Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Decoding EEG With Spiking Neural Networks on Neuromorphic Hardware
Authors: Neelesh Kumar, Guangzhi Tang, Raymond Yoo, Konstantinos P. Michmizos
TMLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We deployed our SNN on Intel s Loihi neuromorphic processor, and show that our method consumed 95% less energy per inference than the state-of-the-art DNN methods on NVIDIA Jeston TX2, while achieving similar levels of classification performance. Finally, we interpreted the SNN using a network perturbation study to identify the spectral bands and brain activity that correlated with the SNN outputs. The results were in agreement with the current neurophysiological knowledge implicating the activation patterns in the low-frequency oscillations over the motor cortex for hand movement and imagery tasks. |
| Researcher Affiliation | Academia | Neelesh Kumar EMAIL Department of Computer Science Rutgers University Guangzhi Tang EMAIL Department of Computer Science Rutgers University Raymond Yoo EMAIL Department of Computer Science Rutgers University Konstantinos P. Michmizos EMAIL Department of Computer Science Rutgers University |
| Pseudocode | No | The paper describes the model architecture and mathematical equations for its components (e.g., equations 1-16) but does not present a dedicated pseudocode or algorithm block. |
| Open Source Code | Yes | 1Code available at https://github.com/combra-lab/snn-eeg |
| Open Datasets | Yes | We validated our method on an in-house, IRB-approved, dataset for classifying complex movement compo-Published in Transactions on Machine Learning Research (06/2022)nents, namely reaction time (RT) and directions, and the publicly available eegmmidb dataset for classifying motor imagery and execution tasks. ... We evaluated the SNN on the publicly available eegmmidb dataset which comprises of EEG recordings from 109 subjects acquired using a 64-channel system during motor movement and imagery tasks (Schalk et al., 2004). |
| Dataset Splits | Yes | We used the leave-k-out technique for evaluation, where data from all but k subjects were used for training. The evaluation was done on the data from the k left-out subjects. ... For the in-house dataset, k = 1, i.e. there were 11 training subjects and 1 test subject. For the eegmmidb dataset, k = 10, which meant that there were 90 training subjects and 10 test subjects. The list of hyperparameters for training is provided in Table 1. |
| Hardware Specification | Yes | We deployed our SNN on Intel s Loihi neuromorphic processor, and show that our method consumed 95% less energy per inference than the state-of-the-art DNN methods on NVIDIA Jeston TX2, while achieving similar levels of classification performance. |
| Software Dependencies | No | The paper describes the methodologies and models but does not provide specific version numbers for software dependencies like programming languages, libraries, or frameworks used for implementation. |
| Experiment Setup | Yes | Table 1: List of hyperparameters Hyperparameters Reaction Time Directions EEGMMIDB EEGMMIDB w/Loihi Segmentation intervals -0.5-0s -0.5-1.5s 0-1s 0-1s Initial voltage decay (Dv) 0.1 0.1 0.1 0.1 (Fixed) Initial current decay (Dc) 0.1 0.1 0.1 0.1 (Fixed) Initial voltage threshold (Vth) 0.1 0.1 0.1 0.2 (Fixed) Feedforward timesteps 125 500 160 160 Convolutional layer architecture 64C1-128C2-256C3 64C1-128C2-256C3 64C1-128C2-256C3 4C1-8C2-128C3 Timesteps for temporal convolutional(win) 3 3 3 3 Pseudograd amplifier (a1) 0.3 0.3 0.3 0.3 Pseudograd window (a2) 0.3 0.3 0.3 0.3 Weight learning rate 1e-4 1e-4 1e-4 1e-4 Decays learning rate 1e-3 1e-3 1e-4 1e-4 Batch size 64 64 64 64 |