Alec Hoyland, Journeyman Neuromancer

I’m a computational scientist at Yurts AI and a PhD student in Biomedical Engineering at Worcester Polytechnic Institute. At Yurts, I develop and deploy the Yurts platform, an LLM technology stack that automagically ingests and indexes factual documents, enabling an LLM to use knowledge from those documents and reducing hallucinations. At WPI, I work in The Brain, Behavior, and Computation Lab, where I use advanced signal processing and machine learning to improve biosignal data collection and analysis.

My research interests include autoencoders, vector embeddings, parameter optimization, signal processing, and differentiable programming.

Table of Contents

Contact Me

alec.hoyland@posteo.net | GitHub | LinkedIn | CV

Code & Projects

Unfortunately, a lot of the work I've done is classified or proprietary, but here are some snapshots of open-source projects I've worked on.

TinnitusReconstructor.jl

Most people think of tinnitus as "ringing in the ears," but the space of tinnitus percepts is much more diverse than that. About 20-50% of tinnitus percepts are non-tonal sounds. There's strong evidence that knowing a patient's tinnitus percept can help treat the condition via sound therapy.

We devised a reverse correlation experiment where subjects listen to randomized noise to see if they recognize their tinnitus sound within it. We can use compressed sensing and regression algorithms to reconstruct detailed spectral representations of what their tinnitus percept sounds like.

See also: tinnitus-reconstruction.

xolotl

Xolotl is a fast single- and multi-compartment neuronal network simulator written in C++ with a MATLAB interface you'll actually like. We built it after being disappointed by the neuronal network simulation software already out there.

Since we were simulating thousands of models with hundreds of stiff equations, we needed something fast, and we wanted it to be usable too. As far as I am aware, xolotl is still the fastest simulator out there, though Conductor.jl is only a bit slower. if you take the guardrails off.

Xolotl has built in parallelization and parameter optimization, as well as interactivity for hand-tuning parameters live.

Hand-tuning parameters using sliders

ReducedOrderModelsProject.jl

In this project, I derive a formulation for a reduced order model that exploits the structure of a partial differential equation to solve it at high accuracy at arbitrary parameter values, using far fewer FLOPs to do so than with traditional linear system solvers, numerical simulation, or approximation by neural networks (surrogates).

A reduced order model (ROM) is a smaller model that mimics the results of a larger model. Often in mathematical modeling and numerical simulation, you run into issues with increasing computational complexity where the state spaces gets too large, or numerical simulations become infeasible. ROMs alleviate this problem by providing a smaller model that can be used in simulations or numerical computations instead. A good ROM has a small approximation error compared to the full order model and conserves the properties of the larger model.

I construct a rational Krylov subspace, where the vectors in the subspace are solutions to the partial differential equation I'm simplifying at different parameter values. A solution for an arbitrary parameter value is a linear combination of the solutions from the subspace. Thus, the Krylov subspace forms a basis for the solutions of the PDE system.

It turns out that we don’t need to know all the vectors in the subspace; we only need about 10. The ROM is constructed by looking at the differences between solutions and accurately captures the dynamics with only a few examples. This method has far better performance than a neural network function approximator because vanilla neural networks don’t exploit the known dynamics for the system. Here, we can write the equation of state, so we can take advantage of them directly.

Audio-ML

I developed some demos of supervised and unsupervised machine learning for audio using PyTorch.

In one notebook, I developed an unsupervised ML model for speech detection using open-unmix for denoising, Mel-frequency cepstral coefficients for features, dimensionality reduction using UMAP, and clustering using k-means.

I also use whispercpp to transcribe speech.

In a literate programming script, I build a convolutional autoencoder that uses 2-D convol ution and max pooling to reduce the dimensionality of an audio signal (represented as a Mel-frequency spectrogram) in the latent space.

MLP-Demo

I wrote this demo for middle school after-school program students who were interested in robotics and machine learning. We got to play around with it interactively as they tried to figure out how it worked and how to break it.

BandwidthEstimator

This package implements a maximum-likelihood leave-one-out cross-validated bandwidth parameter estimation for smoothing a neural spike train. The bandwidth parameter tells you what intrinsic time-scale the spike train encodes.

This package works with RatCatcher, which is a MATLAB utility for parsing data and passing analysis scripts from a local machine to a high-performance computing cluster to run in parallel, before re-stitching the analyses together afterwards.

