This paper presents a 32-channel analog filterbank for front-end signal processing in sound classification systems. It employs a passive N-path switched capacitor topology to achieve high power efficiency and reconfigurability. The circuit's unwanted harmonic mixing products are absorbed by the machine learning model during training. To enable a systematic pre-silicon study of this effect, we develop a computationally efficient circuit model that can process large machine learning datasets on practical time scales. Measured results using a 130 nm CMOS prototype IC indicate competitive classification accuracy on datasets for baby cry detection (93.7% AUC) and voice commands (92.4% average precision), while lowering the feature extraction energy compared to digital realizations by approximately 2× and 10×, respectively. The 1.44 mm 2 chip consumes 800 nW, which corresponds to the lowest normalized power per simultaneously sampled channel in recent literature.
15 August 2019 in IEEE Transactions on Biomedical Circuits and Systems
Neural interfaces of the future will be used to help restore lost sensory, motor, and other capabilities. However, realizing this futuristic promise requires a major leap forward in how electronic devices interface with the nervous system. Next generation neural interfaces must support parallel recording from tens of thousands of electrodes within the form factor and power budget of a fully implanted device, posing a number of significant engineering challenges. In this paper, we exploit sparsity and diversity of neural signals to achieve simultaneous data compression and channel multiplexing for neural recordings. The proposed architecture uses wired-OR interactions within an array of single-slope A/D converters to obtain massively parallel digitization of neural action potentials. The achieved compression is lossy but effective at retaining the critical samples belonging to action potentials, enabling efficient spike sorting and cell type identification. Simulation results of the proposed architecture using data obtained from primate retina ex-vivo with a 512-channel electrode array show average compression rates up to ~40x while missing less than 5% of cells. In principle, the techniques presented here could be used to design interfaces to other parts of the nervous system.
29 December 2023 in IEEE Journal of Solid-State Circuits
This article presents a data-compressive neural recording IC for single-cell resolution high-bandwidth brain–computer interfaces (BCIs). The IC features wired-OR lossy compression during digitization, thus preventing data deluge and massive data movement. By discarding unwanted baseline samples of the neural signals, the output data rate is reduced by 146$\times $ on average while allowing the reconstruction of spike samples. The recording array consists of pulse-position modulation (PPM)-based active digital pixels (ADPs) with a global single-slope (SS) analog-to-digital conversion scheme, which enables a low-power and compact pixel design with significantly simple routing and low array readout energy. Fabricated in a 28-nm CMOS process, the neural recording IC features 1024 channels (i.e., 32 $\times $ 32 array) with a pixel pitch of 36 $\mu \text{m}$ that can be directly matched to a high-density micro-electrode array (MEA). The pixel achieves 7.4-$\mu \text{V}_{\text {rms}}$ input-referred noise with a −3-dB bandwidth of 300 Hz–5 kHz while consuming only 268 nW from a single 1-V supply. The IC achieves the smallest area per channel (36 $\times $ 36 $\mu \text {m}^{{2}}$ ) and the highest energy efficiency among the state-of-the-art neural recording ICs published to date.
4 July 2023 in IEEE Transactions on Biomedical Circuits and Systems
Future high-density and high channel count neural interfaces that enable simultaneous recording of tens of thousands of neurons will provide a gateway to study, restore and augment neural functions. However, building such technology within the bit-rate limit and power budget of a fully implantable device is challenging. The wired-OR compressive readout architecture addresses the data deluge challenge of a high channel count neural interface using lossy compression at the analog-to-digital interface. In this paper, we assess the suitability of wired-OR for several steps that are important for neuroengineering, including spike detection, spike assignment and waveform estimation. For various wiring configurations of wired-OR and assumptions about the quality of the underlying signal, we characterize the trade-off between compression ratio and task-specific signal fidelity metrics. Using data from 18 large-scale microelectrode array recordings in macaque retina ex vivo , we find that for an event SNR of 7-10, wired-OR correctly detects and assigns at least 80% of the spikes with at least 50× compression. The wired-OR approach also robustly encodes action potential waveform information, enabling downstream processing such as cell -type classification. Finally, we show that by applying an LZ77-based lossless compressor (gzip) to the output of the wired-OR architecture, 1000× compression can be achieved over the baseline recordings.
1 October 2025 in IEEE Journal of Solid-State Circuits
This article presents a 1024-channel ultra-low-power spike sorting chip featuring event-driven spike detection and spatial clustering for large-scale neural recording. To address power and scalability constraints in brain–computer interfaces (BCIs), the design integrates a compressive analog-to-digital converter (ADC) with a two-stage spike detector that significantly reduces memory and processing activity. Spatial features derived from high-density micro-electrode array (MEA) enhance cluster separability, enabling robust performance even under neural signal distortion or probe drift, particularly when recordings are obtained using planar MEAs. A modified self-organizing map (SOM) algorithm clusters spikes in the spatial domain with minimal memory access, supporting on-chip training and real-time operation with low latency. Fabricated in 40-nm CMOS, the chip achieves 0.00029-mm2/channel area and 74-nW/channel power consumption, with over $1000\times $ data compression. Performance is validated across synthetic and ex vivo datasets containing up to 500 neurons, demonstrating competitive accuracy and robust drift tracking compared to state-of-the-art solutions with much lower data bandwidth, processing, and power demands.