Wei Deng / Homepage / Twitter

Links for me and for you

• Andrej Risteski (Generative Model & MCMC)
• almost stochastic
• arg min
• Greg
• inFERENCe
• Inverse Probability - Neil Lawrence
• I’m a bandit
• Djalil Chafaï
• Eric Jiang, Flow in JAX
• Hopfield Networks is All You Need
• Lil'Log
• LLM
• Ilya V. Schurov
• Machine Learning Research
• Muon Optimizer - Keller Jordan
• Piecewise Deterministic Monte Carlo
• Probably Approximately Wrong
• Sam Power (MCMC)
• Sander Dieleman (Diffusion Models)
• Sebastian's Slow Blog
• Going Deeper Into RL: Fundamentals of Policy Gradients
• The Information Structuralist
• Thinking Machines
• Transformer Circuits Thread
• Tuan Anh Le
• VAIBHAV PATEL
• Shunyu Yao
• Yuling Yao's Blog
• The Annotated S4
• Simplifying S4
• KL Divergence
• 苏剑林 科学空间
• 紫气东来 LLM

Research

• Contour Sampler is the first algorithm that achieves the real free explorations and exploitations in MNIST dataset [Demo]; most other samplers cannot escape the high-energy barriers and only optimize well locally.

• Non-reversible Parallel Tempering: cosine learning rates work well empirically because it mimics non-reversible parallel tempering

Application

• NLP in Finance leverages sentiment analysis to predict stock price movement based on News headlines

• DeepLight ever won SOTA for CTR prediction in Ad Serving