Generalization in data-driven models of primary visual cortex
Authors: Konstantin-Klemens Lurz, Mohammad Bashiri, Konstantin Willeke, Akshay Jagadish, Eric Wang, Edgar Y. Walker, Santiago A Cadena, Taliah Muhammad, Erick Cobos, Andreas S. Tolias, Alexander S Ecker, Fabian H. Sinz
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We train our network on neural responses from mouse primary visual cortex (V1) and obtain a gain in performance of 7% compared to the previous state-of-the-art network. We then investigate whether the convolutional core indeed captures general cortical features by using the core in transfer learning to a different animal. |
| Researcher Affiliation | Academia | 1 Institute for Bioinformatics and Medical Informatics, University of Tübingen, Germany 2 International Max Planck Research School for Intelligent Systems, Tübingen, Germany 3 Bernstein Center for Computational Neuroscience, University of Tübingen, Germany 4 Department for Neuroscience, Baylor College of Medicine, Houston, TX, USA 5 Center for Neuroscience and Artificial Intelligence, Baylor College of Medicine, Houston, TX, USA 6 Department of Computer Science / Campus Institute Data Science, University of Göttingen, Germany |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | To facilitate this, we share the weights of the trained representation together with its code online2 to allow others to predict neural responses with it. 2https://github.com/sinzlab/Lurz_2020_code |
| Open Datasets | Yes | The data used in our experiments consists of pairs of neural population responses and grayscale visual stimuli sampled and cropped from Image Net... Additionally, we also share the dataset that we evaluate our core on (Fig. 1, blue) so that other representations can be tested and compared with ours on the same data3. 3https://gin.g-node.org/cajal/Lurz2020 |
| Dataset Splits | Yes | We used the repeated images for testing, and split the rest into 4500 training and 500 validation images. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running experiments. |
| Software Dependencies | No | The paper mentions software components and algorithms (e.g., Adam optimizer, ELU nonlinearity) but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | The networks were trained to minimize Poisson loss 1m Pm i=1 ˆr(i) r(i) log ˆr(i) where m denotes the number of neurons, ˆr the predicted neuronal response and r the observed response. We used early stopping on the correlation between predicted and measured neuronal responses on the validation set (Prechelt, 1998): if the correlation failed to increase during any 5 consecutive passes through the entire training set (epochs), we stopped the training and restored the model to the best performing model over the course of training. We found that this combination of Poisson objective and early stopping on correlation yielded the best results. After the first stop, we decreased the learning rate from 5 10 3 twice by a decay factor of 0.3, and resumed training until it was stopped again. Network parameters were iteratively optimized via stochastic gradient descent using the Adam optimizer (Kingma & Ba, 2015) with a batch size of 64. Once training completed, the trained network was evaluated on the validation set to yield the score used for hyper-parameter selection. The hyper-parameters were then selected with a Bayesian search (Snoek et al., 2012) of 100 trials and subsequently kept fixed throughout all experiments. |