Saurabh Gupta’s Post

Reinforcement Learning agents used to struggle to complete simple multi-step tasks. Without a basic understanding of the world, the only way to make progress was exhaustive trial and error. Kamyar Azizzadenesheli explains how LLMs changed everything. 🎧/🎥: https://bit.ly/3wajpqP

Multi agent Reinforcement learning https://lnkd.in/dn7XqQHT 2024 version MIT Press

🚀 Excited to share our latest blog post on learning to watermark LLM-generated text using reinforcement learning. We explore the potential of embedding detectable signals into LLM-generated text to track misuse and offer a new approach to watermark design space. Our model-level watermark, embedding signals into LLM weights, demonstrates enhanced accuracy, robustness, and adaptability, even allowing for open-sourcing of the watermarked model. Discover more and explore our open-source code: [Read the full post here](https://bit.ly/3IIuCSl). #SocialMediaMarketing #ReinforcementLearning #LLM #Watermarking

“We empirically show that a GPT-style transformer exhibits a transition from in-distribution to out-of-distribution generalization as the number of pre-training tasks increases.” Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks https://lnkd.in/gewg9T_h source: https://lnkd.in/g2JZvuKi

Hey all, in my previous blog, I covered the basics of RL, covering the scenarios of Model based (Planning), Model Free ( Learning), Function Approximation ( e.g. deep-Q-network/ DQN). Here I tried cover policy gradients, importance sampling, natural policy gradients, TRPO and then finally PPO — which is very popularly used for RLHF ( Reinforcement Learning from Human Feedback) for LLM these days. The mathematical details and derivations get really complicated in some of these topics; I have tried to keep them as simple as possible. While reading the individual papers, from policy gradient ( NIPS , 1999 paper) to PPO (2017), I have tried to focus on what are the key ideas they introduced and what are new improvements. #reinforcement_learning #rl #dl #deeprl #markov_decision_process #deeplearning #policy_based_rl #npg #trpo #ppo

Our paper on a utility-based approach to reinforcement learning was accepted for presentation at AAMAS later this year, and is available for reading on arxiv https://lnkd.in/dktJb4Qa

🌟 Excited to share our latest blog post on Action-constrained Policy Gradient with Normalizing Flows for reinforcement learning! 🌟 Action-constrained reinforcement learning (ACRL) is crucial for addressing safety-critical and resource-allocation related decision making problems. Our new approach utilizes a normalizing flow model to learn a mapping between the feasible action space and a distribution on a latent variable. We tackle the challenge of action sampling for convex and non-convex constraints, and integrate the learned normalizing flow with the DDPG algorithm. The result? Significantly fewer constraint violations and improved speed on continuous control tasks. Access the full post here: https://bit.ly/3OC57p7 #ReinforcementLearning #PolicyGradient #AIResearch #SocialMediaMarketing

For a patient introduction to reinforcement learning, look no further than Oliver S's ongoing series; in part 2, we dive deep into the inner workings of Markov decision processes.

Leveled up my RL skills! After some trial and error, finally trained an agent to ace the Acrobot-v1 challenge. Reinforcement Learning is truly fascinating! Talk about defying gravity! Video is recorded at 15FPS instead of ideal 30FPS. Here is the repository if you have some tinkering in mind :) https://lnkd.in/g7cWineW #ReinforcementLearning #DeepLearning #GymEnvironment #NeuralNetworks #MachineLearning

From Clinical Problem Solvers– Spaced Learning Series Episode 4 – Recurrent Presyncope https://lnkd.in/g-2V66cJ