本文へジャンプ

成果報告書詳細
管理番号20190000000438
タイトル*平成30年度中間年報 高効率・高速処理を可能とするAIチップ・次世代コンピューティングの技術開発 革新的AIエッジコンピューティング技術の開発 ソフトテンソルプロセッサによる超広範囲センシングAIエッジ技術の研究開発
公開日2019/6/22
報告書年度2018 - 2018
委託先名沖電気工業株式会社 公立大学法人会津大学 綜合警備保障株式会社 ジャパンマリンユナイテッド株式会社
プロジェクト番号P16007
部署名IoT推進部
和文要約
英文要約Title: Project for Innovative AI Chips and Next-Generation Computing Technology Development/Development of innovative AI edge computing technologies/Wide-area Large-scale Sensing AI edge Technologies with Soft Tensor Processor (FY2018-FY2020) FY2018 Annual Report

(1) Development of soft tensor processor
Based on the analysis of workload, data size, concurrency of convolutional neural networks, we proposed a brain-inspired architecture, called soft tensor processor (STP). It is based on a 4-dimensional tensor structure consisting of multiple chained tensor processing elements (TPEs). Like a human-brain, STP uses the extremely high levels of parallelism and data reusing. We implemented a TPE for a convolutional layer as BinaryNet. We also developed a stub version of an SDK tool to convert deep neural network description written in Chainer framework to control codes. This tool analyzes neural network models dynamically by executing their inference processes using dummy input data. It supports Convolution 2D, Linear, Sigmoid, Reshape, Maxpooling2D layers and will be extended in the next year.

(2) High efficiency modeling technology
We developed the high efficiency modeling technology for deep neural networks and evaluated its effectiveness to representative network models. We pruned channels based on attention statics that enables to evaluate the importance of channels. After pruning channels, we quantized both the activations and weights into 8 bits. Furthermore, we investigated the impact of our pruning channel techniques on the sensitivity maps created by gradient descent method.

(3) Wide-area large-scale recognition technologies
This section discusses the development of object detection methodologies from high resolution camera. We have prepared to collect videos of actual shoplifter from some stores, and studied features relate to human movements. We are also working on the development to detect "fire" and "traffic jam". To improve the detection performances, we extracted visual features of smoke videos for fires and vehicle density and tailback for traffic jam. We achieved high detection accuracy in both fire and traffic jam with some false positives that need to be improved.
For moving object detection, we set 2 goals: robustness for small objects detection and occlusion. For small objects detection, we have developed 2 effective methods namely Upsampling and Subnetwork. For the goal of occlusion handling, we have developed a method to re-assign trajectories in order to effectively perform occluded object tracking governed by Tracking-by-Detection approach. In addition, we have confirmed the possibility of recognition of pedestrian’s conditions. We also have developed the deep learning algorithm to improve the performance to detect objects on the sea in the distance.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る