This repository contains the corresponding code, system- and circuit-level information, all simulated models, and detailed hardware evaluation results for the paper Toward A Formalized Approach for Spike Sorting Algorithms and Hardware Evaluation., which is openly accessible here.
Moreover, our synthetic dataset entitled The Synthetic Simulations Of Extracellular Recordings (SSOER) is made publicly available for download here.
Abstract: Spike sorting algorithms are used to separate extra-cellular recordings of neuronal populations into single-unit spike activities. The development of customized hardware implement- ing spike sorting algorithms is burgeoning. However, there is a lack of a systemic approach and a set of standardized evaluation criteria to facilitate direct comparison of both software and hardware implementations in the literature. In this paper, we formalize a set of standardized criteria and a publicly available synthetic dataset entitled Synthetic Simulations Of Extracellular Recordings (SSOER), which was constructed by aggregating existing synthetic datasets with varying Signal-To-Noise Ratios (SNRs). Furthermore, we present a benchmark for future comparison, and use our criteria to evaluate a simulated Resistive Random-Access Memory (RRAM) In-Memory Computing (IMC) system using the Discrete Wavelet Transform (DWT) for feature extraction. Our system consumes approximately (per channel) 10.72mW and occupies an area of 0.66mm2 in a 22nm FDSOI Complementary Metal–Oxide–Semiconductor (CMOS) process.
The SSOER dataset is made publicly available under a Creative Commons Attribution 4.0 International Licence. The dataset can be downloaded from https://zenodo.org/record/6214550.
This dataset is comprised of two files: data.npy and labels.csv.
- data.npy contains 14,400,000 sampled voltage values, from a single channel, taken at a sampling rate of 24 kHz.
- labels.csv contains the timestep, spike class, amplitude (SNR), and firing rate associated with each spiking event.
The Dataset folder includes the 5 sets of synthetic raw extracellular recordings from [1] in .mat data format, and a python script used to generate the SSOER dataset files.
The SpikeDetection folder includes AmpThrDet which performs amplitude threshold detection on raw signal. To compare the performance of various feature extraction algorithms, the folder also includes the DetThrGT which matches the detected spikes with the ground truth (GT) labels. Only detected spikes are only kept if it corresponds with a ground truth label are kept and false positives are removed. This is to ensure the spike sorting accuracy can be determined fairly using ground truth labels across various algorithms and confounding variables are minimized. ThrDetMaxAmpAlign.mlx and ThrDetMaxDeriAlign.mlx operate similarly to ThrDetGT, except they also perform spike alignment at the maximum amplitude and maximum derivative, respectively.
The FeatureExtraction folder includes the code of integer filter feature extraction and crossbar wavelet transform feature extraction, as well as crossbar accuracy, power, area and latency simulation associated with our proposed wavelet crossbar implementation. For simplicity, the extracted spikes and ground truths from the previous step are provided as matlab variables in the Extracted Spikes and Extracted Spikes Labels folders, and can be loaded directly into the python codes using the loadmat function. Note that the Extracted Spike folder contains unaligned (GTthrspike), aligned at maximum amplitude (GTthrAigned) and aligned at maximum derivative (GTthrAlignedDeri) waveforms to compare the alignment requirements for each algorithm.
Classification and accuracy testing folder includes the Classification_Accuracy.py file, which performs K-means clustering followed by accuracy evaluation which outputs the number of correctly sorted spikes vs. number of incorrectly sorted spikes. ICV.py calculates the ICV values for each cluster. The "group" variable in ICV.py refers to the output of K-means classifier in the form of a numpy array.
[1] J. Martinez, C. Pedreira, M. J. Ison, and R. Quian Quiroga, “Realistic simulation of extracellular recordings,” Journal of Neuroscience Methods, vol. 184, no. 2, pp. 285–293, Nov. 2009, doi: 10.1016/j.jneumeth.2009.08.017.
This study uses a semi-passive crossbar model proposed by Primeau et al. 2021 [28], which is based on existing semi-passive crossbar models [26], [29]. This model accounts for a variety of device non-idealities.
The parameters used for accuracy simulations are as follows:
- Wordline (WL) resistance: 20Ω
- Bitline (BL) resistance: 20Ω
- Size of crossbar: 64 by 64
- Size of tile: 8 by 8
- Voltage Supplied (Vdd): 0.2V
- DAC resolution: 4 bits
- ADC resolution: 14 bits
- CMOS line resistance: 600Ω
- CMOS transistor resistance: 20Ω
- R_on (High resistive state): 1e4
- R_off (Low resistive state): 1e5
- Bias Scheme: 1/3 bias scheme (unselected WLs biased to Vdd/3, unselected BLs biased to 2Vdd/3)
- Probability of memristor stuck on High resistive state: 0.01
- Probability of memristor stuck on Low resistive state: 0.01
The parameters used for power consumption simulations are as follows:
- RRAM cell area: 1e-14
- RRAM cell read voltage: 0.2V
- RRAM cell read latency: 6e-9s
- DAC area: 0.05e-6
- DAC power: 2e-4W
- DAC frequency: 100e6 Hz
- ADC area: 1.1e-3 by 0.6e-3
- ADC power: 10e-3W
- ADC frequency: 10e6 Hz
[2] C. Lammie, W. Xiang, B. Linares-Barranco, and M. R. Azghadi, “MemTorch: An Open-source Simulation Framework for Memristive Deep Learning Systems,” Neurocomputing, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0925231222002053.
[3] A. Amirsoleimani, “CODEX: Stochastic Encoding Method to Relax Resistive Crossbar Accelerator Design Requirements,” 2021.
[4] A. Chen, “A Comprehensive Crossbar Array Model With Solutions for Line Resistance and Nonlinear Device Characteristics,” IEEE Transactions on Electron Devices, vol. 60, no. 4, pp. 1318–1326, 2013.