I can sense that this may be coming in a year or two. Afterall, google NASNET is an example \n that machined design network is as good as hand-crafted network.
\n\nhttps://research.googleblog.com/2017/11/automl-for-large-scale-image.html\nhttps://blog.acolyer.org/2017/09/11/learning-transferable-architectures-for-scalable-image-recognition/
\n\nJust wondering if anyone is using machine learning driven architecture search (like [1]) for low complexity model for special prize of raspberry phi3?
\n\n[1] \"Neural architecture search with reinforcement learning.\" - Barret Zoph, Quoc V. Le
", "messages": [ { "id": "268078", "postDate": "01/13/2018 07:42:04", "content": "\n\nI can sense that this may be coming in a year or two. Afterall, google NASNET is an example \n that machined design network is as good as hand-crafted network.
\n\nhttps://research.googleblog.com/2017/11/automl-for-large-scale-image.html\nhttps://blog.acolyer.org/2017/09/11/learning-transferable-architectures-for-scalable-image-recognition/
\n\nJust wondering if anyone is using machine learning driven architecture search (like [1]) for low complexity model for special prize of raspberry phi3?
\n\n[1] \"Neural architecture search with reinforcement learning.\" - Barret Zoph, Quoc V. Le
", "rawMarkdown": "https://singularityhub.com/2018/01/03/ai-uses-titan-supercomputer-to-create-deep-neural-nets-in-less-than-a-day/#sm.0001y64y04ow0d25upt2atyjwt59p\n\nI can sense that this may be coming in a year or two. Afterall, google NASNET is an example \n that machined design network is as good as hand-crafted network.\n\nhttps://research.googleblog.com/2017/11/automl-for-large-scale-image.html\nhttps://blog.acolyer.org/2017/09/11/learning-transferable-architectures-for-scalable-image-recognition/\n\nJust wondering if anyone is using machine learning driven architecture search (like [1]) for low complexity model for special prize of raspberry phi3?\n\n\n[1] \"Neural architecture search with reinforcement learning.\" - Barret Zoph, Quoc V. Le", "votes": null }, { "id": "268149", "postDate": "01/13/2018 14:37:23", "content": "lol It can be! But not now at least!
", "rawMarkdown": "lol It can be! But not now at least!", "votes": null }, { "id": "268174", "postDate": "01/13/2018 16:26:20", "content": "The article is very interesting.... but, it is a little misleading. A lot of the difference lies in compute capability than anything else.... a supercomputer with 18000 GPUs? 500,000 networks in a day? I couldn't train 50 networks during this whole competition.
\n\nThat said... as computing power continues to skyrocket, who knows? Computing power may not be the limiting factor in the future.
", "rawMarkdown": "The article is very interesting.... but, it is a little misleading. A lot of the difference lies in compute capability than anything else.... a supercomputer with 18000 GPUs? 500,000 networks in a day? I couldn't train 50 networks during this whole competition.\n\n That said... as computing power continues to skyrocket, who knows? Computing power may not be the limiting factor in the future.", "votes": null }, { "id": "268175", "postDate": "01/13/2018 16:29:28", "content": "you may want to check slides from goggle brain:\nhttp://ucsc-sip.org/wp-content/uploads/2017/12/JeffDeanSIPreunion1217.pdf
\n\npage 57 onwards:
\n\nAutomated machine learning\n(“learning to learn”)
\n\nCurrent:\n Solution = ML expertise + data + computation
\n\nCan we turn this into:\n Solution = data + 100X computation
", "rawMarkdown": "you may want to check slides from goggle brain:\nhttp://ucsc-sip.org/wp-content/uploads/2017/12/JeffDeanSIPreunion1217.pdf\n\npage 57 onwards:\n\nAutomated machine learning\n(“learning to learn”)\n\n\nCurrent:\n Solution = ML expertise + data + computation\n\nCan we turn this into:\n Solution = data + 100X computation", "votes": null }, { "id": "268194", "postDate": "01/13/2018 17:36:41", "content": "That's fascinating, thank you!
