Brain–Spine Interfaces (BSIs) have recently demonstrated the ability to restore walking after spinal cord injury by decoding motor intentions from electrocorticography (ECoG) recordings and translating them into spinal cord stimulation patterns. However, current decoding pipelines rely on conventional digital processors, limiting power efficiency and preventing the development of fully implantable systems. Neuromorphic hardware offers an attractive alternative through event-driven computation and ultra-low-power processing, but learning compact and informative neural representations directly on such hardware remains an open challenge.

This project aims to develop and evaluate neuromorphic representation learning algorithms capable of extracting low-dimensional latent representations of motor cortical activity recorded from participants implanted with WIMAGINE ECoG devices during the STIMO Brain–Spine Interface clinical trial. The long-term objective is to create compact latent representations that preserve movement-related information while remaining compatible with future fully implantable brain–spine interfaces.

 

Methodology

1. Implement and evaluate alternative spike encoding strategies for transforming continuous ECoG recordings into event-based representations while preserving motor-related information.

2. Develop neuromorphic autoencoder architectures incorporating biologically inspired

3. Compare different reconstruction objectives (e.g., spike-timing-aware losses, contrastive objectives, or task-aware latent constraints) to determine which best preserve movement-related information.

4. Evaluate the learned latent representations through visualization, dimensionality analysis, and downstream movement decoding, comparing performance against conventional dimensionality reduction techniques and previous neuromorphic approaches.

 

Available Material

• Human ECoG recordings from the EPFL Brain–Spine Interface clinical study.

• Neuromorphic simulation framework compatible with mixed-signal neuromorphic hardware.

• Existing spiking autoencoder implementation and baseline models.

• Access to neuromorphic hardware available at the Institute of Neuroinformatics.

 

Requirements

• Python programming.

• Interest in machine learning, neuromorphic engineering, and brain–computer interfaces.

• Basic knowledge of neural networks is beneficial.

 

Contacts

Elisa Donati (elisa@ini.uzh.ch)

 

References

[1] Lorach, H., et al. Nature, 618, 126–133 (2023).

[2] Lee, D.-H., Zhang, S., Fischer, A., & Bengio, Y. Difference Target Propagation. Machine Learning and Knowledge Discovery in Databases (ECML PKDD), 498–515 (2015).