Source-linked AI summary
Interpretable Deep Neural Networks for Single-Trial EEG Classification
Irene Sturm, Sebastian Bach, Wojciech Samek, Klaus-Robert Müller
TL;DR
DNNs remain difficult to interpret in cognitive neuroscience, where neurophysiological insight is important alongside classification. This paper applies DNNs with LRP to EEG motor-imagery BCI, finding detailed, plausible single-trial explanations and potential benefits from subject-to-subject transfer despite no overall performance advantage over CSP-LDA.
Problem
DNNs’ black-box character limits insight into neurophysiological phenomena underlying decisions in cognitive neuroscience.
Method
The study applies DNNs with LRP to EEG motor-imagery BCI and compares classification with CSP-LDA across two datasets.
Results
LRP produces neurophysiologically plausible relevance heatmaps for single time points in single trials, while DNN performance does not outperform CSP-LDA overall.
Takeaways & Limitations
LRP adds high-resolution explanatory information to DNN-based EEG analysis, including accounts of relevant information when individual trials must be examined.
Abstract
from arXiv · showhide
Background: In cognitive neuroscience the potential of Deep Neural Networks (DNNs) for solving complex classification tasks is yet to be fully exploited. The most limiting factor is that DNNs as notorious 'black boxes' do not provide insight into neurophysiological phenomena underlying a decision. Layer-wise Relevance Propagation (LRP) has been introduced as a novel method to explain individual network decisions. New Method: We propose the application of DNNs with LRP for the first time for EEG data analysis. Through LRP the single-trial DNN decisions are transformed into heatmaps indicating each data point's relevance for the outcome of the decision. Results: DNN achieves classification accuracies comparable to those of CSP-LDA. In subjects with low performance subject-to-subject transfer of trained DNNs can improve the results. The single-trial LRP heatmaps reveal neurophysiologically plausible patterns, resembling CSP-derived scalp maps. Critically, while CSP patterns represent class-wise aggregated information, LRP heatmaps pinpoint neural patterns to single time points in single trials. Comparison with Existing Method(s): We compare the classification performance of DNNs to that of linear CSP-LDA on two data sets related to motor-imaginery BCI. Conclusion: We have demonstrated that DNN is a powerful non-linear tool for EEG analysis. With LRP a new quality of high-resolution assessment of neural activity can be reached. LRP is a potential remedy for the lack of interpretability of DNNs that has limited their utility in neuroscientific applications. The extreme specificity of the LRP-derived heatmaps opens up new avenues for investigating neural activity underlying complex perception or decision-related processes.
I. INTRODUCTION
DNNs offer powerful nonlinear classification, but their black-box character limits neurophysiological insight in cognitive neuroscience. LRP addresses this gap by making individual DNN decisions interpretable through input-level relevance, enabling detailed single-trial EEG explanations.
- DNNs are powerful for complex classification, yet cognitive neuroscience and BCI still rely largely on linear methods.
- Black-box DNN decisions are limiting in cognitive neuroscience because neurophysiological insight matters beyond classification performance.
- LRP explains individual DNN decisions by decomposing the output in terms of input variables.
- DNN with LRP can combine nonlinear classification for high-dimensional EEG with detailed single-trial accounts of decision-relevant information.
- The study applies DNN with LRP to EEG motor-imagery BCI, compares performance with CSP-LDA, and produces spatio-temporal relevance heatmaps for individual trials.
A. Model Details
The model uses two linear sum-pooling layers with nonlinear activation or normalization, mapping vectorized EEG features to two class probabilities. Training uses minibatch error backpropagation and stops after 3000 iterations.
- The first layer maps 33,518-dimensional EEG inputs to a 500-dimensional tanh-activated representation.For subjects od-obx, the input dimensionality is 17,458.
- The second layer reduces the 500-dimensional representation to two outputs, followed by softmax to produce class probabilities.
- Training uses standard error backpropagation with batches of five randomly drawn samples.
- Training terminates after 3000 iterations, yielding the reported prediction accuracy.
B. Interpretability
LRP decomposes a DNN’s classification score into input-level relevance values and propagates them backward through local redistribution rules. The resulting relevances identify each input component’s contribution to a decision while satisfying layer-wise conservation.
- For each input sample, the DNN produces a classification score that LRP decomposes into input-component relevances.Each relevance value represents an input component’s share in the classification decision.
- LRP backpropagates relevance values from the network output to the input layer using a local redistribution rule.
- The redistribution rule assigns relevance across connected neurons using learned pair-specific parameters.
- The redistribution rule fulfills layer-wise conservation and is closely related to a deep variant of Taylor decomposition.
A. Experimental Setup and Preprocessing
The study evaluates DNN with LRP on two motor-imagery EEG datasets and compares it with CSP-LDA. Data are filtered, epoched, baseline-corrected, vectorized, and then used to compute trial-level relevance maps.
- Experiments use cued right-hand-versus-foot motor imagery from five subjects and no-distraction left-versus-right-hand motor imagery from another five-subject subset.
- All datasets are downsampled to 100 Hz and bandpass filtered from 9–13 Hz.
- CSP uses [1000 4000] ms post-cue epochs with three spatial-filter pairs, followed by regularized LDA with analytical shrinkage.
- DNN processing uses post-cue epoch envelopes, subtracts a [0 300] ms pre-cue baseline, and vectorizes each epoch’s spatio-temporal features.
- Trial-level relevance maps are calculated from the two-valued DNN output according to the interpretability procedure.
B. Results
DNN classification generally underperformed CSP-LDA, but improved accuracy for the lowest-performing subjects. LRP heatmaps exposed plausible, time-resolved motor patterns in individual trials and resembled class-level CSP patterns when averaged.
- DNN performance was generally lower than CSP-LDA, except for subjects ay and njy, where it increased classification accuracy.
- Inter-subject DNN performed worse than single-subject DNN in 6/10 subjects but substantially increased accuracy in the remaining four.
- LRP heatmaps mapped relevance across each EEG channel and time point, with scalp topographies available for individual time points.
- Single-trial LRP maps revealed typical lateralized motor activation patterns linked to individual time points and resembled CSP patterns after class-wise averaging.
IV. DISCUSSION
The discussion emphasizes LRP as the paper’s main contribution: it makes DNN decisions neurophysiologically inspectable at single instants of single trials. The study also reports potential benefits from inter-subject training, especially for low-performing subjects, while identifying broader applications for detailed trial-level EEG interpretation.
- The relatively simple DNN did not outperform CSP-LDA, but successive training across subjects improved classification accuracy in some cases.
- Inter-subject training substantially increased accuracy for a subject with particularly low accuracy, suggesting potential for transferring learned neural representations.
- LRP produced neurophysiologically plausible, textbook-like motor-imagery explanations at single instants of single trials.
- Unlike CSP-LDA’s class-ensemble summaries, LRP can diagnose influences associated with low-confidence or erroneous individual network decisions.
- Outside BCI, the authors suggest LRP may support detailed single-trial neurophysiological interpretation in clinical and other EEG settings.
V. CONCLUSION
The paper concludes that LRP adds an explanatory layer to DNN-based EEG/BCI analysis by detailing relevant information in high-dimensional data at the level of individual trials.
- LRP provides highly detailed accounts of relevant information in high-dimensional EEG data when individual trials must be analyzed separately.