A curated list of interesting stuff
","icon":"https://cdn.drexl.dev/feed/production/images/channel-0c5fee2ad0e633031076b623c4e52d02.jpeg","favicon":"https://feed-adg.pages.dev/assets/default/favicon.png","language":"en-us","items":[{"id":"a3CKJ2jIPh9","title":"Acquisition of Chess Knowledge in AlphaZero","attachments":[{"url":"https://cdn.drexl.dev/feed/production/media/document-e85e285c4088ac76040d21955b008d8f.pdf","mime_type":"application/pdf","size_in_byte":8988673}],"url":"https://feed.drexl.dev/i/by-chessbase-acquisition-of-chess-knowledge-a3CKJ2jIPh9/","content_html":"Abstract: What is being learned by superhuman neural network agents such as AlphaZero? This question is of both scientific and practical interest. If the representations of strong neural networks bear no resemblance to human concepts, our ability to understand faithful explanations of their decisions will be restricted, ultimately limiting what we can achieve with neural network interpretability. In this work we provide evidence that human knowledge is acquired by the AlphaZero neural network as it trains on the game of chess. By probing for a broad range of human chess concepts we show when and where these concepts are represented in the AlphaZero network. We also provide a behavioural analysis focusing on opening play, including qualitative analysis from chess Grandmaster Vladimir Kramnik. Finally, we carry out a preliminary investigation looking at the low-level details of AlphaZero's representations, and make the resulting behavioural and representational analyses available online.
","content_text":"Abstract: What is being learned by superhuman neural network agents such as\nAlphaZero? This question is of both scientific and practical interest. If the\nrepresentations of strong neural networks bear no resemblance to human concepts,\nour ability to understand faithful explanations of their decisions will be\nrestricted, ultimately limiting what we can achieve with neural network\ninterpretability. In this work we provide evidence that human knowledge is\nacquired by the AlphaZero neural network as it trains on the game of chess. By\nprobing for a broad range of human chess concepts we show when and where these\nconcepts are represented in the AlphaZero network. We also provide a behavioural\nanalysis focusing on opening play, including qualitative analysis from chess\nGrandmaster Vladimir Kramnik. Finally, we carry out a preliminary investigation\nlooking at the low-level details of AlphaZero's representations, and make the\nresulting behavioural and representational analyses available online.","date_published":"2023-01-09T02:21:35.848Z","_microfeed":{"is_audio":false,"is_document":true,"is_external_url":false,"is_video":false,"is_image":false,"web_url":"https://feed-adg.pages.dev/i/acquisition-of-chess-knowledge-in-alphazero-a3CKJ2jIPh9/","json_url":"https://feed-adg.pages.dev/i/a3CKJ2jIPh9/json/","rss_url":"https://feed-adg.pages.dev/i/a3CKJ2jIPh9/rss/","guid":"a3CKJ2jIPh9","status":"published","itunes:episodeType":"full","date_published_short":"Sun Jan 08 2023","date_published_ms":1673230895848}}],"_microfeed":{"microfeed_version":"0.1.2","base_url":"https://feed-adg.pages.dev","categories":[],"subscribe_methods":[{"name":"RSS","type":"rss","url":"https://feed-adg.pages.dev/rss/","image":"https://feed-adg.pages.dev/assets/brands/subscribe/rss.png","enabled":true,"editable":false,"id":"etofHjqsmOi"},{"name":"JSON","type":"json","url":"https://feed-adg.pages.dev/json/","image":"https://feed-adg.pages.dev/assets/brands/subscribe/json.png","enabled":true,"editable":false,"id":"EkEFmBnzBg7"}],"description_text":"A curated list of interesting stuff","itunes:type":"episodic","items_sort_order":"newest_first"}}