Low power multi-bit weight vector operated SRAM cell array machine learning classifier for an era of AI anywhere
| Project/Area Number |
18K11230
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Multi-year Fund |
|---|
| Section | 一般 |
|---|
| Review Section |
Basic Section 60040:Computer system-related
|
|---|
| Research Institution | Fukuoka Institute of Technology |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2018-04-01 – 2021-03-31
|
| Project Status |
Completed (Fiscal Year 2020)
|
| Budget Amount *help |
¥4,550,000 (Direct Cost: ¥3,500,000、Indirect Cost: ¥1,050,000)
Fiscal Year 2020: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Fiscal Year 2019: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Fiscal Year 2018: ¥1,950,000 (Direct Cost: ¥1,500,000、Indirect Cost: ¥450,000)
|
|---|
| Keywords | 省電力機械学習 / SRAM内機械学習 / メモリ内機械学習 / 量子化機械学習 / １ビット機械学習 / 機械学習識別器 / SRAMセルアレイ機械学習 / 重み量子化アルゴリズム / 入出力セル型 / スケーリング / ミックスドシグナル機械学習 / 機械学習器 / SRAM機械学習器 / 多ビット重み |
|---|
| Outline of Final Research Achievements |
This research demonstrated that decoupling read-port from write-port is the key to a better stability for the machine learning operation because the product operations between the input vectors and weight value can be isolated form the storage nodes of the SRAM cells. Both (1) the number of multiple parallel connected transistors for the read-port and (2) the pulse width applied to the transistors can be used for the adjustment of amount of the product value without any stability issues. Because BL can be shorten without any stability issues. This removes the requirement for the suppressing WL voltage level, thus lower voltage operation is enabled. New algorithm for binary quantization reduced the instability of the weight learning curve thanks to regularizing the weight variation range.
The weights are regularized within the same WL unit. This reduced the accuracy degradation due to binary quantization error. This eliminated the requirements for the expensive ensemble learning.
|
| Academic Significance and Societal Importance of the Research Achievements |
本研究は「特徴ベクトル×1-bit重みベクトルの内積計算をSRAMセルアレイの読み出し動作で可能にする機械学習識別器に関するもので提案技術により「特徴ベクトル×多bit重みベクトルの内積和を短絡ビット線電流に反映できる。量子化アルゴリズムは「特徴ベクトル行ごとに量子化することで1/0に丸められる確率を減らし、多bitの本来の精度を維持できる」結果、コストを犠牲にしても必要だった「精度補償用の集団学習」を不要にし、大量のセルアレイを削減できる。これにより、本研究の目指す「どこでもAIに向けて必須の、省電力化機械学習識別器」に必要な、精度とコスト（消費電力、面積）のトレードオフの関係が改善される。
|
Report
(4 results)
Research Products
(13 results)
