Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning
Authors: Wei Wu, Can Liao, Zizhen Deng, Zhengrui Guo, Jinzhuo Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate the efficacy of our method, we first conducted tests on simulated neuronal population dynamics data generated by the Izhikevich model. The results confirmed that our approach accurately captured the neuron types as defined by the preset hyperparameters. Subsequently, we applied our method to two real world neuron dynamics datasets... The representations learned from our model not only accurately predicted neuron types and locations but also demonstrated robustness when tested on out-of-domain data (data from unseen animals). Table 1: This table presents the performance metrics precision (Prec.), recall (Rec.), and F1 score across five methods (PCA, UMAP, Neur Print, LOLCAT, and Neur PIR) for different neuron types... |
| Researcher Affiliation | Academia | Wei Wu Peking University EMAIL, Can Liao University of Georgia EMAIL, Zizhen Deng Chinese Academy of Sciences Institute of Automation EMAIL, Zhengrui Guo The Hong Kong University of Science and Technology Beijing Institute of Collaborative Innovation EMAIL, Jinzhuo Wang Peking University EMAIL |
| Pseudocode | Yes | To enhance the reproducibility of this study, we provide an Appendix section comprising 4 subsections that offer detailed supplementary information. Appendix A.3 presents the pseudo-code of Synthetic Data. Appendix A.5 presents the pseudo-code of sownstream task. |
| Open Source Code | Yes | Code is available at https://github.com/ww20hust/Neur PIR. |
| Open Datasets | Yes | In recognition of the noise and complexity in real-world neuronal datasets, we turn to a publicly available dataset of mouse brain neuron populations... Real Data Bugeon: We utilized a rare, real-world multimodal dataset Bugeon et al. (2022)... Real Data Steinmetz: The Steinmetz dataset comprises 39 Neuropixels recordings... Steinmetz et al. (2019). |
| Dataset Splits | Yes | Evaluation on Simulated Data: ... (ii) Employ a 5-fold cross-validation approach to use these neuron representations as input to a classifier... Evaluation on Real Data Bugeon: (iii) Implement a 4-fold (with the folds based on the identity of the mice) cross-validation approach... Out-of-Domain Evaluation Steinmetz dataset: (iii) Implement a 10-fold (with the folds based on the identity of the mice) cross-validation approach... |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. It mentions technologies for data acquisition like 'two-photon imaging technology' and 'Neuropixels recordings' but not for the computational experiments. |
| Software Dependencies | No | The paper mentions software tools like VICReg and CEBRA (Schneider et al., 2023b) and uses the Izhikevich model (Izhikevich, 2003). It also includes Python code in the appendix for data handling. However, it does not provide specific version numbers for any of these key software components or libraries, which is required for reproducibility. |
| Experiment Setup | Yes | In the implementation, we adopt VICReg, which only uses positive pairs while indirectly separating dissimilar samples through regularization terms... First, we use CEBRA Schneider et al. (2023b) to integrate the single neuronal peripheral information... The dimensions of Xst are n + 1: The n dimensions are the activity of the surrounding neurons relative to the neuron you re considering, because there will be coordinates for each neuron in the dataset, so for each neuron, we ll take the 47 nearest neurons in the experiment. The last dimension is visual stimuli... The dimension of Xbe is 1: running speed each time point. The dimension of Xse is 1: session number. The dimensions of Xsi are 10: The total dimension of Xbe, Xse and Xst is 50, we use CEBBRA project them to the low dimension 10. We then denote the pair (Xsi, Xseg)for the same time as X in the following sections. The encoder G( ) maps the extracted input X into a latent representation H , which is then aggregated into time-invariant feature embeddings H using adaptive average pooling. These embeddings H are further mapped into a lower-dimensional space Z by a projection layer P( )... We applied to synthetic data where we can access the ground-truth intrinsic property. To make the synthetic data exhibit dynamics similar to that of real neurons, we simulated the data following the Izhikevich model... a , b , c and d can be regarded as intrinsic properties of each neuron. ... Employ a 5-fold cross-validation approach to use these neuron representations as input to a classifier... |