Syuichi Tai

GaAs/AlGaAs Optical synaptic Interconnection Device for Neural Networks

A GaAs/AlGaAs optical synaptic interconnection device for neural networks is reported for the first time to our knowledge. This device consists of a light-emitting-diode array, an interconnection matrix, and a photodiode array, which are integrated into a hybrid-layered structure on a GaAs substrate. The device structure and characteristics are reported in detail. The fabricated device can simulate a 32-neuron system. Experimental results of the Hopfield associative memory with three stored vectors are also described.

Variable-sensitivity photodetector that uses a metal-semiconductor-metal structure for optical neural networks.

A novel type of photodetector called a variable-sensitivity photodetector has been developed for optical implementation of neural networks. It utilizes a metal-semiconductor-metal structure whose quantum efficiency can be modulated by an applied bias voltage. A linear dependence of the sensitivity on the bias voltage was obtained with the bipolar current flow. This device operated as a multiplier of the incident light intensity and the bias voltage. It is shown that this device is suitable for achieving dynamic synaptic interconnections. A 4 x 4 array device was fabricated and demonstrated.

Optical learning neurochip with internal analog memory

An optical neurochip with learning capability and a memory function is reported for the first time, to our knowledge. The neurochip is a three-dimensional optoelectronic integrated circuit consisting of a light-emitting diode array and a variable-sensitivity photodetector array. The principle of operation and the fundamental characteristics of the neurochip are described in detail. The synaptic weights, which, are acquired through learning, are stored in the variable-sensitivity photodetector as the detection sensitivity with analog values. Both the positive and the negative synaptic weights are memorized with one variable-sensitivity photodetector element by changing the polarity of the electric signal applied to the photodetector. A storage time of ~20 min. was obtained. With a fabricated optical neurochip that had 32 neurons and 32 x 32 synapses, experiments of on-chip learning for pattern classification were performed successfully.

Variable sensitivity photodetector for optical neural networks

A novel type of photodetector called a variable sensitivity photodetector (VSPD) has been developed for an optical implementation of a neural network. The function of the VSPD is a photodetector with functions of both a spatial light modulator and a photodetector. The VSPD with a metal-semiconductor-metal structure is shown to be suitable for the implementation of the analog synaptic weight with a bipolar value. A two-dimensional 8*8 array of VSPDs was fabricated on a GaAs substrate and integrated with a light-emitting-diode array with eight lines in hybrid form. This chip can simulate an eight-neuron system with 64 variable synaptic connections. The application to pattern classification was demonstrated by using the chip. The chip was used to classify 12 patterns into three categories by the backpropagation learning algorithm in a three-layered network. The authors studied how the nonlinear and nonuniform characteristics of the VSPD array device affect the learning performance. >

Optical neurochip based on a three-layered feed-forward model

We report on a GaAs/AlGaAs optical neurochip based on a three-layered feed-forward model. The optical neurochip consists of a light-emitting diode array with 66 elements, a fixed interconnection matrix, and a photodiode array with 110 elements. The interconnection matrix is determined by the backpropagation learning rule with three quantized levels. There are 35, 29, and 26 neurons, respectively, in the input, hidden, and output layers. The excitatory and inhibitory synapses are integrated on one chip. By using the chip and external electronics, we have succeeded in the recognition of 10 characters with 5 x 7 bits.

Proposal of an Optical Neurochip with Internal Analogue Memory and Its Fundamental Characteristics

An optical neurochip incorporating an analogue memory has been reported. The proposed neurochip is a three-dimensional optoelectronic integrated circuit consisting of a light-emitting diode array and a variable sensitivity photodetector array. In this letter, the preliminary experimental results are described. First, it is shown that the synaptic weights are stored in the variable sensitivity photodetector as the detection sensitivity with analogue values. Both the positive and the negative synaptic weights are memorized by changing the polarity of the electric signal which is applied in the write phase. Second, a novel type of matrix addressing method using the memory function is demonstrated for the realization of a large-scale optical neurochip.

Optoelectronic associative memory using an advanced optical neurochip.

A drastic improvement in the performance of the associative memory was achieved using an optical neurochip. The experimental recognition rate agreed well with computer simulation for the associative memory.

Characteristics of an optoelectronic analogue memory in a GaAs metal-semiconductor-metal photodetector*

The characteristics of a novel optoelectronic analogue memory effect in GaAs metal-semiconductor-metal photodetectors (MSM-PDs) were investigated. The MSM-PDs consisted of GaAs epitaxial layers and aluminum electrodes. It was found that the optoelectronic analogue memory effect was caused by the interfacial states which were formed by defects introduced close to the interface of the GaAs layer during the aluminum deposition, and we can control the characteristics of the effects by the conditions of crystal growth and aluminum evaporation.

Character recognition using a dynamic opto-electronic neural network with unipolar binary weights

A novel quantized learning rule with unipolar binary weights which is useful to simplify the artificial neural hardware is reported. An input-dependent thresholding operation is also proposed to remove the unwanted effect due to insufficient contrast ratio of spatial light modulations as a synaptic connection device. The recognition of 26 characters of the alphabet by the single set of an optoelectronic three-layered network was demonstrated experimentally