Ring Attractor Dynamics in Small Spiking Systems
The fly head-direction system is known to encode a continuous variable in the form of a heading signal. However, such continuous representations are only understood to emerge in large networks (large N) unless requiring fine tuning. Likewise, this fine-tuning of small ring networks to produce ring attractors has only been shown in rate networks. I am investigating how ring attractor dynamics can emerge in small spiking networks, when that breaks down, and what needs to be considered to preserve the ring attractor.
How Neuronal Density Gradients Facilitate Modular, Functional Computation W
We are considering the role that cell-type gradients play in shaping reward and odor encoding in the anterior piriform cortex. In particular, we use a 3 population spiking model with local plasticity and a gradient of SST (inhibitory) neuron density along an axis. This produces a gradient of response properties and self-organized modular networks
Homeostatic Plasticity Enables Stable, Flexible, and Tunable Assemblies
Strongly interconnected neuronal populations, called assemblies, dynamically form through synaptic plasticity mechanisms and are thought to be a substrate for memories in the brain. Many assembly formation models use Hebbian excitatory-to-excitatory plasticity, where coordinated activity strengthens recurrent structure. However, these models typically yield binary assembly outcomes: networks with either weak (no assembly) or maximally strong (assembly) connectivity. We consider networks with a combination of Hebbian excitatory-to-excitatory plasticity and inhibitory-to-excitatory synapses with plasticity that homeostatically stabilizes excitatory neuron firing at a target value. When we set excitatory-to-excitatory plasticity to be homeostatically compliant, in that potentiation and depression are balanced at the homeostatic target firing rate, we find a stable continuum of synaptic strengths, and assembly structure is no longer binary. We use a recurrent network of spiking neuron models and an associated mean-field theory to identify this continuum as a line attractor in synaptic weight space. While along the attractor, homeostasis ensures that neuronal firing rates are invariant, the dynamical response properties of the network are quite malleable, with strongly coupled networks having high gain and longer timescale responses. Using our mean-field theory we show how correlated stochastic spiking activity among the excitatory neurons can destroy the line attractor, yet this can be mitigated when correlated inputs are shared across the excitatory and inhibitory neurons. Altogether, we provide an alternative learning framework based on homeostasis, where a tunable and flexible assembly structure is possible.
Ensemble remodeling supports memory-updating (with 1 layer network with RL)
[2022-2023] PhD Student UChicago in David J Freedman’s lab; Collaborating with Denise Cai’s lab at Mt. Sinai
This project was the beginning of our collaboration with the Cai lab. Experimental results from the Cai lab suggested that task-updating remodels neural ensembles. The experimental setup entails a mouse in a circle track. Along the track there are 8 possible water reward ports. On any given day, two of the ports are rewarded with water. The mice are water-deprived and enter the track every day for 20 minutes. On the fifth day, the location of the water reward switches to two new ports. To model this and ask questions about neural activity during memory updating, I built a custom 1-layer artificial neural network (ANN) to model a reinforcement learning agent in a virtual track. I studied the network activations as the model agent learns the water port task across 4 “days” in the model. On “day” 5, the ports switch to two new locations. The model re-captures the behavior of the real animal by reaching peak hit rate, correct rejection rate, and discriminability index across sessions, which then drop and begin to improve again upon task switching. The network also exhibits reward port over-representation. I found the model relies on a stable spatial representation of co-active neurons to be able to modulate the output activity to action. Further, I find a large proportion of co-active neurons which “fade” in their co-activity upon task updating. We are in the process of writing this result in a paper, which we hope to discern whether these fading ensembles represent reward representations switching.
Results: Paper in Progress Pending re-analysis of experimental data and a Poster at SfN 2023
Artificial Neuronal Ensembles with Learned Context Dependent Gating
[2022] Rotating PhD student in David J Freedman’s lab at University of Chicago
I began this project during a rotation in the Freedman lab prior to joining. Biological neural networks are
capable of recruiting subsets of neurons to encode different memories. However, when training artificial NNs (ANNs) on a set of tasks, because ANNs have no such mechanism, they suffer from catastrophic forgetting, in which their performance rapidly deteriorates as tasks are learned sequentially. We expanded upon a prior continual learning model called Context Dependent Gating (XDG) in which subnetworks of weights are allocated in a random way. We call our method Learned Context Dependent Gating (LXDG). We introduced 3 new regularization terms, allowing subnetworks to be learned by supporting changing of old weights for new tasks, keeping old weights, and maintaining sparsity in the network. We found the model was able to learn how to allocate subnetworks of neurons effectively relative to control models on a continual learning benchmark called rotated MNIST. This method effectively mitigated catastrophic forgetting. This produced a paper which was accepted by ICLR (2023) and a poster at the Society for Neuroscience (SfN 2022.
