Learning Place Cell Representations and Context-Dependent Remapping
Authors: Markus Pettersen, Frederik Rogge, Mikkel Lepperød
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we propose a similarity-based objective function that translates proximity in space, to proximity in representation. We show that a neural network trained to minimize the proposed objective learns place-like representations. We also show that the proposed objective is easily extended to include other sources of information, such as context information, in the same way. When trained to encode multiple contexts, networks learn distinct representations, exhibiting remapping behaviors between contexts. |
| Researcher Affiliation | Academia | Markus Pettersen Department of Numerical Analysis and Scientific Computing Simula Research Laboratory Oslo, Kristian Augusts Gate 23 markusb@simula.no Frederik Rogge Department of Biosciences University of Oslo Oslo, Blinderveien 31 frederik.rogge@ibv.uio.no Mikkel Elle Lepperød Department of Numerical Analysis and Scientific Computing Simula Research Laboratory Oslo, Kristian Augusts Gate 23 mikkel@simula.no |
| Pseudocode | No | The paper describes the models and training procedures in detail within the text and equations, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Figures were created using Bio Render.com, and code to reproduce all findings and figures is available at https://github.com/bio AI-Oslo/Conjunctive Representations. |
| Open Datasets | No | At training time, data was created on the fly due to the low computational cost. The input to the feedforward network consisted of minibatches of continuous Cartesian coordinates, sampled randomly and uniformly within a 2 × 2 square enclosure. For the conjunctive objective (2), the input to the network was a concatenation of randomly sampled Cartesian coordinates x, and uniformly sampled scalar context signals c, i.e. input = cat(x, c). Context signals were sampled uniformly in the interval c ∈ [−2, 2]. Trajectories were generated by creating boundary-avoiding steps successively. |
| Dataset Splits | No | The data used for training and evaluation is generated on the fly rather than being a pre-defined dataset with explicit training, validation, and test splits. While there's a distinction between training and evaluation data, no separate validation split is explicitly mentioned for hyperparameter tuning. |
| Hardware Specification | No | The paper states that the 'model is minimal, and runs on most modern laptops' in the NeurIPS checklist, but it does not provide specific hardware details such as GPU or CPU models, memory, or cloud computing specifications used for the experiments. |
| Software Dependencies | No | Models were implemented and trained using the Pytorch library [Paszke et al., 2019]. We used the special_ortho_group from the Sci Py library [Virtanen et al., 2020]. We used the Adam optimizer [Kingma and Ba, 2017]. UMAP [Mc Innes et al., 2018], with default parameters. While several software libraries are mentioned, no specific version numbers are provided for PyTorch, SciPy, or UMAP. |
| Experiment Setup | Yes | The feedforward network featured two densely connected layers with 64 and 128 hidden units, followed by an output layer containing np = 256 units. Every layer was equipped with the Re LU activation function. The recurrent network consisted of a single vanilla recurrent layer equipped with the Re LU activation function, without added bias. Like the feedforward network, this network featured np = 256 recurrent units. The trajectory length was taken to be T = 10 timesteps. All networks were trained for a total of 60000 training steps, with a batch size of 64. For each model, we used the Adam optimizer [Kingma and Ba, 2017] with a learning rate of 10−4. Unless otherwise specified, all models were trained with λ = 0.1, β = 0.5 and σ = 0.25. |