Abstract:RESUME Les interfaces cerveau-machines (ICM) sont de plus en plus importantes dans la recherche biomedicale et ses applications, tels que les tests et analyses medicaux en laboratoire, la cerebrologie et le traitement des dysfonctions neuromusculaires. Les ICM en general et les dispositifs d'enregistrement neuronaux, en particulier, dependent fortement des methodes de traitement de signaux utilisees pour fournir aux utilisateurs des renseignements sur l’etat de diverses fonctions du cerveau. Les dispositifs d'enregistrement neuronaux courants integrent de nombreux canaux paralleles produisant ainsi une enorme quantite de donnees. Celles-ci sont difficiles a transmettre, peuvent manquer une information precieuse des signaux enregistres et limitent la capacite de traitement sur puce. Une amelioration de fonctions de traitement du signal est necessaire pour s’assurer que les dispositifs d'enregistrements neuronaux peuvent faire face a l'augmentation rapide des exigences de taille de donnees et de precision requise de traitement. Cette these regroupe deux approches principales de traitement du signal - la compression et la reduction de donnees - pour les dispositifs d'enregistrement neuronaux. Tout d'abord, l’echantillonnage comprime (AC) pour la compression du signal neuronal a ete utilise. Ceci implique l’usage d’une matrice de mesure deterministe basee sur un partitionnement selon le minimum de la distance Euclidienne ou celle de la distance de Manhattan (MDC). Nous avons comprime les signaux neuronaux clairsemmes (Sparse) et non-clairsemmes et les avons reconstruit avec une marge d'erreur minimale en utilisant la matrice MDC construite plutot. La reduction de donnees provenant de signaux neuronaux requiert la detection et le classement de potentiels d’actions (PA, ou spikes) lesquelles etaient realisees en se servant de la methode d’appariement de formes (templates) avec l'inference bayesienne (Bayesian inference based template matching - BBTM). Par comparaison avec les methodes fondees sur l'amplitude, sur le niveau d’energie ou sur l’appariement de formes, la BBTM a une haute precision de detection, en particulier pour les signaux a faible rapport signal-bruit et peut separer les potentiels d’actions recus a partir des differents neurones et qui chevauchent. Ainsi, la BBTM peut automatiquement produire les appariements de formes necessaires avec une complexite de calculs relativement faible.----------ABSTRACT Brain-Machine Interfaces (BMIs) are increasingly important in biomedical research and health care applications, such as medical laboratory tests and analyses, cerebrology, and complementary treatment of neuromuscular disorders. BMIs, and neural recording devices in particular, rely heavily on signal processing methods to provide users with nformation. Current neural recording devices integrate many parallel channels, which produce a huge amount of data that is difficult to transmit, cannot guarantee the quality of the recorded signals and may limit on-chip signal processing capabilities. An improved signal processing system is needed to ensure that neural recording devices can cope with rapidly increasing data size and accuracy requirements. This thesis focused on two signal processing approaches – signal compression and reduction – for neural recording devices. First, compressed sensing (CS) was employed for neural signal compression, using a minimum Euclidean or Manhattan distance cluster-based (MDC) deterministic sensing matrix. Sparse and non-sparse neural signals were substantially compressed and later reconstructed with minimal error using the built MDC matrix. Neural signal reduction required spike detection and sorting, which was conducted using a Bayesian inference-based template matching (BBTM) method. Compared with amplitude-based, energy-based, and some other template matching methods, BBTM has high detection accuracy, especially for low signal-to-noise ratio signals, and can separate overlapping spikes acquired from different neurons. In addition, BBTM can automatically generate the needed templates with relatively low system complexity. Finally, a digital online adaptive neural signal processing system, including spike detector and CS-based compressor, was designed. Both single and multi-channel solutions were implemented and evaluated. Compared with the signal processing systems in current use, the proposed signal processing system can efficiently compress a large number of sampled data and recover original signals with a small reconstruction error; also it has low power consumption and a small silicon area. The completed prototype shows considerable promise for application in a wide range of neural recording interfaces.

Deep Compressive Autoencoder for Action Potential Compression in Large-Scale Neural Recording

High-fidelity compression of extracellular recordings from motor cortex

Compression of EMG Signals Using Deep Convolutional Autoencoders

Efficient In Vivo Neural Signal Compression Using an Autoencoder-based Neural Network

An Event-based Neural Compressive Telemetry with >11× Loss-less Data Reduction for High-bandwidth Intracortical Brain Computer Interfaces

Large-scale electrophysiology: acquisition, compression, encryption, and storage of big data

Selective and Compressive Sensing for Energy-Efficient Implantable Neural Decoding.

Selective CS: an Energy-Efficient Sensing Architecture for Wireless Implantable Neural Decoding.

A high‐performance asynchronous readout circuit with cascade function for a neural recording micro‐electrode array

Robust Transcoding Sensory Information with Neural Spikes

A Hybrid Autoencoder Framework of Dimensionality Reduction for Brain-Computer Interface Decoding

Electroencephalogram Sensor Data Compression Using An Asymmetrical Sparse Autoencoder With A Discrete Cosine Transform Layer

Online spike sorting via deep contractive autoencoder

Efficient Compressed Sensing for Wireless Neural Recording: A Deep Learning Approach

Computationally efficient simulation of extracellular recordings with multielectrode arrays

Compressive Sensing and Multichannel Spike Detection for Neuro-Recording Systems

Fast 2D Bicephalous Convolutional Autoencoder for Compressing 3D Time Projection Chamber Data

Multi-channel LFP Recording Data Compression Scheme Using Cooperative PCA and Kalman Filter.

EM-Compressor: Electron Microscopy Image Compression in Connectomics with Variational Autoencoders

Design Of A Low-Noise, High Power Efficiency Neural Recording Front-End With An Integrated Real-Time Compressed Sensing Unit

Efficient Data Compression for 3D Sparse TPC via Bicephalous Convolutional Autoencoder