Schedule of Events | Symposia

Modeling biophysically interpretable meso-scale latent dynamics with filtered point processes

Poster Session F - Tuesday, April 1, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Patrick Bloniasz¹ (pblonias@bu.edu), Shohei Oyama¹, Emily Stephen¹; ¹Boston University

Quantifying low-dimensional latent dynamics of neural systems is a powerful approach for explaining information processing across scales. However, classic unsupervised latent state estimation approaches struggle to identify the underlying network elements generating the dynamics (e.g., different cell or synapse types). We propose a generative modeling framework to complement these methods by modeling relevant network elements as latent stochastic processes. To gain information about latent network elements from data, we exploit an increasingly recognized property of electrophysiological recordings (0–5000 Hz): the time-varying broadband power spectral shape contains rich information about firing rates of different cell type populations and synaptic activity in the tissue. For example, different cell types predictably affect power in the spike-frequency range (>2000 Hz) based on their extracellular waveform shapes, and excitatory/inhibitory synaptic dynamics affect the local field potential (LFP) frequency range (<200 Hz). Building on prior "kernel" methods modeling recordings as sums of filtered point processes (FPPs), our framework assigns a latent subprocess to each event type and links their dynamics to the (cross-)spectrogram of the recordings. We can use these models in both forward (testing whether proposed subprocesses explain spectral effects) and inverse directions (fitting latent subprocess dynamics from an observed spectrogram). We demonstrate the FPP model's value by their ability to capture spectral effects across timescales (e.g., sub-second cross-frequency coupling), interface with existing forward models (e.g., biophysical models), and provide interpretable estimation of latent states (e.g., generalized linear models; GLMs).

Topic Area: METHODS: Electrophysiology