- Placement and Design
Scalable technologies for deep neural networks
Provide AI models suited to calculation capacity in execution environments without retraining.
- Toshiba developed scalable AI that enables the deployment of learned AI in a variety of systems with different computation complexity, while minimizing performance loss.
- Once large-scale, high-performance AI is trained, AI engines can be standardized for different applications, reducing the lead time required for the development of AI engines, and increasing management efficiency.
- Scalable AI training clarifies the relationship between computation complexity and performance, making it easy to choose applicable processors.
Applications
- Applications in which deep learning models for the same tasks are operated on processors and platforms with different computation capacity (e.g., smart phones, security cameras, Automatic Guided Vehicles (AGV)).
Benchmarks, strengths, and track record
- In an evaluation using a 50-layer convolutional network (ResNet-50) with the public data set ImageNet (1,000 categories of image clustering tasks), we confirmed that when the computation complexity is reduced to 1/2, 1/3, and 1/4 of the full-size model, the reduction rate in classification performance can be kept to 1.1% (2.7%), 2.1% (3.9%), and 3.3% (5.0%) respectively (figure in parentheses represents the conventional method), achieving the world’s highest level of performance.
Inquiries
Contact the Toshiba Corporate Research & Development Center
Please include the title “Toshiba AI Technology Catalog: Scalable technologies for deep neural networks” or the URL in the inquiry text.
Please note that because this technology is currently the subject of R&D activities, immediate responses to inquiries may not be possible.
References:
- A. Yaguchi et al., “Decomposable-Net: Scalable Low-Rank Compression for Neural Networks,” Proc. of 30th International Joint Conference on Artificial Intelligence (IJCAI), 2021.
- Toshiba and Riken Develop a World-class Scalable AI Technology that, After Learning, Can Adjust AI Performance and Computational Complexity According to Usage Environment; News release: August 20, 2021
