Brainware Large-Scale Vision Hardware Based on Parallel Stochastic Computing

2018 Fiscal Year Final Research Report

• Project/Area Number: 16K12494
• Research Category: Grant-in-Aid for Challenging Exploratory Research
• Allocation Type: Multi-year Fund
• Research Field: Soft computing
• Research Institution: Tohoku University
• Principal Investigator: ONIZAWA Naoya 東北大学, 電気通信研究所, 助教 (90551557)
• Project Period (FY): 2016-04-01 – 2019-03-31
• Keywords: ソフトコンピューティング / 確率的演算

Outline of Final Research Achievements

Recently, brain-inspired vision processing provides better accuracy than human begins in specific applications. However, the current brain-inspired computing works well in only one application previously trained, and hence is lack of flexibility in unknown various applications, unlike human beings.
In this study, a brain-inspired vision-processing filter is realized that can extract millions of different features, like human beings. In addition, as the hardware is implemented using stochastic computing that exhibits a similar behavior to human neurons, the power dissipation is reduced by 97% while maintaining the throughput in comparison with a traditional binary implementation.

Free Research Field

計算機科学

Academic Significance and Societal Importance of the Research Achievements

本研究課題で実現した人間的視覚処理ハードウェアは，人間の脳のように幅広いパターン（1000万パターン以上）の特徴を抽出することが出来る．この特徴データを用いることで，将来的には汎用的な脳型処理・人工知能実現に繋がるものと考えられる．というのも，現在実現されている人工知能（一般的にニューラルネットワーク）は，特徴をタスクに応じて絞っているために，処理の汎用性が失われていると考えられる．つまり将来，汎用人工知能実現に向けて，本研究課題で実現したハードウェアは，その初期段階の処理に当たる部分が実現されたものと考えられ，学術的に重要な意義を有する．