Learning low-dimensional generalizable natural features from retina using a U-net
Authors: Siwei Wang, Benjamin Hoshal, Elizabeth de Laittre, Thierry Mora, Michael Berry, Stephanie Palmer
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This work focuses on using the retinal encoding of natural movies to determine the presumably behaviorallyrelevant features that the brain represents. In our end-to-end training, an encoder learns a compressed latent representation from a large population of salamander retinal ganglion cells responding to natural movies, while a decoder samples from this compressed latent space to generate the appropriate future movie frame. By comparing latent representations of retinal activity from three movies, we find that the retina has a generalizable encoding for time in the natural scene: the precise, low-dimensional representation of time learned from one movie can be used to represent time in a different movie, with up to 17 ms resolution. |
| Researcher Affiliation | Academia | Siwei Wang1, Benjamin Hoshal1, Elizabeth A de Laittre2, Olivier Marre3, Michael J Berry II4, and Stephanie E Palmer1 1Department of Organismal Biology and Anatomy, University of Chicago 2Committee on Computational Neuroscience, University of Chicago 3Sorbonne Université, INSERM, CNRS, Institut de la Vision 4Princeton Neuroscience Institute, Princeton University |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | You can find our Pytorch implementation at https://github.com/sepalmer/VU-net |
| Open Datasets | No | Our dataset contains retinal recordings of 93 cells responding to repeated, 20 second presentations of three natural movies at 60 frames per second. ... The data will be published in an experimental paper in the near term. we will release our code/data by then (or by the camera ready deadline if our submission is successful) |
| Dataset Splits | Yes | We train one predictive encoder-decoder for a specific movie using 40,000 training samples (see Supplementary Information for details) with a 90%/10% training/validation split. The reconstruction is from an additional held-out test set of 10,000 samples (100 frames, 100 PSTH patterns in each frame, see Supplementary Information). |
| Hardware Specification | No | The paper states in its reproducibility checklist (3.d): "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 Supplementary Information". Since the information is not in the main paper, it is considered not provided in the primary text. |
| Software Dependencies | No | The paper mentions using "Pytorch implementation", "Res Net18" and "VGG19 network" but does not specify version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We train one predictive encoder-decoder for a specific movie using 40,000 training samples (see Supplementary Information for details) with a 90%/10% training/validation split. ... Thus, for further analyses, we use dim Z = 10 unless otherwise stated. ... To be specific, we have σ2 > 0 for p(X|Z). This is also dubbed a committed rate for the encoder. In our case, we choose log σ2 = 1.0 to further ensure numerical stability of the mutual information estimator that we use [38]. |