[{"data":1,"prerenderedAt":287},["ShallowReactive",2],{"project-25683808":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":34,"aspects":35,"strip":41,"post_text":42,"what":43,"said":73,"maker_in_thread":148,"shape":157,"grid":160,"attempts":248,"same_week":249,"maker_shows":32,"maker_hidden":279,"reader_version":280,"card":281},"25683808","Ecco","See what your NLP language model is “thinking”","title","Show HN: Ecco – See what your NLP language model is “thinking”","https:\u002F\u002Fwww.eccox.io\u002F","eccox.io","jalammar",1610107727,2021,185,34,252,2533,10,"notable","ai","model interpretability","library",[24,25,26,27],"transformer interpretability","neuron activations","language model visualization","hugging face","loved",0.531,{"positive":18,"neutral":31,"mixed":32,"critical":33},5,2,0,17,{"docs":36,"design":38,"idea":40},{"positive":37,"critical":33},3,{"positive":39,"critical":33},1,{"positive":31,"critical":33},[33,31,33,33,33,33,31,37,33,33,31,33,33,37,33,37,37,31,33,37,33,31,33,31,33,33,37,33,37,31,31,31,31,33],null,{"problem":44,"audience":42,"pricing":42,"motivation":51,"time_to_build":42,"stack":55,"openness":71,"stage":72,"team_size":72},{"value":45,"source":46,"evidence":47},"Examining neuron activations inside Transformer language models to see what they are thinking","maker",[48],{"id":49,"by":11,"text":50},"25683813","Hi HN, Author here. I had been fascinated with Andrej Karpathy's article (https:\u002F\u002Fkarpathy.github.io\u002F2015\u002F05\u002F21\u002Frnn-effectiveness\u002F) -- especially where it shows neurons being…",{"value":52,"source":46,"evidence":53},"Inspired by Karpathy's RNN article showing neurons responding to brackets and indentation",[54],{"id":49,"by":11,"text":50},[56,63,67],{"value":57,"source":58,"evidence":59},"Python","inferred",[60],{"id":61,"by":11,"text":62},"25690285","I actually started with PCA. But NMF proved more understandable since negative dimensions in PCA are hard to interpret. I didn't consider UMAP, but would be interested to see how…",{"value":64,"source":46,"evidence":65},"Hugging Face transformers",[66],{"id":49,"by":11,"text":50},{"value":68,"source":46,"evidence":69},"scikit-learn NMF",[70],{"id":61,"by":11,"text":62},"open_source","unknown",{"praise":74,"objections":109,"requests":117,"questions":129,"alternatives":141},[75,91,98],{"value":76,"source":77,"evidence":78},"Visualizations are clear and helpful for learning transformers","commenter",[79,83,87],{"id":80,"by":81,"text":82},"25688286","yowlingcat","Wow, love the NNMF visualization. Like all great visualizations, it does a very good job of showing and not telling me what's going on. More of this, please. One question: how…",{"id":84,"by":85,"text":86},"25684149","anfal_alatawi","Thank you, Jay! I appreciate the addition of the colab notebooks with code examples. I can't wait to play around with this and investigate how language models _speak_.",{"id":88,"by":89,"text":90},"25688876","Grimm1","This is fantastic, I used your earlier transformers article to first get a real grasp on the architecture. I hope you expand this to accommodate other modes of attention outside…",{"value":92,"source":77,"evidence":93},"NMF factorization of activations is a good approach",[94],{"id":95,"by":96,"text":97},"25685259","shenberg","NMF for factorizing activations is brilliant!",{"value":99,"source":77,"evidence":100},"Useful for organizations adopting AI",[101,105],{"id":102,"by":103,"text":104},"25684872","indymike","Helping people understand \"what the ai is thinking\" is really important when you are trying to get organizations to adopt the technology. Great work.",{"id":106,"by":107,"text":108},"25687056","ZeroCool2u","Fantastic work. This is the kind of stuff we need to get these models actually adopted and integrated into non-tech organizations.",[110],{"value":111,"source":77,"evidence":112},"Factorization results require interpretation and may hide structure",[113],{"id":114,"by":115,"text":116},"25685453","GistNoesis","Interesting. The non-negative matrix factorization on the first level kinda highlight some semantic groupings : paragraph, verbs, auxiliaries, commas, pronouns, nominal…",[118,122],{"value":119,"source":77,"evidence":120},"Support other attention modes beyond transformers",[121],{"id":88,"by":89,"text":90},{"value":123,"source":77,"evidence":124},"Add