Operative dimensions in unconstrained connectivity of recurrent neural networks
Authors: Renate Krause, Matthew Cook, Sepp Kollmorgen, Valerio Mante, Giacomo Indiveri
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here we study how network dynamics are related to network connectivity in RNNs trained without any specific constraints on several tasks previously employed in neuroscience. Despite the apparent high-dimensional connectivity of these RNNs, we show that a low-dimensional, functionally relevant subspace of the weight matrix can be found through the identification of operative dimensions, which we define as components of the connectivity whose removal has a large influence on local RNN dynamics. We perform our analyses on vanilla RNNs trained without regularization terms, using the standard RNN equation: τ xt = xt + Wrt + But + σt (1) |
| Researcher Affiliation | Academia | Institute of Neuroinformatics, University of Zurich and ETH Zurich, Zurich, Switzerland Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland |
| Pseudocode | No | The paper describes mathematical equations and procedural steps but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | All the code and models required to reproduce the main analyses supporting our conclusions are available in matlab and python at: https://gitlab.com/ neuroinf/operative Dimensions |
| Open Datasets | Yes | The RNNs are trained separately on two previously proposed tasks: context-dependent integration [19] and sine wave generation [20] (see appendix section A.3.8 for additional results on sequential MNIST). |
| Dataset Splits | Yes | Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] Exact code to reproduce results is provided, including all parameters. Additionally, details are explained in respective sections of text and appendix. |
| Hardware Specification | Yes | Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See section A.2.7. |
| Software Dependencies | No | The paper states: 'We used Matlab to perform the data analysis.' However, it does not specify a version number for Matlab or any other specific software libraries with version numbers, which is required for reproducibility. |
| Experiment Setup | Yes | We perform our analyses on vanilla RNNs trained without regularization terms, using the standard RNN equation: τ xt = xt + Wrt + But + σt (1) where xt RN are the linear activities of the N hidden units over time t with rt = tanh(xt), W RN N is the recurrent weight matrix of the hidden units and τ R is the time constant (τ = 10 ms, dt = 1ms). We consider RNNs of N = 100 noisy units, where each element of σt is drawn from a Gaussian distribution N(µ = 0, σ = 3.1623 dt 0.1). All network weights (B, W, Y) are randomly initialized, and networks are trained to minimize the summed costs across all conditions. For both tasks we trained 20 RNNs (different random initial connectivity) with gradient-based optimization to minimize the cost (Eq. 3; for details see section A.1.2). |