J.-M. Seo

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.

Measurement of ocular torsion using digital fundus image

Computer-based objective measurement of the ocular cyclotorsion using digital fundus photograph was developed. Color digital fundus photographs acquired with the field angle of 60/spl deg/, 1520 /spl times/ 1080 in resolution were analyzed. Optic disc and macula were segmented by the program developed on MATLAB, which executed the serial analysis of the Otsu threshold, labeling, Canny edge. The angle between the horizontal line that bisects the optic disc and the line connecting the center of optic disc and macula was measured and compared with the torsion determined by the specialist. Optic disc and macula were segmented and the mean of the calculated angle was 3.02/spl plusmn/1.24/spl deg/. The mean of the torsion determined by the specialist was 3.13/spl plusmn/1.98/spl deg/ and there was no difference between the two. The measurement of the cyclotorsion using computer program showed good coincidence with that of the specialist and it can be a good candidate as a tool helping precise diagnosis and the objective evaluation of the disease for the physicians.

Automated quantification of retinal nerve fiber layer atrophy in fundus photograph

Quantitative analysis of the retinal nerve fiber layer (RNFL) defect is prerequisite in the early detection and management of glaucoma. A new automatic quantification method to evaluate the degree of RNLF defect has been proposed in this paper. Simple image processing technique is applied to locate optic disc and intensity of the pixels around optic disc is plotted. The area with RNFL defect can be easily determined by comparing the intensity plot of the RNFL and the first derivative of the intensity plot. Through analysis of the plot, thickness of RNFL also can be postulated.

Natural magnetic nanoparticle containing droplet for smart drug delivery and heat treatment

Biodegradable polymer droplet containing natural magnetic nanoparticle is composed for smart drug delivery and heat treatment. For selective and efficient drug delivery to the target tissue, direct high magnetic field will be applied near the target tissue. For drug release control and heat treatment, alternative high magnetic field will be applied. Magnetosome, natural magnetic nanoparticle, is extracted from magnetotactic bacteria, AMB-1. Mixture of magnetosome and sodium alginate composes into droplet using the microfluidic device applied Plateau-Rayleigh instability principle. The magnetosome contained droplet selected its rout at the bifurcate microchannels by direct high magnetic field. High alternative magnetic field generating circuit is designed with 18 mT and 4 Hz magnetic wave. The generated magnetic wave was applied to the extracted magnetosomes so that temperature of the magnetosomes increased from 15.2°C to 17.6°C.

Long-term In vivo Biocompatibility Evaluation of MEMS Retinal Tack for Retinal Prosthesis

In this paper, long-term (1 year) in vivo biocompatibility evaluation of silicon retinal tack is presented. Such a long-term in vivo experiment of biocompatibility has not been reported so far in the retinal prosthesis research field. Microscopic histological examinations of retina indicate no inflammatory infiltrates, and scanning electron microscope (SEM) images reveal neither chemical reactions nor mechanical fractures of the device. In addition, for the necessity of soft and elastic retinal tack, poly-methyl-methacrylate (PMMA) retinal tack is fabricated and mechanically characterized. The results indicate that PMMA retinal tack is also suitable for mechanical fixation of micro-electrode array and an in vivo experiment is in progress.

A new approach to binarizing neural networks

As deep neural networks grow larger, they suffer from a huge number of weights, and thus reducing the overhead of handling those weights becomes one of key challenges nowadays. This paper presents a new approach to binarizing neural networks, where the weights are pruned and forced to take degenerate binary values. Experimental results show that the proposed approach achieves significant reductions in computation and power consumption at the cost of a slight accuracy loss.