Result: Paper accepted by ICLR 2023 and Poster at SfN 2022
Skills: model architecture development, and weight optimization. .
Investigation of Divisive Normalization (DN) in Image Classification
[2019-2021] Research Assistant in Ken Miller’s Lab Columbia University, Center for Theoretical Neuroscience
A year after college, I joined Ken D. Miller’s lab upon deciding I wanted to switch into computational neuroscience. While there, I investigated divisive normalization (DN), a phenomenon observed in the cortex where neuron responses can be inhibited by nearby neurons (e.g. lateral inhibition). My work involved incorporating a custom DN technique into a 5-layer convolutional NN, aiming to understand its impact on performance and learning. The key findings of my research, which resulted in a paper at the International Conference for Learned Representations (ICLR), revealed that DN differs from traditional normalization methods, such as Batch, Group, and Layer norm, by inducing competition among neurons. Through parameter tuning, I achieved optimal results and compared the DN model to models with combinations of normalization, which outperformed on both ImageNet and CIFAR100 datasets. Furthermore, I explored changes in neural manifolds, observed increased sparsity influenced by DN, and identified shifts in the radial profile of Fourier power, potentially selecting for large-scale structures. This project honed my skills in customizing NN architectures, analyzing their activity, and presenting research findings effectively at conferences.
Result:
● Paper accepted for ICLR 2022; Attended Cold Spring Harbour NAISys conference 2020 (online)
Effect of Space Charge on Oscillation Frequency Distributions
[2018-2021] Accelerator Physics Masters Student: advised by Steven Lund through Indiana University/USPAS
Upon graduating from Brown, I was offered a scholarship to continue the US Particle Accelerator School (USPAS) and to enroll in their joint Master’s in Physics with Indiana University, which I had been attending part time since freshman year. It’s a unique intensive program offering accelerator physics courses every 6 months around the country. For the thesis project,my research focused on understanding the behavior of particle beams in high-intensity particle accelerators. I utilized a particle-in-cell code to study the effect of different distributions of space charge (e.g. charged particles) on the stability of the particle beam. I coded (in Python) a custom Fast Fourier Transform to accumulate the particle trajectories. Through these simulations, I found that when these particle forces are very strong, it causes the particle beam to become more stable, as opposed to unstable, which was historically thought to be the case. This research is important for accelerator physics, as it showed that high magnitude electromagnetic forces in accelerators can actually help stabilize the particle beams, preventing what is called “beam blowup”. It provides valuable insights into how we can control and optimize these high-intensity particle accelerators, which are vital tools in the world of physics.
Result:
Masters Thesis : Effect of Self-Consistent Space Charge on the Distribution of Particle Oscillation Frequencies in Continuously Focused Beams
Gained skills working with particle-in-cell sims running simulation tests; Custom algorithm development.
Comparative Resilience of Machine Learning Models in Inference
[2019] Summer Research Student: Los Alamos National Laboratory
In order to make the transition to work in computational neuroscience, I attended the Radiation Effects Summer School at Los Alamos National Laboratory. I loaded a neural network (NN) model (MobileNet) trained on ImageNet onto a Movidius Intel chip interfacing with an NVIDIA GPU and CPU. We irradiated the configuration with 14.1MeV neutrons to test the resilience of the model during the inference stage of image classification. My goal was to test whether different modifications to the NN architecture would make it more resilient to the single-event neutron radiation damage, which would be informative for ascertaining relevant architectures. Specifically, I tested whether dropout is a useful regularization method to mitigate these radiation effects, which we found did not statistically mitigate these effects relative to the control.
Result:
Wrote up project summary and presented work at end of summer symposium to division-wide scientists
Gained skills for optimizing neural network architectures and motifs which are robust to radiation effects