Titanium Media reported on April 22, recently, the research result of the joint team of the fully automatic AI mid-stage provider "Deep Endowment" and Xiamen University Professor Ji Rongrong "Evolving Fully Automated Machine Learning via Life-Long Knowledge Anchors" "(Automatic Machine Learning Based on Knowledge Anchor Evolution) was accepted by IEEE TPAMI.
This paper innovatively proposes a new type of fully automatic machine learning framework, which for the first time breaks the independent design of each search space in the existing automatic machine learning, and uses dataset knowledge anchors and evolutionary algorithms to speed up the search and solve the problem. It solves the design problem of searching for an optimal solution in a very large space.
It is reported that "Deep Endowment Intelligence" was established in 2019 and is dedicated to the algorithm research of MetaAI technology (making AI with AI) and the implementation of business scenarios. The founder and CEO of the company, Wu Chenglin, graduated from the Department of Computer Science of Xiamen University. Before founding DeepFuzhi, Wu Chenglin was the chief scientist of Stupid Bird Social (B+ round) and the head of AI Lab, and won the honor of Forbes 30 under 30 in 2018. . Since the establishment of
one year ago, DeepFuzhi has completed the development of multiple core modules in the SaaS platform, providing fully automatic AI middle-end SaaS services for B-end customers in retail/shoes and apparel, industrial manufacturing, trading platforms, finance and other industries. In reducing the labor cost of AI application landing. In October last year, DeepFuzhi announced the completion of an angel round of financing of tens of millions of yuan invested by Meihua Venture Capital.
In short, deep learning allows machines to learn from a large amount of data and apply them. It has achieved amazing results in many tasks such as image classification and sequence labeling. However, this process requires a lot of manual intervention, such as feature extraction, model selection, parameter tuning, etc., which is time-consuming and labor-intensive.
And the author of this paper innovatively proposes a new type of fully automatic machine learning framework, which introduces an automatic model to let the machine "learn how to learn" by itself, that is, AI replicates AI.
According to the description of the paper, the knowledge anchors in this framework use a new meta-feature and probability sampling method, which greatly reduces labor and alleviates the overfitting of to in the search process. The framework realizes the automation of the whole process, which greatly reduces the threshold for machine learning applications. Users only need to perform "fool-like" operations according to the prompts, which greatly reduces the user's learning cost. At the same time, it can be found from the experimental results that the effect of the framework on typical datasets of five modalities of image/audio/video/text/table is far superior to the current optimal result (SOTA) scheme.
Deep Wisdom said that the company has now applied this achievement to its self-developed Tianji automatic machine learning platform, and has provided multiple head customers in retail/shoes and apparel, industry, trading platforms, finance and other industries. corresponding service. customer data shows that the entire set of solutions of this platform can greatly improve the effect of tasks such as decision-making, recommendation, and search, and improve business indicators by 40% to 60%, achieving more than 60% labor cost savings and significant net profit for customers. Interest rates rise.
In April 2020, the company used a new fully automatic machine learning framework to win the world championship of the NeurIPS-AutoDL series competition finals, the top international competition in the field of automatic machine learning. It has stably obtained eight firsts and equal firsts on each dataset, proving the universality of this framework in different scenarios.
In terms of talents, the in-depth intelligence team comes from 985/211 colleges and universities, as well as first-class talents from colleges and universities including Stanford, CMU, Oxford, IC, UCL, Tsinghua University, Peking University, Xiamen University, etc. Senior researcher and architect of Baidu, Huawei and other companies, with AI landing experience of billions of users and hundreds of billions of data.
Wu Chenglin said that in the next period of time, DeepFuzhi will give full play to its rich algorithm resources, in-depth technical cooperation support, constantly improving ecological resources and other advantages, continue to improve MetaAI technology, continue to build a fully automatic machine learning platform, and More developers and service providers will jointly launch more AI solutions for end customers.
(Titanium Media App first published in this article, author | Zhijia Lin)
