Mozes Jacobs

I am a third-year computer science PhD candidate at Harvard University, advised by Professor Demba Ba. I am supported by the Kempner Institute Graduate Fellowship.

Representation LearningInterpretabilityFoundation Models

I study the foundations of deep learning, with a focus on representation learning and interpretability. My research aims to understand what neural networks learn and the principles that govern why those representations emerge. I pursue both a mechanistic understanding of how networks compute and a normative understanding of why they learn what they do. Ultimately, I aim to use these insights to improve the efficiency, generalization, and safety of neural systems.

Previously, I worked at the AI Institute in Dynamic Systems with Nathan Kutz and Ryan Raut. I earned my B.S. in computer science from the Allen School at the University of Washington, where I worked with Rajesh Rao and William Noble.

Email  /  CV  /  Google Scholar  /  Twitter  /  LinkedIn

* denotes equal contribution

Attention sinks share a visual signature but hide two distinct algorithms: nop, where a head suppresses its update by routing to a null token, and broadcast, where a sink aggregates and redistributes global information. Each mechanism leaves distinct traces — nop sinks have negligible value norms; broadcast sinks induce low-rank outputs — which we use to derive practical diagnostics. Applied to pretrained vision transformers, we find both mechanisms coexist at scale. Gating and registers, the two dominant interventions, each implicitly target only one mechanism; combining them yields complementary gains. Training LeJepa with both gating and registers improves downstream semantic segmentation performance beyond either alone.

We introduce the Block-Recurrent Hypothesis (BRH), arguing that trained ViTs admit a block-recurrent depth structure. To validate this, we train recurrent surrogates called Raptor. We demonstrate that a Raptor model can recover 96% of DINOv2 ImageNet-1k linear probe accuracy in only 2 blocks while maintaining equivalent runtime. We leverage our hypothesis to perform dynamical interpretability, revealing directional convergence into class-dependent basins, token-specific trajectory dynamics, and low-rank attractor structure in late layers.

We investigate how traveling waves of neural activity enable spatial information integration in convolutional recurrent networks. Our models learn to generate traveling waves in response to visual stimuli, effectively expanding receptive fields of locally connected neurons. This mechanism significantly outperforms local feed-forward networks on semantic segmentation tasks requiring global spatial context, achieving comparable performance to non-local U-Nets while using significantly fewer parameters.

HyperSINDy is a deep generative framework for discovering stochastic governing equations from data. A variational encoder and hypernetwork produce sparse differential equations — learned via a trainable binary mask — whose coefficients are driven by Gaussian white noise. HyperSINDy accurately recovers ground-truth stochastic dynamics and provides uncertainty quantification that scales to high-dimensional systems.

Website template from Jon Barron