Self-Supervised Learning of Representations for Space Generates Multi-Modular Grid Cells
Authors: Rylan Schaeffer, Mikail Khona, Tzuhsuan Ma, Cristobal Eyzaguirre, Sanmi Koyejo, Ila Fiete
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Experimental Results |
| Researcher Affiliation | Academia | Rylan Schaeffer Computer Science Stanford University rschaef@cs.stanford.edu Mikail Khona Physics MIT mikail@mit.edu Tzuhsuan Ma Janelia Research Campus Howard Hughes Medical Institute mat@janelia.hhmi.org Cristóbal Eyzaguirre Computer Science Stanford University ceyzaguirre@stanford.edu Sanmi Koyejo Computer Science Stanford University sanmi@cs.stanford.edu Ila Rani Fiete Brain and Cognitive Sciences MIT fiete@mit.edu |
| Pseudocode | No | The paper describes the proposed framework and loss functions using text and mathematical equations, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | Our code will be made publicly available upon publication. |
| Open Datasets | No | For each gradient step, we sample a sequence of T velocities (v1, v2, ..., v T ), with vt i.i.d. p(v), then construct a batch by applying B randomly sampled permutations {πb}B b=1, πb : [T] [T] to the sequence of velocities to obtain B permuted velocity trajectories; doing so ensures many intersections between the trajectories exist in each batch (SI Fig. 8b). (No mention of a publicly available dataset to download). |
| Dataset Splits | No | The paper describes how training data batches are constructed ('For each gradient step, we sample a sequence of T velocities...'), and mentions 'training distribution' and 'evaluation trajectories', but does not provide explicit train/validation/test dataset splits with percentages or sample counts. |
| Hardware Specification | No | The paper does not specify the hardware used for running experiments (e.g., GPU models, CPU types, or cloud computing instances). |
| Software Dependencies | No | Our code was implemented in Py Torch [46] and Py Torch Lightning [20]. (No version numbers provided for these software components). |
| Experiment Setup | Yes | Appendix A: Experimental Details and Table 1: Hyperparameters used for training the networks. (This table lists specific values for Batch size, Trajectory length, Velocity sampling distribution, RNN nonlinearity, Number of RNN units, Number of MLP layers, Spatial length scale σx, Neural length scale σg, Separation loss coefficient λSep, Invariance loss coefficient λInv, Capacity loss coefficient λCap, Optimizer, Optimizer scheduler, Learning rate, Gradient clip value, Weight decay, Accumulate gradient batches, Number of gradient descent steps). |