Loading...

Research & projects

Selected work in machine learning research and research-oriented engineering.

Research experience

Stochastic optimization for discrete world models

Intelligent Dynamics Lab · Advisor: Roy Fox

Developed a PyTorch library of gradient estimators and control variates. I also replaced straight-through estimators in DreamerV3 with Gumbel-Softmax relaxations for discrete latents and actions.

PyTorchReinforcement learningOptimization

Double descent and model merging

Machine Learning Research Group · Advisor: Alexander Ihler

Studied how scale, distillation, and regularization affect minima flatness and generalization. I also merged pruned CNNs by combining similar filters and evaluating unique filters through grouped convolutions.

GeneralizationCNNsModel compression

Surrogate models for thermal simulation

California Institute for Telecommunications and Information Technology

Developed physics-informed neural networks as efficient surrogates for finite-element simulations used to evaluate thermal-storage systems.

Scientific MLSimulationPython

Selected projects

Molecular generation

Deep generative models · 2026

Built a graph variational autoencoder with a property-prediction head, inverse autoregressive flow, and autoregressive transformer decoder to generate drug-like molecules.

Result: Fréchet ChemNet Distance of 1.81 on ZINC-250k.

Graph VAETransformersNormalizing flows

Semi-supervised code translation

Programming language translation · 2026

Fine-tuned CodeT5 for Python–C++ translation using paired examples alongside backtranslation and denoising autoencoding on unpaired code.

Result: Approximately 76% CodeBLEU and 20% HumanEval-X.

CodeT5Semi-supervised learningNLP

More implementation details and experiments are available on GitHub , with a complete experience history in my Resume .

© 2026 Elian Hijmans Malessy. All Rights Reserved.