Accurate deep neural network inference using computational phase-change memory
- Submitting institution
-
King's College London
- Unit of assessment
- 12 - Engineering
- Output identifier
- 139419768
- Type
- D - Journal article
- DOI
-
10.1038/s41467-020-16108-9
- Title of journal
- Nature Communications
- Article number
- 2473
- First page
- -
- Volume
- 11
- Issue
- 1
- ISSN
- 2041-1723
- Open access status
- Compliant
- Month of publication
- May
- Year of publication
- 2020
- URL
-
-
- Supplementary information
-
-
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- No
- Number of additional authors
-
9
- Research group(s)
-
-
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- This industrially co-authored paper experimentally demonstrated for the first time that imperfect but energy-efficient nanoscale memory devices can be used for on-chip inference achieving close to software-equivalent accuracy. It discusses novel methods to incorporate experimentally-measured device noise and compensation schemes during software training of deep neural networks. The paper reported the highest accuracy experimentally reported to-date by any analogue resistive memory hardware (based on the CIFAR-10 image classification dataset) and the feasibility to implement large image classification problems such as Imagenet. The paper was covered in a blog by IBM Research. The first author is Rajendran’s PhD student at KCL.
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -