Predictive coding in balanced neural networks with noise, chaos and delays
Authors: Jonathan Kadmon, Jonathan Timcheck, Surya Ganguli
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare the theoretically predicted decoder bias ˆx x and variance (δˆx)2 with numerical experiments, obtaining an excellent match (see Fig. 1D and Figures below). |
| Researcher Affiliation | Academia | Jonathan Kadmon Department of Applied Physics Stanford University,CA kadmonj@stanford.edu Jonathan Timcheck Department of Physics Stanford University,CA Surya Ganguli Department of Applied Physics Stanford University, CA |
| Pseudocode | No | The paper describes the model and derivations using mathematical equations (e.g., Equation 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13) and prose, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any statements about releasing open-source code, nor does it include links to a code repository. |
| Open Datasets | No | The paper describes numerical simulations of a theoretically tractable neural network model, not experiments conducted on a publicly available dataset. Therefore, there is no mention of a dataset being publicly available or access information provided. |
| Dataset Splits | No | The paper performs numerical simulations of a theoretical model rather than training on and evaluating against external datasets. Consequently, it does not describe training, validation, or test dataset splits. |
| Hardware Specification | No | The paper does not specify the hardware (e.g., GPU models, CPU types) used for running the numerical simulations. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, or scientific computing libraries) used for the numerical simulations. |
| Experiment Setup | Yes | We compare the theoretically predicted decoder bias ˆx x and variance (δˆx)2 with numerical experiments, obtaining an excellent match (see Fig. 1D and Figures below). (C) The mean input-output transfer function ˆx as a function of x obtained by solving (6) (solid curves) and numerical simulations of (3) (points) with N = 1400, σ = 0.75 and g = d = 0 for 3 values of b. (D) The decoder bias |x ˆx | (top) and standard deviation p (δˆx)2 (bottom) as a function of balance b for theory (curves) and simulations (points). σ = 0.75, g = 0 for noise (blue), σ = 0, g = 1.6 for chaos (orange). In both cases N = 1400 and x = 0.2. (Figure 2) Left: The critical balance bc (blue curve) as a function of the delay (with σ2 = 2). Right: sample firing rates ri(t) (grey) from simulations of (3) with N = 1000, with parameters corresponding to points in the left two panels. (Figure 3) Optimally balanced network with delay, φ = tanh and x = 0. Points reflect simulations of (3) with N = 1400 and curves reflect theory. Left: Decoder standard deviation ( p N) as a function of balance b with σ = 0.75. Right: Same as center but with deterministic chaos (g = 1.6, σ = 0). |