Title:
NEURAL NETWORK WEIGHT DISCRETIZING METHOD, SYSTEM AND DEVICE, AND READABLE STORAGE MEDIUM
Document Type and Number:
WIPO Patent Application WO/2019/006976
Kind Code:
A1
Abstract:
Provided are a neural network weight discretizing method, system and device, and a readable storage medium, the method involving: acquiring a weight value range and the number of discrete weight states; acquiring a weight state of the previous time step and a weight increment of the current time step; according to the weight increment of the current time step, using a directional function to acquire a state switching direction; and according to the weight state at the previous time step, the weight increment at the current time step, the state switching direction, the weight value range and the number of discrete weight states, acquiring a weight state of the current time step. The method ensures that the weight values are always constrained in the same discrete-state space, without the need for storing an additional virtual continuous-state implicit weight. On the premise that the computation performance of the neural network is ensured, the consumption of a storage space is greatly reduced, and the computation complexity is reduced.
Inventors:
LI GUOQI (CN)
WU ZHENZHI (CN)
PEI JING (CN)
DENG LEI (CN)
WU ZHENZHI (CN)
PEI JING (CN)
DENG LEI (CN)
Application Number:
PCT/CN2017/114667
Publication Date:
January 10, 2019
Filing Date:
December 05, 2017
Export Citation:
Assignee:
UNIV TSINGHUA (CN)
International Classes:
G06N3/04
Foreign References:
| CN106845632A | 2017-06-13 | |||
| CN102263636A | 2011-11-30 | |||
| CN106875003A | 2017-06-20 | |||
| US7814038B1 | 2010-10-12 |
Attorney, Agent or Firm:
ACIP LAW OFFICES (CN)
Download PDF: