Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics
Authors: William Qian, Jacob Zavatone-Veth, Ben Ruben, Cengiz Pehlevan
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here we show that observing only a subset of neurons in a circuit can create mechanistic mismatches between a simulated teacher network and a data-constrained student, even when the two networks have matching single-unit dynamics. In particular, partial observation of models of low-dimensional cortical dynamics based on functionally feedforward or low-rank connectivity can lead to surrogate models with spurious attractor structure. |
| Researcher Affiliation | Academia | William Qian1,2, Jacob A. Zavatone-Veth3,4,5, Benjamin S. Ruben1, Cengiz Pehlevan2,3,4 1Biophysics Graduate Program, 2Kempner Institute for the Study of Natural and Artificial Intelligence, 3John A. Paulson School of Engineering and Applied Sciences, 4Center for Brain Science, 5Department of Physics, Harvard University Cambridge, MA 02138 |
| Pseudocode | No | No pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | Yes | All code is available at https://github.com/wqian0/Data Constrained RNNs/. |
| Open Datasets | No | We constructed a model sensory integration task, which networks of both architectures could effectively solve (Fig. 1). |
| Dataset Splits | No | In the examples of Fig. 1, we generate ground truth network activity by iterating the dynamics for a duration T = 5000Δt. |
| Hardware Specification | Yes | They were not computationally-intensive, and required less than 12 hours in total to run on a consumer Dell XPS laptop equipped with an Intel Core i7-13700H processor. |
| Software Dependencies | Yes | All of our numerical simulations are implemented in Python 3.9.18 using Num Py 1.26.2 [68], Sci Py [69], and Py Torch [70]. |
| Experiment Setup | Yes | We integrate the student and teacher RNN dynamics via Euler integration with a timestep Δt = 0.01. Under the discretization scheme of B.4, in all experiments, we select the noise parameters of the student and teacher dynamics as ση = σξ = 0.02. For MAP inference, we use a regularization parameter ρ = 0.001 in all experiments. |