Deep Neural Network (DNN)

Zoom

Deep neural network (DNN) has been applied to various fields such as computer vision, natural language processing, recommendation system, and more. The latest DNN models have better performance than humans, but they require huge amount of computational complexity and data storage space. Due to its large input datasets and parameters for operations, a high-speed and low-power(energy) processor design to accelerate DNN operations becomes very important. Our research focus is a DNN accelerator design for energy efficient inference/training of various DNN models by applying the following techniques.

• Dataflow which maximizes throughput per energy consumption
• Sparsity-aware hardware architecture
• Approximate computing techniques for energy efficient edge/mobile applications
• Various hardware design techniques based on domain change
• Training architecture optimization
• Reconfigurable DNN accelerator which can adapt to various task specifications

Spiking Neural Network (SNN)

Zoom

Spiking neural network (SNN) is a brain-inspired neural network to realize energy-efficient neuromorphic systems for Edge and IoT applications. Over the last decades, SNN researches have been trying to improve energy efficiencies of neural networks by utilizing sparse, event-driven characteristics found in the biological brain. The main research focus of SNN team ranges from hardware-friendly SNN algorithm design to low cost SNN accelerator chip implementation.

• Neuron model and neural coding optimization for hardware
• Approximate techniques for low energy inference
• Lightweight training methods for efficient on-chip learning hardware
• Preprocessing method for improving the accuracy of the networks
• Dataflow and hardware architecture for low power, area, and latency implementation