Kyoung-hoon Kim

Dynamic energy-accuracy trade-off using stochastic computing in deep neural networks

This paper presents an efficient DNN design with stochastic computing. Observing that directly adopting stochastic computing to DNN has some challenges including random error fluctuation, range limitation, and overhead in accumulation, we address these problems by removing near-zero weights, applying weight-scaling, and integrating the activation function with the accumulator. The approach allows an easy implementation of early decision termination with a fixed hardware design by exploiting the progressive precision characteristics of stochastic computing, which was not easy with existing approaches. Experimental results show that our approach outperforms the conventional binary logic in terms of gate area, latency, and power consumption.

Highly reliable vertical NAND technology with biconcave shaped storage layer and leakage controllable offset structure

The performance and reliability of 3-D NAND cells fabricated by TCAT (Terabit Cell Array Transistor) technology have been improved significantly via a damascened metal gates and a controlled offset between BL contact and select transistor. The damascened metal gate providing sufficiently low resistance is achieved by adopting a novel metal process. Highly suppressed disturbance property is achieved by the appropriate offset which reduces the leakage current through the select transistor. It is proved that the TCAT NAND is a manufacturable technology in terms of reliability as well as performance in a channel hole with a diameter of 90nm.

A Cell-permeable Fusion Protein for the Mineralization of Human Dental Pulp Stem Cells:

Human dental pulp stem cells (hDPSCs) are the only mesenchymal stem cells in pulp tissue that can differentiate into osteoblasts, odontoblasts, and adipose cells. The transcriptional co-activator with PDZ-binding motif (TAZ) protein has been reported to modulate osteogenic differentiation in mouse MSCs. Therefore, we examined whether the TAZ protein plays the same role in human pulp stem cells. In this study, TAZ was applied to cells directly with low-molecular-weight protamine (LMWP) as a cell-penetrating peptide (CPP). The LMWP-TAZ fusion proteins were expressed in an E. coli system with a pET-21b vector and efficiently transferred into hDPSCs without producing toxicity in the cells. The efficient uptake of TAZ was shown by Western blot with an anti-TAZ antibody, fluorescence-activated cell sorting, and confocal microscopy in live cells. The delivered TAZ protein increased osteogenic differentiation, as confirmed by alkaline phosphatase (ALP) staining, RT-PCR, and Western blotting. In addition, TAZ also inhibited adipogenic differentiation, regulating peroxisome proliferator-activated receptor-γ (PPAR-γ), lipoprotein lipase (LPL), and adipocyte fatty acid-binding protein (aP2) mRNA levels. These in vitro studies suggest that cell-permeable TAZ may be used as a specific regulator of hard-tissue differentiation.

Approximate de-randomizer for stochastic circuits

De-randomizer is one of the most important components in stochastic computing. We suggest an approximate parallel counter for the de-randomizer generating a small number of errors, which outperforms a conventional parallel counter in terms of area, delay, and power.

An energy-efficient random number generator for stochastic circuits

Stochastic circuits provide very high efficiency in terms of gate area and power consumption compared with conventional binary logic. However, they require random bit streams generated by stochastic number generators (SNGs), which account for a significant portion of area and energy offsetting their merits. In this paper, we propose a new SNG that significantly reduces area and energy while improving accuracy in progressive precision. Experimental results show that the proposed SNG reduces energy by more than 72% compared to the state-of-the-art designs.

A design framework for hierarchical ensemble of multiple feature extractors and multiple classifiers

It is well-known that ensemble of classifiers can achieve higher accuracy compared to a single classifier system. This paper pays attention to ensemble systems consisting of multiple feature extractors and multiple classifiers (MFMC). However, MFMC increases the system complexity dramatically, leading to a highly complex combinatorial optimization problem. In order to overcome the complexity while exploiting the diversity of MFMC, we suggest in this paper a hierarchical ensemble of MFMC and its optimizing framework. By constructing local groups of feature extractors and classifiers and then combining them as a global group, the approach achieves a better scalability. Both reinforcement machine learning and Bayesian networks are adopted to enhance the accuracy. We apply the proposed method to vision based pedestrian detection and recognition of handwritten numerals. Experimental results show that the proposed framework outperforms the previous ensemble methods in terms of accuracy. Optimization of MFMC (multiple feature-extractor, multiple classifier) systems.Presentation of a general design framework for an ensemble of MFMC.Proposing a hierarchical approach for reducing the complexity of MFMC optimization.Proposing a new approach that integrates reinforcement learning and Bayesian network.Experimental results show that the proposed framework outperforms previous approaches.

An Efficient and Accurate Stochastic Number Generator Using Even-distribution Coding

Stochastic computing (SC) is a promising approach for low-power and low-cost applications with the added benefit of high error tolerance. However, the high overhead of generating stochastic bitstreams can offset the advantages of SC especially when a large number of bitstreams are needed. In this paper, we propose a new stochastic number generator (SNG) that significantly reduces area and energy while improving accuracy. Experimental results show that the proposed SNG can reduce energy by more than 72% compared with the state-of-the-art designs.

Concept-aware ensemble system for pedestrian detection

For pedestrian detection in ADAS, using multiple classifiers generally performs better than using a single classifier in terms of accuracy since the classifiers can be made to complement one another. On the other hand, such a pedestrian detector needs to be tuned dynamically to the variation of real-world environment such as different poses of pedestrians and variable background. Thus the system is requested to incrementally accept new information while retaining the old one. This paper presents an environment-adaptive ensemble system that performs incremental learning for pedestrian detection. It combines a pedestrian detector comprised of multiple classifiers with a front-end concept recognizer that selectively turns on and off the member classifiers adaptively according to the recognized concept of the input image. It adopts an incremental learning algorithm to add a new classifier, which is trained with a newly added batch of dataset, to the existing ensemble. With the intervention of the front-end concept recognizer, the system can retain good accuracy for old environments while not losing the focus on current environment.

SoC Architecture for Automobile Vision System

Advanced Driver Assistance System (ADAS) is becoming more and more popular and increasing its importance in a car with the advancement in electronics and computer engineering that provides key enabling technologies for such a system. Among others, vision is one of the most important technologies since the current practice of automotive driving is mostly, if not entirely, based on vision. This chapter discusses architectural issues to be considered when designing Systems-on-a-Chip (SoC) for automobile vision system. Various existing architectures are introduced together with some analysis and comparison.