Writings

Differential Responses to Neuromodulation in Model Neurons of the Crustacean Stomatogastric Ganglion

Master of Neuroscience thesis at Brandeis University

Neuronal networks must produce stable circuit output for sustained periods of time despite environmental perturbation. In addition, they must be sensitive to key endogenous signaling to produce differing output. The STG manages these competing objectives while remaining degenerate to ion channel density. Neuromodulators can produce a diverse set of network states using the same cellular and synaptic morphology. In particular to the STG, the dense, tangled neuropil and gradations in reversal potential render neurons isopotential with respect to the somata. Neuromodulators, then, play the role of maintaining and switching network activity. For stable and responsive biological activity, degenerate networks must still be robust to environmental perturbation and responsive to intentional modulation. In this thesis, I describe red pigment-concentrating hormone (RPCH) acting as a neuromodulator on a computational model of a rhythmic motor circuit.

History of Ideas minor thesis at Brandeis University

I presented this paper at a mini-conference on the intersection of genealogy, science, and social justice held at Brandeis University. What philosophical and sociological conclusions can we gather from current neuroscientific research? This work is broken into three parts. First we consider the scientific and philosophical implications of anomalous self-experience in schizophrenia. Second, we consider an examination of agency in schizophrenic individuals. Finally, we consider the implications of the first and second sections on issues of social justice concerning sexual consent.

Publications

Bibliography
[1]
A. Hoyland et al., “Reverse Correlation Uncovers More Complete Tinnitus Spectra,” IEEE Open Journal of Engineering in Medicine and Biology, pp. 1–3, 2023, doi: 10.1109/OJEMB.2023.3275051.
[2]
A. Hoyland et al., “Characterizing Complex Tinnitus Sounds Using Reverse Correlation: A Feasibility Study,” in Association for Research in Otolaryngology, Orlando FL, Feb. 2023.
[3]
A. Hoyland et al., “Reverse Correlation Uncovers More Complete Tinnitus Spectra.” bioRxiv, p. 2022.12.23.521795, Jan. 06, 2023. doi: 10.1101/2022.12.23.521795.
[4]
H. Dannenberg, H. Lazaro, P. Nambiar, A. Hoyland, and M. E. Hasselmo, “Effects of visual inputs on neural dynamics for coding of location and running speed in medial entorhinal cortex,” eLife, vol. 9, p. e62500, Dec. 2020, doi: 10.7554/eLife.62500.
[5]
M. E. Hasselmo et al., “The Unexplored Territory of Neural Models: Potential Guides for Exploring the Function of Metabotropic Neuromodulation,” Neuroscience, Apr. 2020, doi: 10.1016/j.neuroscience.2020.03.048.
[6]
H. Dannenberg, C. Kelley, A. Hoyland, C. K. Monaghan, and M. E. Hasselmo, “Speed coding by entorhinal cortex speed cells differs across behaviorally relevant timescales and is independent of cholinergic modulation,” presented at the Society for Neuroscience, in 508.27. San Diego, CA, 2018.
[7]
W. Ning, J. H. Bladon, J. Chen, S. Steinwenter, A. Hoyland, and M. E. Hasselmo, “A cortical-hippocampal network supporting the temporal organization of memory,” in 2019 Neuroscience Meeting Planner, in 164.05. Chicago, IL, 2019.
[8]
A. Hoyland, “Differential Responses to Neuromodulation in Model Neurons of the Crustacean Stomatogastric Ganglion,” Thesis, Brandeis University, 2018. Accessed: Aug. 14, 2019. [Online]. Available: http://bir.brandeis.edu/handle/10192/35686
[9]
H. Dannenberg, C. Kelley, A. Hoyland, C. K. Monaghan, and M. E. Hasselmo, “The Firing Rate Speed Code of Entorhinal Speed Cells Differs across Behaviorally Relevant Time Scales and Does Not Depend on Medial Septum Inputs,” J. Neurosci., vol. 39, no. 18, pp. 3434–3453, May 2019, doi: 10.1523/JNEUROSCI.1450-18.2019.
[10]
S. Gorur-Shandilya, A. Hoyland, and E. Marder, “Xolotl: An Intuitive and Approachable Neuron and Network Simulator for Research and Teaching,” Front. Neuroinform., vol. 12, 2018, doi: 10.3389/fninf.2018.00087.
CC BY-SA 4.0 Alec Hoyland. Last modified: October 27, 2023. Website built with Franklin.jl and the Julia programming language.