Back to the Continuous Attractor

Authors: Ábel Ságodi, Guillermo Martín-Sánchez, Piotr Sokol, Memming Park

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We train recurrent neural networks on analog memory tasks to support the appearance of these systems as solutions and their generalization capabilities. (Abstract), 4 Numerical Experiments on Task-optimized Recurrent Networks (Section 4 title)
Researcher Affiliation Academia Champalimaud Centre for the Unknown Champalimaud Foundation, Lisbon, Portugal
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code for this project is publicly available at https://github.com/catniplab/back_to_the_continuous_attractor.
Open Datasets No Building upon prior work, which has shown their capabilities on such tasks, we trained RNNs to either (1) estimate head direction through integration of angular velocity92,93 (Fig. 4A1) or (2) perform a memory-guided saccade task for targets on a circle100,101 (Fig. 4A2, details in Sec. S7.1 and see Sec. S7.3 for how RNNs relate to Eq. 2). The paper describes how data for these tasks are sampled but does not provide a link or specific citation to a publicly available dataset.
Dataset Splits No The paper mentions using 'a validation set' and specifies training parameters like 'batch size of 64' and 'training was run for 5000 gradient updates', but it does not provide specific dataset split percentages or sample counts for training, validation, and test sets.
Hardware Specification No Training a single network took around 10 minutes on a CPU and occupied 10 percent of an 8GB RAM. This description does not provide specific CPU models or detailed hardware specifications.
Software Dependencies Yes We trained RNNs with Py Torch 146 (Section S7.5)
Experiment Setup Yes Adam optimization with β1 = 0.9 and β2 = 0.999 was employed with a batch size of 64 and training was run for 5000 gradient updates. The best learning rate 10 2 was chosen from a set of values {10 2, 10 3, 10 4, 10 5} by 5 initial runs for all nonlinearity and size pairings with the lowest average loss after 100 gradient steps. (Section S7.5)