arXiv subjects

Spencer LaVere Smith

Publications and source records attributed to Spencer LaVere Smith.

4 recordsLinked to original sources

arXiv2026

Attenuation scaling and error analysis of 2f and 4f architectures for free-space optical matrix-vector multiplication

Free-space optical computing has been suggested as a scalable, high speed, and energy efficient platform for performing matrix-vector multiplication (MVM). We present two free-space optical approaches for MVM, called the 2f and 4f architectures, and model them using wave optics simulations. After constraining the optical modulator in our models to have a maximum gain limit, we use our simulations to compare 2f and 4f MVM performance in terms of computational error and optical signal attenuation per MVM. We examine how 2f and 4f signal attenuation per MVM scales with increasing MVM problem size for different statistical distributions of matrix elements and compare to the expected attenuation from a universal multiport interferometer (UMI), commonly used in integrated photonics for MVM. We find that the 2f and 4f architectures scale more favorably to large problem sizes, experiencing many orders of magnitude less attenuation than UMIs for matrix dimension above a thousand elements. We furthermore examine how varying modulator space-bandwidth product and output slit aperture affect 2f and 4f attenuation and computational error across different distributions of matrix elements. We conclude that the preference of 2f or 4f MVM depends on the statistics of the matrix used, but that 4f may provide more flexibility than 2f.

physics.optics↗

arXiv2025

Mouse vs. AI: A Neuroethological Benchmark for Visual Robustness and Neural Alignment

Visual robustness under real-world conditions remains a critical bottleneck for modern reinforcement learning agents. In contrast, biological systems such as mice show remarkable resilience to environmental changes, maintaining stable performance even under degraded visual input with minimal exposure. Inspired by this gap, we propose the Mouse vs. AI: Robust Foraging Competition, a novel bioinspired visual robustness benchmark to test generalization in reinforcement learning (RL) agents trained to navigate a virtual environment toward a visually cued target. Participants train agents to perform a visually guided foraging task in a naturalistic 3D Unity environment and are evaluated on their ability to generalize to unseen, ecologically realistic visual perturbations. What sets this challenge apart is its biological grounding: real mice performed the same task, and participants receive both behavioral performance data and large-scale neural recordings (over 19,000 neurons across visual cortex) for benchmarking. The competition features two tracks: (1) Visual Robustness, assessing generalization across held-out visual perturbations; and (2) Neural Alignment, evaluating how well agents' internal representations predict mouse visual cortical activity via a linear readout. We provide the full Unity environment, a fog-perturbed training condition for validation, baseline proximal policy optimization (PPO) agents, and a rich multimodal dataset. By bridging reinforcement learning, computer vision, and neuroscience through a shared, behaviorally grounded task, this challenge advances the development of robust, generalizable, and biologically inspired AI.

q-bio.NC↗

arXiv2025

Tracking the topology of neural manifolds across populations

Neural manifolds summarize the intrinsic structure of the information encoded by a population of neurons. Advances in experimental techniques have made simultaneous recordings from multiple brain regions increasingly commonplace, raising the possibility of studying how these manifolds relate across populations. However, when the manifolds are nonlinear and possibly code for multiple unknown variables, it is challenging to extract robust and falsifiable information about their relationships. We introduce a framework, called the method of analogous cycles, for matching topological features of neural manifolds using only observed dissimilarity matrices within and between neural populations. We demonstrate via analysis of simulations and \emph{in vivo} experimental data that this method can be used to correctly identify multiple shared circular coordinate systems across both stimuli and inferred neural manifolds. Conversely, the method rejects matching features that are not intrinsic to one of the systems. Further, as this method is deterministic and does not rely on dimensionality reduction or optimization methods, it is amenable to direct mathematical investigation and interpretation in terms of the underlying neural activity. We thus propose the method of analogous cycles as a suitable foundation for a theory of cross-population analysis via neural manifolds.

q-bio.NC↗

arXiv2024

Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data

State-of-the-art systems neuroscience experiments yield large-scale multimodal data, and these data sets require new tools for analysis. Inspired by the success of large pretrained models in vision and language domains, we reframe the analysis of large-scale, cellular-resolution neuronal spiking data into an autoregressive spatiotemporal generation problem. Neuroformer is a multimodal, multitask generative pretrained transformer (GPT) model that is specifically designed to handle the intricacies of data in systems neuroscience. It scales linearly with feature size, can process an arbitrary number of modalities, and is adaptable to downstream tasks, such as predicting behavior. We first trained Neuroformer on simulated datasets, and found that it both accurately predicted simulated neuronal circuit activity, and also intrinsically inferred the underlying neural circuit connectivity, including direction. When pretrained to decode neural responses, the model predicted the behavior of a mouse with only few-shot fine-tuning, suggesting that the model begins learning how to do so directly from the neural representations themselves, without any explicit supervision. We used an ablation study to show that joint training on neuronal responses and behavior boosted performance, highlighting the model's ability to associate behavioral and neural representations in an unsupervised manner. These findings show that Neuroformer can analyze neural datasets and their emergent properties, informing the development of models and hypotheses associated with the brain.

q-bio.NC↗