Machine Learning

I designed and performed a transient SPICE simulation of a circuit capable of training a multi-layer perceptron fully in analog, using the Skywater 130nm PDK. It uses crossbar arrays built out of modified 2T1C dram cells to perform (nonlinear) matmuls and opamps as activation functions. Exact gradient updates over multiple layers (taking into account device non-idealities) are computed and applied fully in analog, using the equilibrium propagation algorithm.

Training SOTA imitation learning algorithms, like diffusion policy and ACT, has become easier with recent open source libraries like Huggingface’s LeRobot. But building the pipeline to teleoperate an arbitrary pair of robot arms in real time, recording demonstrations for imitation learning in an intuitive way, requires some expertise to set up. I built a browser app to make this easy: WebXR-based handtracking, connecting via WebRTC to you laptop, which solves the inverse kinematics for arbirary URDFs in WASM in the browser, exporting trajectories containing movements and webcam-vision that are ready for imitation learning. All you need to write is a little python server forwarding joint angles from localhost to your robot.

This paper develops the statistical theory (i.e. derives convergence rates and asymptotic normality) of adversarial estimators, which estimate quantities of interest by pitting two models against each other. The paper proves traditional statistical guarantees even for semiparametric settings involving deep neural networks, and points out interesting information-theoretic connections between various estimators: from traditional econometric methods such as Empirical Likelihood, over Generative AI methods such as GANs, to modern nonparametric causal inference estimators.

Objective benchmarks based on real-world datasets have been crucial for methodological progress in machine learning. These do not exist in causal inference, which aims to predict counterfactuals that are not observed in the data. Here, we show that GANs can be used to learn data generating processes which closely resemble real data in terms of observables, while containing the unobserved counterfactuals necessary to evaluate causal inference methods. We show how causal inference practitioners can use this approach to evaluate methodological progress, and to select appropriate methods for specific data sets.