Hayson Cheung

A light budget head per layer decides how many of the top-8 routed experts a token needs rather than which: distilled against the unmodified model, it cuts 8 active experts to 2.5 on WikiText and 5.8 in chat at router KL ≈ 0.05. Trained for Gemma 4 and DiffusionGemma. Achieved a near-lossless performance with 69% parameter reduction

Surface EMG decoders degrade across users because electrode placement and muscle anatomy shift the input distribution. Conditioning a pose backbone on a learned per-user embedding absorbs that shift, beating the state of the art on all three EMG2POSE splits from under 45 seconds of calibration and with no retraining at deployment time.

A neural surrogate for flows where advection and diffusion both matter. Splitting the dynamics into a deterministic neural ODE and a variational stochastic controller keeps the mean field physically consistent while still modelling the noise, which is worth an 80–96% accuracy gain over the purely deterministic baseline across flow regimes.

Accompaniment has to track a vocal line locally while staying coherent globally. A soft alignment attention that reweights local against global temporal dependencies as a function of the diffusion timestep gets both from a 15M-parameter latent diffusion model: 220× fewer parameters and 52× faster inference at competitive quality.

Predictive representations for embodied agents: learning latent dynamics and planning in a compressed state space.

Tree search depth at an exponential cost, the alternative here is to learn the value landscape as an energy over a continuous manifold of positions, so a plan comes from gradient descent on that energy instead of from expanding a frontier: branching becomes a fixed number of refinement steps, and those steps batch on a GPU where a search serializes.

NEAT grows topology and weights together from a minimal network, and training is done via evolution of genomes that defined the networks. I had it play smash bros and pongs.

Test-Image-to-3D generation lifts a photo into a 3D model, and structured decoding turns the resulting geometry into an ordered, step-by-step build sequence. First thing that got me into 3D generation.

A sequence-to-sequence translator built around an LSTM bottleneck, first thing that got me into represenation learning and neuro networks.

Real time tracking of pendulum from video, where the bob is found by cosine similarity in RGB and cross frame physical heuristics. Period and damping are then read off the trajectory. First-year physics project, and the first thing that got me into computer vision.