Baidu Rolls Deep Learning Platform
Two software technologies advancing in parallel are merging in a platform designed to use low-code application development to ease development of machine learning models.
Baidu, the Chinese e-commerce giant and AI pioneer, unveiled its EZDL "service” platform this week during an industry event in San Francisco, promoting it as a way to build custom machine learning models using a drag-and-drop interface. The platform focuses on three primary capabilities: image and sound classification, the latter including voice recognition, along with object detection.
The service platform builds on Baidu’s aggressive push into AI development, including the July release of its Brain 3.0 platform billed as a full AI “tech stack” spanning hardware to deep learning frameworks. The company claims more than 600,000 developers have used AI platform, and the no-code framework released this week seeks to expand its customer base to small and mid-sized companies with limited programming experience and computing resources.
The no-code approach aims to reduce deep learning training to as few as four steps that can then be deployed using “only a small amount of data,” said Yongkang Xie, Baidu’s lead technologist for EZDL. “We seek to create a true ecosystem for AI, democratizing access to AI capabilities.”
The image classification model provides automated classification via custom labels, the Chinese company said. In one industrial application, Baidu (NASDAQ: BIDU) said a keyboard manufacturer used the tool to train an image recognition model based on several hundred sample images. The model was then deployed to detect misplaced or missing assembly parts with an accuracy of about 95 percent.
The object detection capability spots and counts the number of objects in an image based on labeling. Baidu said the capability has been used in the retail and medical sectors, including cell counting in microscopy imagery from blood tests.
Meanwhile, voice recognition is among the applications of Baidu’s sound classification model, including scientific research used to distinguished the sounds made by different animal species.
Baidu has been steadily expanding its AI partnerships with U.S. chip makers. In July, it announced a collaboration with Intel Corp. (NASDAQ: INTC) on FPGA-based workload acceleration and a deep learning framework based on Baidu’s open-source PaddlePaddle platform. Baidu previously announced plans to tailor its open source deep learning framework to Volta GPUs from Nvidia (NASDAQ: NVDA) and bring AI capabilities to the Chinese consumer market.
Baidu claims its no-code approach has boiled down the process of developing deep learning models to four steps: model creation; uploading and labeling image or objects; training and testing the model; and deploying the trained model with, in the case of enterprise applications, a cloud API.
The new AI tool is essentially free, with limitations. After their model is deployed, users can get 500 requests/day and up to 2 queries/second for free. If this cannot meet their needs they can work out an arrangement with Baidu (via ezdl@baidu.com). If the required level is very high, it may be “appropriately charged based on number of requests.” For those concerned about the privacy implications of a free service, Baidu states it “currently will not use/share any of your content (such as images, audio and labels) for any purpose except to provide you with the Cloud API service.” See the company’s EZDL FAQ for more information.
