Abstract
This paper presents a spike sorting processor based on an accurate spike clustering algorithm. The proposed spike sorting algorithm employs an L2-normalized convolutional autoencoder to extract features from the input, where the autoencoder is trained using the proposed spike sorting-aware loss. In addition, we propose a similarity-based K-means clustering algorithm that conditionally updates the means by observing the cosine similarity. The modified K-means algorithm exhibits better convergence and enables online clustering with higher classification accuracy. We implement a spike sorting processor based on the proposed algorithm using an efficient time-multiplexed hardware architecture in a 40-nm CMOS process. Experimental results show that the processor consumes 224.75μW/mm² when processing 16 input channels at 7.68 MHz and 0.55 V. Our design achieves 95.54% clustering accuracy, outperforming prior spike sorting processor designs.
Introduction
In neural prosthetics and brain-machine interfaces (BMIs), researchers must record activity from hundreds or thousands of individual neurons. However, a single electrode often picks up signals from multiple nearby neurons. "Spike sorting" is the computational challenge of separating these overlapping signals into distinct "units."
Implementing this process on-chip (inside an implant) is difficult due to tiny power budgets and limited bandwidth. Sending raw neural data to an external computer is too energy-intensive, while existing on-chip solutions often sacrifice accuracy for simplicity. This paper addresses this gap by implementing a sophisticated deep-learning-based feature extractor and an unsupervised clustering algorithm directly into a 40-nm CMOS processor.
Methods
The authors propose a hardware-software co-design approach consisting of five stages: detection, alignment, initial template generation, feature extraction, and clustering.
Convolutional Autoencoder (CAE) and Weighted Loss
Instead of traditional methods like Principal Component Analysis (PCA), the system uses a 1D-CAE to extract features. The encoder compresses a 50-sample spike into a 4-dimensional hidden feature vector. A key innovation is the "spike-sorting-aware" loss function. While standard training uses Mean Squared Error (MSE), the authors found that the middle of a spike waveform (samples 6–25) contains the most discriminative information. By prioritizing these samples during training, the autoencoder learns more robust features.
Similarity-Based K-Means and Hardware Design
Traditional K-means can be sensitive to noise, causing cluster means to drift. The authors implemented a modified version using cosine similarity to determine if an incoming signal is a valid spike before updating the cluster mean. To stay under the 277 μW/mm² thermal limit for brain tissue, the processor uses an SRAM-less architecture, allowing stable operation at a mere 0.55 V.
Results
The design was evaluated using synthetic Quiroga datasets and real-world human neural data recorded at UCLA.
- Accuracy: The processor achieved 95.54% clustering accuracy on synthetic data, outperforming prior state-of-the-art hardware. The proposed weighted loss function improved accuracy by 3.4% over standard MSE.
- Generalization: On real human neural data, the algorithm achieved 87.03% accuracy, significantly outperforming the popular "Wave-Clus" software (70.87%).
- Efficiency: The processor consumes only 224.75 μW/mm², which is safely below the biological safety limit for heat dissipation in the brain.
Conclusion
This work demonstrates that deep-learning-based feature extraction can be efficiently implemented on-chip for high-accuracy spike sorting. While the autoencoder weights are currently pre-trained offline, the unsupervised clustering enables robust real-time operation. Future work will explore scaling the time-multiplexed architecture to support higher channel counts required for high-density neural probes.
Summary updated