| 研究課題/領域番号 |
21K11809
|
|---|
| 研究種目 |
基盤研究(C)
|
| 配分区分 | 基金 |
|---|
| 応募区分 | 一般 |
|---|
| 審査区分 |
小区分60040:計算機システム関連
|
|---|
| 研究機関 | 奈良先端科学技術大学院大学 |
|---|
研究代表者 |
ZHANG Renyuan 奈良先端科学技術大学院大学, 先端科学技術研究科, 准教授 (00709131)
|
|---|
| 研究分担者 |
木村 睦 奈良先端科学技術大学院大学, 先端科学技術研究科, 客員教授 (60368032)
|
|---|
| 研究期間 (年度) |
2021-04-01 – 2024-03-31
|
| 研究課題ステータス |
交付 (2022年度)
|
| 配分額 *注記 |
4,160千円 (直接経費: 3,200千円、間接経費: 960千円)
2023年度: 1,300千円 (直接経費: 1,000千円、間接経費: 300千円)
2022年度: 1,430千円 (直接経費: 1,100千円、間接経費: 330千円)
2021年度: 1,430千円 (直接経費: 1,100千円、間接経費: 330千円)
|
|---|
| キーワード | Non-deterministic / bisection neural network / re-configurability / efficiency / Continuous domain / parameter reduction / 確率計算 / CGRA / スパイクベース計算 |
|---|
| 研究開始時の研究の概要 |
本研究では、従来のデジタル厳密計算基盤からAI向け超並列曖昧計算基盤まで対応できる時・空再構成可能な演算機構の基礎研究を行う。空間軸再構成に対して独創的な二分木ニューラルネットワークにより製造後任意に解体・組立できる演算器アレイを構築する。時間軸再構成に対して非決定論的計測に基づく確率的スパイクベース計算方式を創出する。精度の制御が可能な仕組みを導入し、両者の一体化を進める。最終には精度を調整できる無駄のない厳密・非厳密混合計算基盤の実現を目指す。さらに、メムキャパシタ等新機能デバイス実装技術を加えた、開発される計算機構の小型化を探索対象とする。
|
| 研究実績の概要 |
In this year, both of time- and space-reconfigurable computing technologies are explored in deep as planned. 1. For time-reconfigurable computing technologies, we focus on the non-deterministic computing applications in various fields such as medicine and wireless communications. By applying the proposed stochastic computing scheme, the quality of service and robustness in some real-world scenarios are both superior to the world top performances. 2. For space reconfigurable computing technologies, the DiaNet series (the third version) have been applied in various AI tasks and archived fair or superior performances with greatly reduced cost.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The current progress fully matches my initial proposal. All of three tiers including mechanism, circuits, and application levels have been explored, and the performances of our developed platforms are superior in some specific features. Moreover, a new technology for data-coding was developed beyond the initial plan, which accelerates the computations greatly. The relevant research progresses were published on world top-class transactions and conference such as IEEE TNNLS and IJCAI. The proposed technologies appear potentials on solving the real-world problems such as medicine and wireless communications. The next step of this project is well indicated on the basis of progress of this year. From the current results, the “calculator free NN inference” becomes feasible as initially planned.
|
| 今後の研究の推進方策 |
Firstly, the quantum-spike coding methodology will be explored. We are going to start from some toy-examples such as conventional neural networks. Two schemes including one-shot observation and statistic observation are verified to perform regression and pattern recognition. Then, we will migrate this coding methodology into our DiaNet. Secondly and simultaneously, more series of DiaNet (so far, till version 3.1) and flash computing architecture are expected to evolve. As soon as above techniques ready, we might migrate some existing tensor computing structures such as systolic ring by partitioning the DiaNet into reasonable pieces. As the further step rooting on this project, it is expected to develop the CMOS-superconductor hybrid computing platforms.
|
