Michelle Miller, Christoph Miehl, Brent Doiron

University of Chicago

Poster at SfN 2024

Recurrent circuits with plasticity are prone to instabilities, thus we aimed to characterize the stability of an EI rate network with E to E and I to E plasticity

We identify a line attractor in weight phase space and characterize it’s stability in terms of network parameters. We derive conditions of stability for the plasticity threshold, and identify a regime in which the threshold can be unbounded.

However, even upon finding stable circuits, when translated to linear Poisson spiking networks, we find these stable dynamics do not occur. We see this occurs due to the covariability in the network and derive the drift terms using mean field equations for a linear poisson model of neurons. We further find that by sharing the noise between the E and I population, this drift is mitigated and synaptic weights can be stably learned

Austin M. Baggetta*, Michelle Miller*, William Mau, Matthew J. Tilley, Zhe Dong, Brian M. Sweis,Denisse Morales-Rodriguez, Zachary T. Pennington, Taylor R. Francisco,Mark G. Baxter. David J. Freedman, Tristan Shuman, Denise J. Cai1

*indicates equal contributions

Poster at Society for Neuroscience 2023

Memory-updating (e.g., updating spatial maps with new reward locations) is critical in dynamic environments as information changes across time. Using Miniscope calcium imaging, we identified neuronal ensembles (co-active neurons) in dorsal CA1 that were spatially tuned and stable across training sessions. When reward locations were moved during a reversal session, a subset of these ensembles decreased their activation strength (i.e., fading ensembles) which correlated with enhanced memory-updating. Middle-aged mice had impaired memory-updating (i.e., reversal learning) which correlated with their deficit in fading ensembles. We developed an artificial neural network, BioANN, to causally test our experimental observations and develop new predictions. This BioANN relies on an RL framework to simulate and agent going through the water port task across sessions, including an updating session akin to the experiment. We found simulated ensembles in the model and found that the task performance correlated with the fading ensembles

Matthew J Tilley** Michelle Miller** Austin Baggetta, Zhe Dong, B. M. Sweiss,Denisse. M. Morales-Rodriguez, David J. Freedman, Denise Cai

**implies equal contributio

Poster at Society for Neuroscience 2023

Biological neural networks can recruit different sets of neurons to encode different memories.

• When training artificial neural networks on a set of tasks, typically, no mechanism is employed for selectively producing anything analogous to these neuronal ensembles.

• Artificial neural networks suffer from catastrophic forgetting, where the network’s performance rapidly deteriorates as tasks are learned sequentially.

• By contrast, sequential learning is possible for a range of biological organisms.

• We introduce Learned Context Dependent Gating (LXDG), a method to flexibly allocate and recall ‘artificial neuronal ensembles’, using a particular network structure and a new set of regularization terms.

Michelle Miller, Johnathan Pober

American Astronomical Society Conference 2017

After the Big Bang the universe was filled with neutral hydrogen that began to cool and collapse into the first structures. These first stars and galaxies began to emit radiation that eventually caused all of the neutral hydrogen in the universe to ionize. However, not much is known about the properties of these first sources. A large goal of studying Reionization is two better constrain how properties of these objects like their masses, photon efficiencies, and distributions within the universe itself. For example, we do not know the precise role galaxies play in this era as a opposed to individual stars. I worked with 21CMMC, a semi-numerical Monte Carlo Markov Chain code that takes simulated boxes of the ionized early universe from another code called 21cmFAST, which simulates a mini universe Mock measurements from 21cmFAST or other simulations. Those measurements are thrown into 21CMMC which will go through the same process as 21cmFAST. Using power spectra from a more complex piece of code than 21cmFAST, I tried to test 21CMMCs capacity to determine the parameters in the simulated universe. We found it stuggled to converge on the mean free path parameters.

Michelle Miller , Philip Mathieu, Chloe Hequet, Johnathan Pober

Poster at the American Astronomical Society Conference (2016)

BURST was a two antenna, single-baseline, radio telescope on the roof of the physics building at Brown University. Myself and two other undergrads worked on assembling this telescope. We sed the same antennas as those proposed for SKA-Low and the same digital infrastructure (i.e. correlator) as HERA. At the time of building it can take measurement over a broad band range of frequencies from 100 to 1000Mhz. The intention at the time was to add more baselines and to narrow down a bandwidth to take measurement in.