a parent) that extreme heat causes pain.", "314220": "But this \"learning\" is a bit intertwined with definitions that we choose ourselves. You could argue that it all boils down to physics which describes reality \"as it is\", but that's an assumption that's rarely ever explicitly mentioned. One related article I liked a lot is https://www.nature.com/news/physics-qbism-puts-the-scientist-back-into-science-1.14912", "314224": "That's true that the type of learning I describe certainly relies on definitions and axioms we choose ourselves. \n\nThanks for sharing the article!", "315208": "The generalizability in physics I think is highly related to what's the fundamental theorems being discovered. I feel that for this competition it will be interesting if a model can discover something more fundamental of whether an object is a nuclei or not. For example, there are some white background and black nuclei images, only the black nuclei is being labeled. And there are also some blue nuclei images, some of them has smaller black kernels, but only the outer blue object being labeled. For an non-expert like me it is somewhat confused, so I think there may be some hidden rules about recognizing nuclei, and it will be interesting machines can discover that automatically." }, "source": "meta" }