• Proceedings of the KIEE Conference
• /
• 2003.11b
• /
• pp.171-174
• /
• 2003

In this paper, a Learning module for a reconfigurable neural network processor(ERNIE) was proposed for an On-chip learning. The existing reconfigurable neural network processor(ERNIE) has a much better performance than the software program but it doesn't support On-chip learning function. A learning module which is based on Back Propagation algorithm was designed for a help of this weak point. A pipeline structure let the learning module be able to update the weights rapidly and continuously. It was tested with five types of alphabet font to evaluate learning module. It compared with C programed neural network model on PC in calculation speed and correctness of recognition. As a result of this experiment, it can be found that the neural network processor(ERNIE) with learning module decrease the neural network training time efficiently at the same recognition rate compared with software computing based neural network model. This On-chip learning module showed that the reconfigurable neural network processor(ERNIE) could be a evolvable neural network processor which can fine the optimal configuration of network by itself.

• Journal of Electrical Engineering and information Science
• /
• v.3 no.3
• /
• pp.342-347
• /
• 1998

An OptiNeural system is developed for optical pattern classification. It is a novel hybrid system which consists of an optical processor and a multilayer neural network. It takes advantages of two dimensional processing capability of an optical processor and nonlinear mapping capability of a neural network. The optical processor with a binary phase only filter is used as a preprocessor for feature extraction and the neural network is used as a decision system through mapping. OptiNeural system is trained for optical pattern classification by use of a simulated annealing algorithm. Its classification performance for grey tone texture patterns is excellent, while a conventional optical system shows poor classification performance.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.33B no.2
• /
• pp.149-158
• /
• 1996

In this paper we describe designing and implementing a digital neural chip and a parallel neural machine for simulating large scale neural netsorks. The chip is a single-chip multiprocessor which has four digiral neural processors (DNP-II) of the same architecture. Each DNP-II has program memory and data memory, and the chip operates in MIMD (multi-instruction, multi-data) parallel processor. The DNP-II has the instruction set tailored to neural computation. Which can be sed to effectively simulate various neural network models including on-chip learning. The DNP-II facilitates four-way data-driven communication supporting the extensibility of parallel systems. The parallel neural machine consists of a host computer, processor boards, a buffer board and an interface board. Each processor board consists of 8*8 array of DNP-II(equivalently 2*2 neural chips). Each processor board acn be built including linear array, 2-D mesh and 2-D torus. This flexibility supports efficiency of mapping from neural network models into parallel strucgure. The neural system accomplishes the performance of maximum 40 GCPS(giga connection per second) with 16 processor boards.

• The Transactions of the Korean Institute of Electrical Engineers
• /
• v.45 no.3
• /
• pp.444-450
• /
• 1996

A digital hardware architecture for artificial neural network with learning capability is described in this paper. It is a modified hardware architecture known as HANNIBAL(Hardware Architecture for Neural Networks Implementing Back propagation Algorithm Learning). For implementing an efficient neural network hardware, we analyzed various type of multiplier which is major function block of neuro-processor cell. With this result, we design a efficient digital neural network hardware using serial/parallel multiplier, and test the operation. We also analyze the hardware efficiency with logic level simulation. (author). refs., figs., tabs.

• ETRI Journal
• /
• v.42 no.4
• /
• pp.491-504
• /
• 2020

We present AB9, a neural processor for inference acceleration. AB9 consists of a systolic tensor core (STC) neural network accelerator designed to accelerate artificial intelligence applications by exploiting the data reuse and parallelism characteristics inherent in neural networks while providing fast access to large on-chip memory. Complementing the hardware is an intuitive and user-friendly development environment that includes a simulator and an implementation flow that provides a high degree of programmability with a short development time. Along with a 40-TFLOP STC that includes 32k arithmetic units and over 36 MB of on-chip SRAM, our baseline implementation of AB9 consists of a 1-GHz quad-core setup with other various industry-standard peripheral intellectual properties. The acceleration performance and power efficiency were evaluated using YOLOv2, and the results show that AB9 has superior performance and power efficiency to that of a general-purpose graphics processing unit implementation. AB9 has been taped out in the TSMC 28-nm process with a chip size of 17 × 23 ㎟. Delivery is expected later this year.

• Proceedings of the KAIS Fall Conference
• /
• 2008.11a
• /
• pp.191-196
• /
• 2008

The intent of this paper is to describe a neural network structure called dynamic neural processor(DNP), and examine how it can be used in developing a learning scheme for computing robot inverse kinematic transformations. The architecture and learning algorithm of the proposed dynamic neural network structure, the DNP, are described. Computer simulations are provided to demonstrate the effectiveness of the proposed learning using the DNP.

• IE interfaces
• /
• v.10 no.3
• /
• pp.145-153
• /
• 1997

This paper deals with constructing a scheduling system for a mold manufacturing factory. The scheduling system is composed of 4 submodules such as pre-processor, neural network training, neural networks and simulation. Pre-processor analyzes the condition of workshop and generates input data to neural networks. Network training module is performed by using the condition of workshop, performance measures, and dispatching rules. Neural networks module presents the most optimized dispatching rule, based on previous training data according to the current condition of workshop. Simulation module predicts the earliest completion date of a mold by forward scheduling with the presented dispatching rules, and suggests a possible issue date of a material by backward tracking. The system developed shows a great potential when applied in real mold factory for automotive parts.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.30B no.2
• /
• pp.44-50
• /
• 1993

In this paper, we will present an array processor for implementation of digital neural networks. Back-propagation model can be formulated as a consecutive matrix-vector multiplication problem with some prespecified thresholding operation. This operation procedure is suited for the design of an array processor, because it can be recursively and repeatedly executed. Systolic array circuit architecture with Residue Number System is suggested to realize the efficient arithmetic circuit for matrix-vector multiplication and compute sigmoid function. The proposed design method would expect to adopt for the application field of neural networks, because it can be realized to currently developed VLSI technology.

• Journal of IKEEE
• /
• v.21 no.1
• /
• pp.92-95
• /
• 2017

In this paper, we proposed a way to improve function of small scale GP-GPU. Instead of using superscalar which increase scheduling-complexity, we suggested the application of simple core to maximize GP-GPU performance. Our studies also demonstrated that simplified Stream Processor is one of the way to achieve functional improvement in GP-GPU. In addition, we found that developing of optimal thread-assigning method in Warp Scheduler for specific application improves functional performance of GP-GPU. For examination of GP-GPU functional performance, we suggested the thread-assigning way which coordinated with Deep-Learning system; a part of Neural Network. As a result, we found that functional index in algorithm of Neural Network was increased to 90%, 98% compared with Intel CPU and ARM cortex-A15 4 core respectively.

• Journal of the Institute of Electronics Engineers of Korea SD
• /
• v.41 no.5
• /
• pp.61-70
• /
• 2004

The operation mode of neural network is divided into learning and recognition process. Learning is updating process of weight until neural network archives target result from input pattern. Recognition is arithmetic process of input pattern and weight. Traditional inner product process is focused to improve processing speed and hardware complexity. There is no hardware architecture to distinguish between loaming and recognition mode of neural network. In this paper we propose the new architecture of low power inner product processor for reconfigurable neural network. The proposed architecture is similar with bit-serial inner product processor on learning mode. It have several advantages which are fast processing base on bit-level, suitability of hardware implementation and pipeline architecture to compute data. And proposed architecture minimizes active units and reduces consumption power on recognition mode. Result of simulation shows that active units is depend on bit representation of weight, but we can reduce active units about 50 precent.