Curious: how are people finding the most important recent AI papers? (I haven't been seriously following, but would like to catch up.)
Show this thread
-
.
@karpathy's arxiv-sanity is a great project, and I wondered how the top papers over the past year (http://www.arxiv-sanity.com/top?timefilter=year&vfilter=all … ) would be. But many of the most-saved papers are elementary tutorials, which makes me wonder about the collective(!) judgement of users on the siteShow this thread
What are your favourite 1-3 AI / ML papers over the past 12-24 months? (And feel free to say what excited you about them?)
-
-
(To clarify the arxiv-sanity remark: nothing against the project or users! But it suggests the collective judgements are from people fairly new to the field - natural, given the influx into deep learning. But curious about the recommendations of ppl who've been at this a while!)
Show this thread -
The suggestions so far are great! Please keep them coming! Also love to hear about classic papers that you particularly like!
Show this thread
End of conversation
New conversation -
-
-
@numenta's Framework for Intelligence and Cortical Function Based on Grid Cells in the Neocortex! (and related papers) It feels like speculation at this point, but a speculation I like :)http://bit.ly/2qznPCw - End of conversation
New conversation -
-
-
- I loved PredNet: https://arxiv.org/abs/1605.08104 (they have a YT lecture too which is awesome) - Neural Turing Machine blew my mind https://www.nature.com/articles/nature20101 … - Curiosity-driven exploration seems very promisinghttps://pathak22.github.io/noreward-rl/
-
Thanks, looks fun!
End of conversation
New conversation -
-
-
There is this transformer hype, starting from: https://arxiv.org/abs/1706.03762
-
And NLP transfer learning (elmo, ULMFit, transformer, BERT)
- 1 more reply
New conversation -
-
-
- ELMo: LM pretraining is key - MAML: Neat framework for meta-learning with gradients only - MINE and DeepInfomax: Estimate Mutual Information with NN magic Lesser known: - "Model Ensemble TRPO" & "DL in a Handful of Trials": Model Based RL can work if uncertainty is handled
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I recommend "Self-critical Sequence Training" [0], which proposed using RL's Policy Gradients directly on non-differentiable metrics [ex. CIDEr, BLUE, ROUGE]. Also, using SCST makes it possible to utilize human-in-the-loop. [0] https://arxiv.org/abs/1612.00563
-
And I double Transformer and esp. curiosity-driven exploration — based on which the latest
@OpenAI model "solved" extremely hard Montezuma’s Revenge [1]. [1]https://blog.openai.com/reinforcement-learning-with-prediction-based-rewards/ …
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.