PhD Scientific Days 2022

Theoretical Medicine (Poster discussion will take place on the terrace of the room during the Coffee Break)

Deep learning-based spike sorting on edge devices

Text of the abstract

Introduction
Today’s multi-channel silicon-based probes can generate large amounts of data, providing finer resolution and enabling better separation of the spike clusters. To cope with the increased complexity, several algorithms (so-called spike sorting algorithms) were developed to provide a faster and more accurate data analysis tool.

Aims
The aim of our project was to assess the performance of deep learning methods and their efficiency in the evaluation of high complexity extracellular neural data on edge devices like a tensor processing unit (TPU).

Method
A supervised deep learning model was trained to identify and classify extracellular signals. The architecture of the model was selected from a list of well-known, optimized architectures used in the field of deep learning computer vision. The training data consisted of samples of 128x128 dimension semi-synthetic data in order to provide the most accurate ground truth to the model during training. After quantization, the model was transferred and tested on a Coral Development Board Mini, containing a Coral Edge TPU.
Results
During the initial supervised training of the model, a mean average precision of 63.04 was achieved on the test dataset. The inference speed of the initial model, with single-precision floating-point weights, could reach 3,17 msec/sample on a high-performance PC, while the quantized model on the development board reached an average speed of 90 msec/sample.

Conclusion
Considering the promising results, a conclusion can be made, that by further optimizing the model in the future, a highly efficient, real-time method can be developed for spike sorting, enabling an efficient and near-state-of-art performance on edge devices.

Funding
Project no. ÚNKP-21-3-II-SE-1 has been implemented with the support provided by the Ministry of Innovation and Technology of Hungary from the National Research, Development and Innovation Fund, financed under the ÚNKP-21-3 funding scheme. This research was also partially funded by Semmelweis 250+ grant (EFOP-3.6.3-VEKOP-16-2017-00009), by the Hungarian Brain Research Program (2017_1.2.1-NKP-2017-00002) and by the Hungarian National Research, Development and Innovation Office (TUDFO/51757-1/2019-ITM). G M is thankful to the Hungarian National Research, Development and Innovation Office (FK132823) grant.