Few-shot transfer learning for individualized braking intent detection on neuromorphic hardware.

Area of Science:

• Neuromorphic computing
• Machine learning for assistive technologies
• Biomedical signal processing

Background:

• Traditional group-level models in electroencephalographic (EEG) data analysis lack individual specificity.
• Developing personalized models for real-time applications like advanced driver-assist systems (ADAS) is challenging.
• Neuromorphic systems offer potential for energy-efficient, on-device learning.

Purpose of the Study:

• To explore a few-shot transfer learning method for training convolutional spiking neural networks (CSNNs) on the BrainChip Akida AKD1000.
• To develop individual-level predictive models for braking intention using EEG data.
• To evaluate the efficacy and energy efficiency of the proposed method on a neuromorphic platform.

Main Methods:

• A group-level convolutional neural network (CNN) was trained on EEG data from participants performing a driving task.
• The CNN was converted for the Akida AKD1000 neuromorphic processor and quantized.
• Few-shot transfer learning was applied, training the final layer on individual data subsets using online Akida edge-learning for personalized CSNN models.

Main Results:

• Individual-specific braking intention models achieved over 90% accuracy, true positive rate, and true negative rate with as few as three training epochs.
• The Akida AKD1000 processor demonstrated over 97% power reduction compared to a traditional CPU, with a minimal 1.3x increase in latency.
• Ablation studies confirmed the robustness of the method with a reduced number of EEG channels.

Conclusions:

• The proposed few-shot transfer learning method enables rapid development of accurate, individual-specific predictive models on energy-efficient neuromorphic hardware.
• This approach is highly relevant for real-time applications requiring personalized, adaptive AI.
• The study highlights the potential of neuromorphic computing for customized, low-power edge AI solutions.