You might've missed it because Twitter sucks, but the LSUV variant often does very well for the parser: https://t.co/OsuN6Gk2Rg
— Matthew Honnibal (@honnibal) September 14, 2018
Honestly parsing, tagging, NER etc are not great value topics to dabble in. They're pretty fiddly, and hard to publish on.
Sure thing! You might skim this: https://t.co/nFI8DsHAIK . Now I just have a NN predicting the actions, instead of an averaged perceptron.
— Matthew Honnibal (@honnibal) September 14, 2018
It's sort of like reinforcement learning. To predict the next state, we need the current state to make the features.
Hamiltonian Descent Methods (https://t.co/rgNkkOgQNM): our generalisation of the momentum method extends the class of convex functions that can be optimised with exponential rates using just gradients. pic.twitter.com/myeFLjB8Zj
— DeepMind (@DeepMindAI) September 14, 2018
Check out the samples of generated music using the new “Relative Transformer” model. My favourite ones are the “Piano-e-Competition Primed Samples” at the bottom of this page: https://t.co/JrKnw2BwE4 https://t.co/0V2FEaExP5
— hardmaru (@hardmaru) September 13, 2018
Histogram of P-values (rounded, presumably, to the nearest 0.01) published in millions of scientific studies. Hmm. Great, provocative talk by @StatGarrett at #earlconf pic.twitter.com/42MegOhjsQ
— David Smith (@revodavid) September 12, 2018
A good analysis of the embeddings we use in the OpenAI Five model: https://t.co/Dx9ZseZYhx. After analyzing the rest of the model, concludes: “All the strategy and tactics must lie in one place – 1024-unit LSTM.” Pretty remarkable what LSTMs (invented in 1997!) can do at scale.
— Greg Brockman (@gdb) September 11, 2018
Technology for interacting with a web browser via natural language instructions—a new challenge for grounding language with dataset, initial neural models by @IcePasupat•Jiang•Liu•@kelvin_guu•@percyliang. #emnlp2018 #NLProc https://t.co/LlYQw9c6AM pic.twitter.com/18r0oiApb2
— Stanford NLP Group (@stanfordnlp) September 10, 2018
Training Millions of Personalized Dialogue Agents
— ML Review (@ml_review) September 9, 2018
w/ @braizh
Shows how to create a very large dataset for persona-based dialogue
Persona-based end-to-end dialogue models outperform counterparts that do not have access to personashttps://t.co/kSV4ZSU6Kl pic.twitter.com/4nmPCOiuAx
Oh that's clever! :) https://t.co/BlEwI71zq7
— Jeremy Howard (@jeremyphoward) September 7, 2018
Congratulations to @julius_adebayo , @_beenkim, @muelly and @mrtz ! "Sanity Checks for Saliency Maps" (on arxiv soon) will be a spotlight at NIPS 2018 pic.twitter.com/EOjVMeQaWC
— Ian Goodfellow (@goodfellow_ian) September 6, 2018
Congratulations to @avitaloliver , @gstsdn , @colinraffel , and @ekindogus ! "Realistic Evaluation of Deep Semi-Supervised Learning Algorithms" will be a spotlight at NIPS 2018. https://t.co/L1wWdAQlYh pic.twitter.com/vck3vc4xbv
— Ian Goodfellow (@goodfellow_ian) September 6, 2018
Happy to announce that our work, ”Bilinear Attention Networks (BAN)” is accepted for #NIPS2018!!! 🎉 SotA on VQA 2.0, Flickr30k Entities, and VizWiz! https://t.co/pgaAEremB7
— Jin-Hwa Kim (@jnhwkim) September 5, 2018