other dimensionality reduction methods such as UMAP",[125],{"id":126,"by":127,"text":128},"25689933","Der_Einzige","This work is awesome! Are there theoretical reason to choose NMF over other dimensionality reduction algorithms, e.g. UMAP? Is it easy to add other DR algorithms? I may submit a…",[130,137],{"value":131,"source":77,"evidence":132},"Novel insights obtained with the method",[133],{"id":134,"by":135,"text":136},"25684436","blackbear_","Any examples of novel insights obtained with this method?",{"value":138,"source":77,"evidence":139},"How it relates to explainable AI",[140],{"id":80,"by":81,"text":82},[142],{"value":143,"source":46,"evidence":144},"LIME and Shapley values",[145],{"id":146,"by":11,"text":147},"25689450","These are AI explanation methods. They belong to the toolbox which would include LIME, Shapley values...etc. Input saliency is a gradient-based explanation method.",{"comments":18,"answered":37,"key":149},[150,151,154],{"id":49,"text":50},{"id":152,"text":153},"25684694","I'd love to look at your group's visualizations! Is it a private repo? because the link doesn't open up. It never stops to blow my mind that we can represent words and concepts…",{"id":155,"text":156},"25690703","Hey, I feel you! I'm an intuitive learner as well. I wouldn't have been able to learn much in ML if it weren't for people who write and visualize and make the methods accessible…",{"commenters":158,"max_depth":31,"maker_replies":18,"first_comment_mins":39,"lifespan_hours":159},21,54,[161,163,167,169,172,176,177,180,184,187,189,191,193,196,198,200,204,206,207,210,212,214,217,219,221,224,225,228,231,234,236,238,242,246],[49,33,33,162,11,39],"maker\u002Fneutral\u002Fidea",[164,33,31,165,166,33],"25684476","maker\u002Fpositive\u002Fdocs","ninjin",[152,33,33,168,11,39],"maker\u002Fneutral\u002Fnone",[170,33,33,171,166,33],"25685307","commenter\u002Fneutral\u002Fnone",[173,33,33,174,175,33],"25690169","commenter\u002Fpositive\u002Fnone","ptd",[155,33,33,165,11,39],[178,33,31,165,179,33],"25691583","tchalla",[181,33,37,182,183,33],"25685311","project\u002Fmixed\u002Fdesign","airstrike",[185,33,33,171,186,33],"25691168","cmrdsprklpny",[188,33,33,171,183,33],"25691241",[88,33,31,190,89,33],"project\u002Fpositive\u002Fidea",[192,33,33,162,11,39],"25689396",[194,33,33,195,89,33],"25690017","other\u002Fneutral\u002Fidea",[114,39,37,197,115,33],"project\u002Fmixed\u002Fidea",[199,39,33,162,11,39],"25689899",[201,39,37,202,203,33],"25709821","project\u002Fneutral\u002Fdocs","mcswell",[205,39,37,202,115,33],"25717660",[102,32,31,190,103,33],[208,32,33,195,209,33],"25685062","nathanyz",[126,37,37,211,127,33],"project\u002Fneutral\u002Fidea",[61,37,33,213,11,39],"maker\u002Fneutral\u002Fexecution",[84,215,31,216,85,33],4,"project\u002Fpositive\u002Fdocs",[218,215,33,168,11,39],"25684521",[80,31,31,220,81,33],"project\u002Fpositive\u002Fdesign",[222,31,33,195,223,33],"25689441","gfody",[146,31,33,162,11,39],[134,226,37,227,135,33],6,"project\u002Fneutral\u002Fnone",[229,226,33,230,11,39],"25684501","maker\u002Fpositive\u002Fidea",[232,226,37,227,233,33],"25690629","amrrs",[106,235,31,190,107,33],7,[95,237,31,190,96,33],8,[239,240,31,190,241,33],"25687359",9,"pizza",[243,18,31,244,245,33],"25683962","project\u002Fpositive\u002Fnone","khalidlafi",[247,18,33,168,11,39],"25684074",[],[250,255,260,265,270,274],{"id":251,"name":252,"points":253,"category":254},"25677272","Synergy",343,"languages",{"id":256,"name":257,"points":258,"category":259},"25749492","Amplication",272,"dev_tools",{"id":261,"name":262,"points":263,"category":264},"25713858","hn-big-threads",248,"consumer_social",{"id":266,"name":267,"points":268,"category":269},"25750807","Vantage",221,"infrastructure",{"id":271,"name":272,"points":273,"category":269},"25671919","I built an intercom for my 6 yo to keep us connected during quarantine",181,{"id":275,"name":276,"points":277,"category":278},"25678568","Leet Resumes",147,"productivity",false,"v2",{"id":282,"name":5,"line":6,"ai":279,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":283,"top":18,"kw":284,"maker":11,"domain":10,"tier":19,"attempts":33,"strip":285,"url":9,"mix":286,"praise":76,"objection":111,"sub":21},25683808,0.53,[24,25,26,27],"0500005300500303350305050030355550",[18,31,32,33],1794036601034]