ELECTRONIC APPARATUS AND COMPRESSION METHOD FOR ARTIFICIAL NEURAL NETWORK

專利類型

發明

專利國別 (專利申請國家)

美國

專利申請案號

16/253,229

專利證號

US 11,270,207

專利獲證名稱

ELECTRONIC APPARATUS AND COMPRESSION METHOD FOR ARTIFICIAL NEURAL NETWORK

專利所屬機關 (申請機關)

國立清華大學

獲證日期

2022/03/08

技術說明

經過深度學習演算法訓練之後，得到的深度神經網路模型，能夠解決很多視覺相關的應用問題，但是通常需要很高的運算量。本技術提出了一個新的壓縮深度神經網路模型的機制，此機制包含一個特別設計的模組，能夠刪除多餘的卷積參數並減少浮點數運算，讓模型得以簡化，如此將有助於 AI 晶片的設計，並且可進一步佈建於終端設備上進行運算。 Techniques based on deep neural networks have been very successful for a wide variety of applications in artificial intelligence, but the good performance often comes at a price. A deep-net model typically comprises a huge number of parameters that may not be efficient for run-time computation and also take up large amounts of storage. Such concerns have not been mitigated by the fact that most of the state-of-the-art deep learning methods tend to adopt deeper and complicated architecture to improve the accuracy. Meanwhile, there is also an immediate need for lightweight and efficient versions of deep networks that can be easily ported to and run on embedded systems or mobile devices. We aim to develop a cost-aware method that can reduce the size of a given network significantly, without compromising its expected performance.

備註

連絡單位 (專責單位/部門名稱)

智財技轉組

連絡電話

03-5715131-62219

網址