Real-time neural pulse-pattern feature extraction and recognition

Abstract

Department of Electrical hgineering, II. FEATURE EXTRACT10 Nlhispapcrprcsentstbealgori~andirqplemnta~onofthenal-timefeatllre extraction and pattctn recognition for signals from culturtd living neural-cell network. FkaWc extraction and pattem recognition are achieved by the application of data compression techniques and an artificial neural network (ANN). Iht implmtation cmists of an 80386-based F C and a TMS32Oc30 digital signal processing @SP) card that is inserted into an expansion dot in the PC. Off-line trainfng of the ANN is done in the PC. R c a l -t i m e~o f t h e p u l s e p a m r n s~m t b e~v i n g n e u r a l c c l l s i s~ in tbe DSP sub-system Iht recognition results are sent from the DSP subsystem to the PC in real time for displaying and recording. -A living neural cell fires by sending electrical impulses to its neighbors through synaptic junctions. With advanced VLSI trchnology, neuroscientists areabletomoahoraadstimulatetheactionpotentialsofcultund,livingne~al cells [1].lbstpdythcbchaviorofneurons, itisessentialu,recognizethefiring pat" especially in real time. & " of tbt fyino patterns can be achieved by the application of a speach c c m g d t h algorithm which is basically a template-matching system. Butitrequinsin~~ecalculationtocomparetheincomingpattans withall t b e~l a t c s t o f~t h e leastma. Italsorequireslargeamouatsofmmoryto store all the sampled data before they are converted to linear predictive code o. To be time-and storage-efficient, a different algorithmis introduced in which data compnssion is done in real time and an ANN [23 is applied to This paper describes the application of an ANN and a digital signal processing @SP) system [3] for faing-pattem recognition. Due to the nature of arallel proccsstng of a neural network and the calculating power of the D& chip, real-& recognition of the signals from the living neural cells becomcs realizable.The irnpltmentation includes two phases. The training phase teaches the supervised ANN to recognize sample pattems. The recognition phase uses the well-trained ANN to idenw pulse trains. F e a m extraction which converts, condenses, and normalizes the incoming signals to a predefined format is used to extract the feature of the firing patterns. The feature of the f h g patterns are catalogued into 16 templates which are U@ to train the supervised ANN. "he ANN is a two-layer multiple-input, multiple-output (MIMO) supervised network. It has 32 nodes in b e input layer, 12 nodes in the hidden layer, and 4 nodes in the output layer.Since binary codes are used as the desired values, four nodes in the output layer can be used to represent 1 6 templates. The off-line recursive training is implemented in the PC.After training, the weights of the trained ANN is used by the ANN of the s a m configuration in the recognition system. The recognition system which is implemnted in the DSP subsystem includes real-time feature extraction and ANN recognition. Feature extraction which is used to prepare the template patterns is also applied in the recognit...