Research Interests

Selected Research

A Geometric View of Why Guidance Improves Diffusion Samples

Research project · Generative modeling, geometry, diffusion models

Developing a geometric explanation for why classifier-free guidance improves diffusion samples. The central idea is that unguided diffusion may favor smoother, easier-to-model regions of the data distribution, while underrepresenting high-curvature or boundary-like regions where the score field varies more rapidly. I study this effect on controlled manifolds such as the catenoid and test whether guided image samples move toward higher-curvature regions in stable and DINO representation spaces.

Implicit Augmentation from Distributional Symmetry in Turbulence Super-Resolution

NeurIPS Workshop paper · ML for physics, turbulence, equivariance, super-resolution

Studied whether convolutional neural networks can acquire rotational equivariance implicitly from turbulence data, without explicit rotation augmentation or specialized equivariant architectures. We found that statistical isotropy in turbulent flows acts as a natural source of augmentation: models trained on more isotropic mid-plane data exhibit lower equivariance error than those trained near anisotropic boundary layers, and increasing temporal or spatial sampling further improves equivariance.

Research Notes / Blog Posts

Mixing Is Not Meaning: Evaluating Soft-Token Reasoning in LLMs

Research project · Language models, representation learning, continuous reasoning

Studied whether language models can meaningfully interpret soft tokens formed as mixtures of token embeddings. We found that pretrained models often collapse these mixtures to their dominant discrete component, suggesting that naive embedding interpolation does not preserve distributional meaning. We introduced diagnostic probes for soft-token understanding and explored preliminary soft-token fine-tuning objectives.

Actions That Do Nothing: A Small Pitfall in Reinforcement Learning

Blog post · Reinforcement learning, action masking, 2048

A short case study from training reinforcement learning agents for 2048. The main lesson is that actions that leave the state unchanged can create surprisingly bad learning dynamics, especially for value-based methods. I discuss why simple reward penalties may not be enough, how action masking helps, and how this connects to invalid-action masking and action-elimination ideas in reinforcement learning.

Course Projects

Encrypt What Matters: ROI-Guided FHE for CNN Inference

Research project · FHE, privacy-preserving ML, efficient encrypted inference

Course project for MIT 6.5610 on privacy-preserving ML inference. We explored a region-of-interest approach to fully homomorphic encrypted CNN inference, encrypting only sensitive input regions and using CNN locality to reduce encrypted computation.

Writing

Coming soon

I am currently preparing blog posts on geometry for ML, information theory for ML, and the role of divergences as learning objectives.

I will be a teaching assistant for Reinforcement Learning (6.7920), and I am writing course notes to make the material more structured and accessible.

Contact

The best way to reach me is by email at abackour@mit.edu.