Skip to content
View msw1age's full-sized avatar
1 follower · 0 following

Achievements

Achievement: Arctic Code Vault Contributor

Achievements

Achievement: Arctic Code Vault Contributor

Block or report msw1age

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
msw1age/README.md

Matthew Helton

Ph.D. candidate in Mathematics at NYU's Courant Institute (advisor: Lai-Sang Young) and NSF Graduate Research Fellow. I build large stochastic models of cortex, and the GPU and statistical tooling that keeps them fast and correct.

Looking for a summer 2027 research internship in machine learning, quantitative research or scientific computing.

matthewhelton.com · matt.helton@nyu.edu · LinkedIn

Current work

  • Spiking network model of macaque V1. 136,000 conductance-based neurons and 36 million synapses (excitatory, PV and SOM cells, short-term plasticity), built to test how SOM interneurons shape surround suppression. Manuscript in preparation; code to be released with the paper.
  • Validated GPU acceleration. I directed AI coding agents to optimize and port the simulator: 814 s → 183 s per simulation on one CPU core → about 8.5 ms on a laptop GPU (with a validated shorter warm-up), roughly 68,000× faster per simulated second. Every change was gated by bitwise or statistical checks against a validated reference; the GPU port matched the CPU code on all 77 per-ring readouts.
  • Bayesian calibration. About 45 parameters fit by staged Bayesian optimization (a GPU implementation of TPE) on a validated reduced surrogate.

Projects

Repository What it is
plainclaude Windows reading aid that rewrites Claude Desktop replies in plain language with a local LLM (Ollama), read from the accessibility tree.
SSL_Vision DINO self-supervised ViT-S/8 trained from scratch on 1.95M images; 6th of 39 teams (NYU Deep Learning, 2025).
TD_AE Time-delay autoencoder with linear (Koopman) latent dynamics for forecasting neural data; tested on Lorenz-96 and widefield imaging.
LND_notebooks Dynamic mode decomposition (exact, optimized, multiresolution and Hankel DMD) of mouse widefield and Neuropixels recordings.
V1 Early 400-neuron Numba prototype of the V1 model, with a neural-network surrogate for parameter search.

Tools

Python (PyTorch, NumPy, SciPy, Numba, scikit-learn, pandas) · GPU simulation · Bayesian optimization · statistical testing · SQL · Git · LaTeX

Pinned Loading

  1. LND_notebooks LND_notebooks Public

    Dynamic mode decomposition (exact, optimized, multiresolution, Hankel) of mouse widefield and Neuropixels data

    Jupyter Notebook

  2. TD_AE TD_AE Public

    Time-delay autoencoder with linear (Koopman) latent dynamics for forecasting neural data

    Jupyter Notebook

  3. V1 V1 Public

    Early Numba prototype of a spiking E/PV/SOM network model of V1 surround suppression, with a neural-network surrogate

    Jupyter Notebook