Learnable Parametric Bandpass Filters for EEG Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning approaches for classifying brain activity signals, such as EEG, require extensive manual feature engineering and deep learning methods need large datasets, while providing limited interpretability.
Innovation Solution
An end-to-end trainable model using parametrised bandpass filters and a differentiable feature module to generate interpretable feature maps, enabling classification with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning approaches are used for classifying brain activity signals, then manual feature engineering can be performed, but important unidentified features may be missed and extensive manual effort is required
Solution Approach 1:
The system automatically learns relevant features from raw EEG signals through the differentiable feature module and trained filters, eliminating the need for manual feature engineering. The model self-identifies important features by optimizing them during training to maximize classification performance.
Solution Approach 2:
Manual feature engineering processes are replaced with an automated differentiable feature module that computes features through mathematical operations. This substitution enables end-to-end training and automatic feature optimization without human intervention.
2Reliability
If deep learning approaches are used for EEG decoding, then classification performance can be improved, but large amounts of data are required and interpretability is limited
Solution Approach 1:
The patent changes the parameters of the filters (center frequencies, bandwidths) to be learnable quantities that adapt during training. This allows the model to automatically discover optimal frequency bands for different classification tasks while maintaining interpretability through the parametric filter structure.
Solution Approach 2:
The differentiable feature module acts as an intermediary between raw EEG signals and the classification model, providing a bridge that maintains interpretability. By using differentiable operations instead of black-box deep learning layers, the system preserves the ability to understand and interpret intermediate features.
3Reliability
If deep learning models are trained on limited EEG data, then training may fail or performance degrades, but collecting large datasets increases time and resource requirements
Solution Approach 1:
The patent incorporates prior neuroscientific knowledge into the model architecture through parametric filters with physically meaningful parameters. This preliminary structuring of the model based on domain knowledge allows effective training with limited data by reducing the search space and providing inductive biases.
Solution Approach 2:
The filter parameters (center frequencies, bandwidths) are made dynamic and adaptable during training rather than fixed. This allows the model to automatically adjust to the specific characteristics of the available data, improving robustness when training data is limited.
Data Source
Figure 1
Figure 2a~2d
Figure 3
AI summary
This specification relates to the classification and/or decoding of brain activity signals, such as electroencephalogram (EEG) signals, using machine-learning techniques. According to one aspect of this specification, there is described a computer implemented method of classifying brain activity signals. The method comprises: receiving a plurality of channels of brain activity signals; generating a plurality of channels of filtered brain activity signals by applying a plurality of filters to the received channels of brain activity signals, wherein the plurality of filters comprises a plurality of learned parameterised bandpass filters; determining, using a differentiable feature module, a plurality of feature maps from the plurality of channels of filtered brain activity signals; and determining, using a classification model, one or more classifications for the received plurality of channels of brain activity signals based on the determined feature maps.