Postdoc with @KanakaRajanPhD working on Neuro-inspired modular reinforcement learning, previously PhD at Princeton University. A better world is possible.
Princeton, NJ
Joined March 2020
- Tweets114
- Following60
- Followers100
- Likes232
Riley Simmons-Edler retweeted
(1/7) New preprint from the Rajan Lab! 🧠🤖 @RyanPaulBadman1 & @SimmonsEdler show–through cog sci, neuro & ethology-how an AI agent with fewer ‘neurons’ than an insect can forage, find safety & dodge predators in a virtual world. Here's what we did
Paper: arxiv.org/pdf/2506.06981
Riley Simmons-Edler retweeted
Our new preprint from Rajan lab (Harvard):
"Deep RL Needs Deep Behavior Analysis: Exploring Implicit Planning by Model-Free Agents in Open-Ended Environments"
Sophisticated exploration, predator evasion, and foraging strategies by standard DRL.
arxiv.org/abs/2506.06981
Riley Simmons-Edler retweeted
New paper published by @ICMLconf #icml2024
@SimmonsEdler, @RyanPaulBadman1, @ShayneRedford & I argue that global embrace of military #AI systems with less human oversight makes conflict more likely, jeopardizing global stability & AI research norms: arxiv.org/abs/2405.01859
Riley Simmons-Edler retweeted
Expanding our research programs & our group! New website & hiring details coming soon…
For now, @RyanPaulBadman1, @SimmonsEdler & I are looking forward to a new chapter in #neuroAI, unlocking the mysteries of the brain using #artificialintelligence & #neurotheory 🚀
Another personal update: I've officially joined Mt. Sinai as a postdoc to work with the brilliant Kanaka Rajan on neuro-inspired and modular RL! Looking forward to doing great things together!
The Rajan lab is excited to welcome Dr. @SimmonsEdler! By way of @SebastianSeung at @PrincetonNeuro, Riley works on neuro-inspired reinforcement learning. We'll be creating modular RL algos that decompose tasks into subtasks, which can be reused in robust ways in new contexts!
A couple days late in tweeting it, but: new blog post! This time I'm taking a holistic look at multitask RL, why it's hard, and how modular learning could improve it: rileyse.org/2022/03/21/why-i…
3
The blogging continues! My latest, on how generalization makes reinforcement learning hard: rileyse.org/2021/11/29/why-i…
1
1
I don't know what secret sauce animals use here, but the generalization gap suggests we need to reconsider how we approach the problem.
1
1
Amusingly, while I was writing this some related work came out which also reexamines generalization (in a complementary way). If this topic interests you, here's a recent survey paper that looks in more depth at the methods and benchmarks currently in use:
arxiv.org/pdf/2111.09794.pdf
1
1
New blog post! I think two posts makes this officially a "blog," right? 😛 rileyse.org/2021/10/06/detec…
1
In this one, I'm taking a closer look at sampling-related bottlenecks from the last post. I develop a high-level framework based on conditional regret for detecting bottlenecks, and then discuss some approaches for estimating that quantity.
1
I started a blog! I've had some RL hot-takes/research ideas swirling about for a while, and this seems like a good way to share them informally: rileyse.org/2021/09/16/rl_bo…
1
3
The first topic: Detecting performance bottlenecks in RL: How do we figure out what's limiting an agent's performance on a task? Usually, this is more of an art than a science. Here are some baby steps towards the latter.
1
1
As this is new analysis it's sadly not eligible for ICLR's new blog post track, my dreams of winning a best blog post award will have to wait. 😢
teapawt.net/iclr_conf/status/14374…
ICLR is happy to announce the call for contributions for our very first "Blog Posts Track". We invite submissions in blog format discussing previously published papers at ICLR. For details on this exciting new experiment in publication models see: iclr.cc/Conferences/2022/Cal…
1
Time for another new paper thread: Introducing "AuraSense: Robot Collision Avoidance by Full Surface Proximity Detection" to appear at IROS next month! arxiv.org/abs/2108.04867
2
1
What's more, this setup is extremely practical: We basically hot glued some $3 transducers to a robot and connected them to a commodity DAC, then piped the signals through a modest CNN. It can be cheaply implemented on any robot, and it's easy to imagine use in production.
1
Thanks for reading! This was work I did at SAIC-NY with my fantastic collaborators Xiaoran (Van) Fan, Daewon Lee, Larry Jackel, Rich Howard, and Dan Lee (none of whom are on twitter, SMH). Check out our paper for more! arxiv.org/abs/2108.04867
1