Title:
ARITHMETIC PROCESSING DEVICE, INFORMATION PROCESSING DEVICE, METHOD AND PROGRAM
Document Type and Number:
WIPO Patent Application WO/2018/139266
Kind Code:
A1
Abstract:
[Problem] To reduce the circuit size and power consumption by precisely executing deep learning in a fixed-point number. [Solution] This arithmetic processing device is provided with: an acquisition circuit which acquires statistical information on a bit distribution in fixed-point number data after executing a command for the fixed-point number data; and an update circuit which updates a decimal point position of the fixed-point number data.
Inventors:
ITO MAKIKO (JP)
TOMONO MITSURU (JP)
ISHIHARA TERUO (JP)
YODA KATSUHIRO (JP)
NOTSU TAKAHIRO (JP)
TOMONO MITSURU (JP)
ISHIHARA TERUO (JP)
YODA KATSUHIRO (JP)
NOTSU TAKAHIRO (JP)
Application Number:
PCT/JP2018/000988
Publication Date:
August 02, 2018
Filing Date:
January 16, 2018
Export Citation:
Assignee:
FUJITSU LTD (JP)
International Classes:
G06F7/499; G06F9/38; G06N3/063
Foreign References:
| JP2006048590A | 2006-02-16 | |||
| JP2016539407A | 2016-12-15 | |||
| JP2016062404A | 2016-04-25 | |||
| JP2006155102A | 2006-06-15 | |||
| JP2016033754A | 2016-03-10 | |||
| JP2012203566A | 2012-10-22 | |||
| JP2010112878A | 2010-05-20 |
Other References:
COURBARIAUX ET AL.: "TRAINING DEEP NEURAL NETWORKS WITH LOW PRECISION MULTIPLICATIONS", ACCEPTED AS A WORKSHOP CONTRIBUTION AT ICLR 2015, INTERNATIONAL CONFERENCE ON LEARNING REPRESENTATIONS(ICLR, 23 September 2015 (2015-09-23)
See also references of EP 3575952A4
See also references of EP 3575952A4
Attorney, Agent or Firm:
MUKOUYAMA Naoki (JP)
Download PDF:
Previous Patent: PROCESSOR, INFORMATION PROCESSING DEVICE, AND PROCESSOR OPERATION METHOD
Next Patent: WIRE HARNESS
Next Patent: WIRE HARNESS