Twitter | Search | |
Rupesh Srivastava
AI research . I want to build a self-improving system smarter than .
287
Tweets
196
Following
1,260
Followers
Tweets
Rupesh Srivastava Jan 14
I remember Juergen explaining the basics of machine learning when I joined the field and it was always about "exploiting algorithmic information about data through compression". IMO distinctions like SL/UL/RL are almost superfluous & much less useful than fundamental principles.
Reply Retweet Like
Rupesh Srivastava retweeted
Jeremy Howard Jan 13
At medical imaging we often need to make a little data go a long way. One under-appreciated approach for this is self-supervised learning. It's almost magical! I've written a little overview to help you get started. Let me know if you try it :)
Reply Retweet Like
Rupesh Srivastava Jan 2
Replying to @srchvrs
"Distilling the Knowledge in a Neural Network" has 2.5K citations, and **does not have a related work section** :)
Reply Retweet Like
Rupesh Srivastava Dec 19
Sure but they better earn it! Okay may be don't.
Reply Retweet Like
Rupesh Srivastava Dec 19
This is a clear explanation of reasons behind the protests in India. It is important to know the details, but it is also obvious what the big picture behind CAA and NRC is.
Reply Retweet Like
Rupesh Srivastava retweeted
DeepAI Dec 9
Training Agents using Upside-Down Reinforcement Learning by Rupesh Kumar Srivastava et al. including
Reply Retweet Like
Rupesh Srivastava Dec 18
Hi Francois, are you familiar with this recent TR from my colleague ? On approximating ∇f with neural networks:
Reply Retweet Like
Rupesh Srivastava retweeted
NNAISENSE Dec 17
Announcing a new round of funding! As we take the next step in bringing to industrial automation, we welcome Samsung Ventures, & Repsol Energy Ventures as we welcome back & Metaplanet.
Reply Retweet Like
Rupesh Srivastava retweeted
Parth Chadha Dec 16
Thanks for suggesting , interesting read! Wrote a quick naive implementation. Check it out here
Reply Retweet Like
Rupesh Srivastava Dec 14
In 15 mins. (2:30pm), I'll present Upside-Down RL at the poster session in the Deep RL workshop (West Exh. Hall C). Come to learn about a new kind of RL!
Reply Retweet Like
Rupesh Srivastava Dec 12
Nice explanatory video by on ⅂ꓤ!🙃
Reply Retweet Like
Rupesh Srivastava Dec 12
Haha yes! Come to my poster at the Deep RL workshop on Saturday!
Reply Retweet Like
Rupesh Srivastava Dec 12
Thanks for your interest!
Reply Retweet Like
Rupesh Srivastava Dec 12
Finally, we run 20 exps with pre-decided seeds, and the plots summarize statistics over them. So we don't *consistently* reach 750 avg. under this eval scheme *within the hyperparam sets fairly evaluated*. The task isn't too hard so we didn't target absolute best perf. (3/3)
Reply Retweet Like
Rupesh Srivastava Dec 12
b) We fix set of possible values of hyperparams and tune them for all algos (including architecture). We use separate env seeds for training, tuning and testing, and also use separate algo. seeds for final evals. Trying to be rigorous and fair but not the best.
Reply Retweet Like
Rupesh Srivastava Dec 12
Hi! Yes, we do get >750 avg., more often if we use larger nets (not in the paper). Reasons our results may not be directly comparable: a) We convert to gray-scale, downsample to 32x32, and use 8 frames instead of RNNs. This is to keep setup simple and fast for more runs 1/
Reply Retweet Like
Rupesh Srivastava retweeted
Sjoerd van Steenkiste Dec 12
Are disentangled representations helpful for abstract visual reasoning? Come find out later today during the afternoon poster session at 17:00, poster board 28. This is joint work with , , and
Reply Retweet Like
Rupesh Srivastava retweeted
Synced Dec 9
"Training Agents using Upside-Down Reinforcement Learning" by Rupesh Kumar Srivastava, Pranav Shyam, Filipe Mutz, Wojciech Jaśkowski and Jürgen Schmidhuber from and The Swiss AI Lab IDSIA. Read the full paper at .
Reply Retweet Like
Rupesh Srivastava retweeted
osdf Dec 10
Transformations tell us a lot about objects they are acting on,so maybe we should have 'latent transformations', too?See idea and first results in our small note 'No Representation without Transformation' (). Joint work with ,
Reply Retweet Like
Rupesh Srivastava retweeted
celestekidd Dec 9
I'm happy to share my comments on the climate for men from my talk:
Reply Retweet Like