Self-Supervised Graph Representation Learning for Neuronal Morphologies
Authors: Marissa A. Weis, Laura Pede, Timo Lüddecke, Alexander S Ecker
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show, in two different species and across multiple brain areas, that this method yields morphological cell type clusterings that are on par with manual feature-based classification by experts, but without using prior knowledge about the structural features of neurons. Moreover, it outperforms previous approaches on quantitative benchmarks predicting expert labels. |
| Researcher Affiliation | Academia | 1Institute of Computer Science and Campus Institute Data Science, University of Göttingen, Germany 2Institute for Theoretical Physics, University of Tübingen, Germany 3Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany *Correspondence: EMAIL |
| Pseudocode | No | The paper describes methods like Graph DINO and AC-Attention through descriptive text and mathematical equations (e.g., Equation 1) and diagrams (e.g., Figure 1), but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://eckerlab.org/code/weis2023/. |
| Open Datasets | Yes | Blue Brain Project (BBP): Rat somatosensory cortex. Available from the Neocortical Microcircuit Collaboration Portal of the Blue Brain Project1 (Ramaswamy et al., 2015)... M1 Patch Seq: Mouse motor cortex... (Scala et al., 2021).2... Allen Brain Atlas (ACT): Mouse visual cortex... (Allen Institute, 2016).3... Brain Image Library (BIL): Whole mouse brain... (Peng et al., 2021).4... Janelia Mouse Light (JML): Whole mouse brain... (Winnubst et al., 2019).5... The Trees dataset (Seidel et al., 2021)... |
| Dataset Splits | Yes | For evaluation, we follow the evaluation protocol and use the same dataset split as Laturnus & Berens (2021). We additionally report the 5-nearest neighbour accuracy of three additional dataset splits... Chen et al. (2022b) remove 955 neurons from the dataset due to reconstruction errors and evaluate on a 80-20% training-test split... Since we did not have access to the exact split, we performed five cross-validations instead and report the average accuracy over folds. |
| Hardware Specification | Yes | All trainings were performed on a NVIDIA Quadro RTX 5000 single GPU. |
| Software Dependencies | No | Graph DINO is implemented in Py Torch (Paszke et al., 2019) and trained with the Adam optimizer (Kingma & Ba, 2015). ... We use scipy for fitting Gaussian Mixture models (GMM) and k-nearest neighbor classifiers (k NN) (Pedregosa et al., 2011). No specific version numbers are provided for these software components, making replication difficult without trial and error. |
| Experiment Setup | Yes | To select hyperparameters we run three grid searches and pick the best hyperparameters according to the lowest average loss over the BBP and M1 Patch Seq dataset. For the optimization, we run a hyperparameter search over batch size {32, 64, 128}, learning rate {10-3, 10-4, 10-5}, and number of training iterations {20,000, 50,000, 100,000}. For the augmentation strength, we run a hyperparameter search over jitter variance σ1 {1.0, 0.1, 0.001}, number of deleted branches n {1, 5, 10}, and graph position variance σ2 {0.1, 1.0, 10.0}. For the architecture, we run a hyperparameter search over latent dimension {16, 32, 64}, number of Graph Attention blocks (depth) {5, 7, 10}, and number of attention heads per block {2, 4, 8}. Tables C.1, C.2, and C.3 provide the specific hyperparameters used for each dataset. |