Kioxia Presented Image Classification System Deploying Memory-Centric AI with High-capacity Storage at ECCV 2022

Kioxia Corporation, the world chief in reminiscence options, has developed a picture classification system primarily based on Memory-Centric AI, an AI know-how that makes use of high-capacity storage. The system classifies pictures utilizing a neural community that refers to information saved in exterior high-capacity storage; this avoids “catastrophic forgetting,” one of many main challenges of neural networks, and permits information to be added or up to date with out the lack of present information. This know-how was introduced on October 25 at the oral session of European Conference on Computer Vision 2022 (ECCV 2022) in Tel Aviv, one of many prime conferences within the area of laptop imaginative and prescient^[1].

This press launch options multimedia. View the complete launch right here: https://www.businesswire.com/news/home/20221101006375/en/

Image classification using high-capacity storage (Graphic: Business Wire)

In typical AI methods, neural networks are educated to amass information by updating parameters known as “weights.” Once totally educated, to be able to purchase new information a neural community should be both re-trained from the start or fine-tuned with new knowledge. The former requires big quantities of time and consumes vital vitality prices, whereas the latter requires parameters to be up to date and faces the catastrophic forgetting drawback of shedding the information acquired previously which ends up in deterioration of classification accuracy.

To deal with the problems of value and accuracy in neural network-based picture classification methods, the brand new resolution shops giant quantities of picture knowledge, labels and picture characteristic maps^[2] as information in a high-capacity storage. The neural community then classifies pictures by referring to that saved information (Figure 1). Using this technique, information may be added or up to date by including newly obtained picture labels and have maps to the saved knowledge. As there isn’t any must re-train or replace weights, which can trigger “catastrophic forgetting,” picture classification may be maintained extra precisely.

Furthermore, by utilizing the info referred from the storage when the neural community classifies pictures, the idea for the classification outcomes may be visualized, which is anticipated to enhance the explainability of AI^[3] and alleviate the black-box drawback^[4], in flip permitting the selective modification of information sources. In addition, by analyzing the referred knowledge, the contribution of every saved knowledge may be evaluated based on the frequency of references.

Guided by its mission of “Uplifting the World with ‘Memory,’” Kioxia will proceed to contribute to the event of AI and storage applied sciences by increasing Memory-Centric AI past picture classification to different areas and selling analysis and growth of AI know-how using high-capacity storage.

About Memory-Centric AI

Introductory Video: https://youtu.be/lw8XKhviGJc

Memory-Centric AI, Part I: How Kioxia’s Top Engineers Are Developing an AI That Relies on Memory https://brand.kioxia.com/en-jp/articles/article25.html

Memory-Centric AI, Part II: An Internet of Memories: Brainstorming Uses for Memory-Centric AI https://brand.kioxia.com/en-jp/articles/article26.html

About ECCV

The European Conference on Computer Vision (ECCV) is among the prime conferences within the area of laptop imaginative and prescient. In latest years, ECCV has established itself as a chief venue for the presentation of AI analysis papers together with picture classification, object detection, and different applied sciences utilizing deep studying. The oral presentation acceptance charge was 2.7% on this yr.

[1] Paper title: Okay. Nakata et. al., “Revisiting a kNN-based Image Classification System with High-capacity Storage”, European Conference on Computer Vision 2022 (ECCV 2022)

[2] Image characteristic maps: multidimensional (e.g., 1,024-dimensional) numerical knowledge obtained by means of neural community operations

[3] Explainability of AI: probably of explaining the idea and causes of outcomes predicted by AI in a manner that may be interpreted by people.

[4] Black-box drawback: the method resulting in the outcomes predicted by AI isn’t interpretable to people, making it a black field drawback.

Company names, product names, and repair names could also be logos of third-party corporations.

About Kioxia

Kioxia is a world chief in reminiscence options, devoted to the event, manufacturing and sale of flash reminiscence and solid-state drives (SSDs). In April 2017, its predecessor Toshiba Memory was spun off from Toshiba Corporation, the corporate that invented NAND flash reminiscence in 1987. Kioxia is dedicated to uplifting the world with reminiscence by providing merchandise, companies and methods that create selection for purchasers and memory-based worth for society. Kioxia’s revolutionary 3D flash reminiscence know-how, BiCS FLASH™, is shaping the way forward for storage in high-density purposes, together with superior smartphones, PCs, SSDs, automotive and knowledge facilities.

View supply model on businesswire.com: https://www.businesswire.com/news/home/20221101006375/en/