On March 4th, I’ll be in Philadelphia for the final day of #AAAI workshops, presenting our work on Sequence Reinforcement Learning at not one, but TWO workshops!
Morning: I will be presenting a poster at GenPlan.
Evening: Giving a talk on the same at PRL
If you’ll be at AAAI, let’s connect!
I’m holding open office hours for anyone who needs Python programming help. These are held in parallel to UMass CS 119 that I am teaching this semester. More details here: www.devdharpatel.com/office-hours
Does anyone you know need Python help? Share this! #Python #LearnToCode #OpenOfficeHours
As an aside, @iclr-conf.bsky.social, how are top reviewers identified? I’d love to nominate two of mine!
Comments from the Program Chair on NeurIPS work:
- We ran only 1 seed (incorrect— even ChatGPT got this right!)
- No code submitted. (They searched for "GitHub" instead of checking the supplemental materials)
While we may have been below the bar for NeurIPS, the reasons cited were outright false.
📝 Sharing my contrasting experiences with the #ICLR and #NeurIPS review processes:
At #ICLR, our paper’s average score jumped from 4.2 to 7 thanks to thoughtful reviewer correspondence during the rebuttal/discussion phase.
In contrast, the #NeurIPS feedback we received felt... less constructive.
Sequence RL model
Difference between Standard RL and Sequence RL
Excited to share that we have two papers accepted to #ICLR2025 #ICLR!
One of them:
🧠 "Overcoming Slow Decision Frequencies in Continuous Control"
✨ Brain-inspired, model-based Sequence RL achieving SOTA performance at human-level decision speeds!
#ReinforcementLearning #AI
This poor quality of assessment not only demotivates researchers, it also brings into question any claims of fairness and accountability made by NeurIPS. (4/4)
Lastly did not check the supplementary materials for code and instead checked our paper for github link (in a double blind review process) and said we made a spurious claim about including our code. (3/4)
For example, despite giving many references to neuroscience papers, the PC clams that none of it is supported by neuroscience or behavioral data. We clearly present results with error bars and averaged over 5 seeds (explicitly mentioned in the paper), yet the PC claimed we presented only 1 seed(2/4)