", "rawMarkdown": "That's fascinating, thank you!", "votes": null }, { "id": "270029", "postDate": "01/17/2018 17:07:59", "content": "AIBot is really comming:
\n\n\n\nhttps://cloud.google.com/automl/
", "rawMarkdown": "AIBot is really comming:\n\n\nhttps://techcrunch.com/2018/01/17/googles-automl-lets-you-train-custom-machine-learning-models-without-having-to-code/\n\nhttps://cloud.google.com/automl/", "votes": null } ], "comments": [ { "id": 268149, "author_name": "learnitdeep", "author_url": "", "post_date": "01/13/2018 14:37:23", "content": "lol It can be! But not now at least!
", "votes": null, "replies": [] }, { "id": 268174, "author_name": "rhgrossm", "author_url": "", "post_date": "01/13/2018 16:26:20", "content": "The article is very interesting.... but, it is a little misleading. A lot of the difference lies in compute capability than anything else.... a supercomputer with 18000 GPUs? 500,000 networks in a day? I couldn't train 50 networks during this whole competition.
\n\nThat said... as computing power continues to skyrocket, who knows? Computing power may not be the limiting factor in the future.
", "votes": null, "replies": [ { "id": 268175, "author_name": "hengck23", "author_url": "", "post_date": "01/13/2018 16:29:28", "content": "you may want to check slides from goggle brain:\nhttp://ucsc-sip.org/wp-content/uploads/2017/12/JeffDeanSIPreunion1217.pdf
\n\npage 57 onwards:
\n\nAutomated machine learning\n(“learning to learn”)
\n\nCurrent:\n Solution = ML expertise + data + computation
\n\nCan we turn this into:\n Solution = data + 100X computation
", "votes": null, "replies": [] }, { "id": 268194, "author_name": "rhgrossm", "author_url": "", "post_date": "01/13/2018 17:36:41", "content": "That's fascinating, thank you!
", "votes": null, "replies": [] } ] }, { "id": 270029, "author_name": "hengck23", "author_url": "", "post_date": "01/17/2018 17:07:59", "content": "AIBot is really comming:
\n\n\n\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "268078": "https://singularityhub.com/2018/01/03/ai-uses-titan-supercomputer-to-create-deep-neural-nets-in-less-than-a-day/#sm.0001y64y04ow0d25upt2atyjwt59p\n\nI can sense that this may be coming in a year or two. Afterall, google NASNET is an example \n that machined design network is as good as hand-crafted network.\n\nhttps://research.googleblog.com/2017/11/automl-for-large-scale-image.html\nhttps://blog.acolyer.org/2017/09/11/learning-transferable-architectures-for-scalable-image-recognition/\n\nJust wondering if anyone is using machine learning driven architecture search (like [1]) for low complexity model for special prize of raspberry phi3?\n\n\n[1] \"Neural architecture search with reinforcement learning.\" - Barret Zoph, Quoc V. Le", "268149": "lol It can be! But not now at least!", "268174": "The article is very interesting.... but, it is a little misleading. A lot of the difference lies in compute capability than anything else.... a supercomputer with 18000 GPUs? 500,000 networks in a day? I couldn't train 50 networks during this whole competition.\n\n That said... as computing power continues to skyrocket, who knows? Computing power may not be the limiting factor in the future.", "268175": "you may want to check slides from goggle brain:\nhttp://ucsc-sip.org/wp-content/uploads/2017/12/JeffDeanSIPreunion1217.pdf\n\npage 57 onwards:\n\nAutomated machine learning\n(“learning to learn”)\n\n\nCurrent:\n Solution = ML expertise + data + computation\n\nCan we turn this into:\n Solution = data + 100X computation", "268194": "That's fascinating, thank you!", "270029": "AIBot is really comming:\n\n\nhttps://techcrunch.com/2018/01/17/googles-automl-lets-you-train-custom-machine-learning-models-without-having-to-code/\n\nhttps://cloud.google.com/automl/" }, "source": "meta" }