Projects

Helped implement Didec, which discovers reusable options (temporally extended skills) by masking parts of a trained neural policy.

A careful re-evaluation of the claim that programmatic policies generalize better than neural ones on out-of-distribution tasks.

Designed a quality metric that builds on multi-modal foundation models (CLIP-IQA, Q-Align) and outperforms state-of-the-art baselines.

A human-aligned colour metric that fixes inconsistencies in ΔE. Using it in the on-device colour-correction pipeline improved performance by more than 20%.

Tuned image-signal-processing pipelines for edge devices with genetic algorithms, sample-efficient Bayesian optimization and reinforcement learning.

Built a fusion-aware component for the tensor-optimization pipeline (at least 10% better) and applied Genetic MCTS to CompilerGym tasks. As an intern, modelled uncertainty in Halide optimization with heteroscedastic regression.

Changed each component of MCTS so it takes errors in the transition model into account. With a good error estimate it can match MCTS with a perfect simulator.

Studied whether heteroscedastic regression gives a reliable estimate of model uncertainty.

How step size and performance interact for different values of λ.

Applied variance-reduction methods to off-policy policy evaluation.

How sequence-tagging models handle words they have never seen.

Translated MS-COCO captions into Persian and trained an attention-based (Show, Attend and Tell) captioning model on them.

Collected and analysed a large word-association dataset for Persian through a survey website.