For anyone worried their LLM might be making stuff up, we made a budget‐friendly truth serum (semantic entropy + Bayesian). See for yourself: https://t.co/gq8oFP5Eqr Paper:

Interested in working on exciting ML ideas for Spotify? Join us! We are looking for a Research Scientist Intern to join our research lab in London for summer 2023. https://t.co/62V4NvEddS @SpotifyResearch

On the other hand, in “traditional” dynamic programming we have to carefully design the order of updates, which is problem-specific. (3/3)

In fact, these meanings are closely related. Q-learning is a type of dynamic programming. The main insight is that, in Q-learning, updates can be made in a random order, but we are still able to solve arbitrary MDPs. (2/3)

People sometimes say that the term “dynamic programming” has two distinct meanings. The first meaning is what you typically encounter in an algorithms textbook, for things like Levenshtein distance etc. The second meaning is about RL and things like Q-learning. (1/3)

Kullback-Leibler divergence is not the same as Leibler-Kullback divergence

The competition aspect drives people to do well and improves science, but may not be the most optimal way to do so. Outcomes are also often pretty noisy. @TmlrOrg looks like an interesting alternative path. (2/2)

Since acceptance rate in major conferences is approximately constant ( https://t.co/EOIRkXYCAt), conferences are in a way zero-sum games. #NeurIPS2022 #iclr2022. (1/2)

Want to do imitation learning in a simple and efficient way? We released code for the ICLR 2022 paper “Imitation Learning by Reinforcement Learning”. See https://t.co/pL8beWndyR.

Interested in Imitation Learning? You can do it using a single call to a Reinforcement Learning oracle. See our #ICLR2022 paper “Imitation Learning by Reinforcement Learning” ( https://t.co/o8X6UIMO7n).

I'm excited to present our work on active reward learning at #NeurIPS2021! We propose a general way to make queries that are informative about the optimal policy. Joint work with @MatteoTurchetta, Sebastian Tschiatschek, @MLciosek, and @arkrause: https://t.co/wSQLoWppHm 👇(1/6)

Like policy gradients? In "Expected Policy Gradients for Reinforcement Learning", we study various quadrature schemes to decrease variance in gradient estimates. Final version is now published in JMLR (@JmlrOrg). See https://t.co/a9gljfYJDf.

Our work on the reproducibility of meta-RL baselines (Bandits + MDPs) with MAML and Reptile is at the Reproducibility in ML workshop (C2) #ICML2018 together with @arianTBD & Simon Guiroy @MILAMontreal

Learn about our work on "Expected Policy Gradients" in this 13min video by Kamil Ciosek - with @MLciosek @shimon8282 https://t.co/G296BlOS48

"Fourier Policy Gradients" by Matthew Fellows, Kamil Ciosek and Shimon Whiteson https://t.co/v2Q5xd9HLh

Our new paper: using Fourier analysis to derive policy gradients: we recast the integrals as convolutions, which a Fourier transform turns into multiplications. The resulting analysis unifies existing policy gradient results.