DeepBlue Technology beats Microsoft and Tsinghua at the PAKDD 2019 AutoML Challenge

DeepBlue Technology, a leading AI company headquartered in Shanghai, took top prize in the automated machine learning (AutoML) competition at the cutting-edge PAKDD 2019, beating Microsoft Research Asia and Tsinghua University in the process. The Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) challenge, staged April 14-17 in Macau, brought together the world’s top teams to compete in designing the best AutoML for Lifelong Machine Learning algorithm. This year, PAKDD 2019 Challenge received over 130 entries and more than 550 submissions from over 46 countries.

This year the competition required participants to design autonomous algorithms that were capable of handling concept drift, a data input method that mimicked real-world data that are acquired gradually over time. The DeepBlue team posted a record-breaking result, scoring a score of 1.2 over the second-placed team at 2.2. This difference is significant because the champion team proposed a more comprehensive solution compared to the rest of the participants.

Currently a hot topic in the AI circles, AutoML has sparked interest from a  larger community since Google launched their Cloud AutoML in January 2018. However, there is still much research that needs to be done to apply the technology to today’s products. 

AutoML overview  

Yitaek Hwang, Director of R&D at Leverege, summarizes AutoML in this way: This field of AI “aims to provide machine learning at the click of a button, or, at the very least, promise to keep algorithm implementation, data pipelines, and code, in general, hidden from view.” AutoML does that by applying automation to certain steps in the data modeling process, thereby reducing the amount of work done by data scientists through simplifying the creation process. 

This simplification provides three main benefits: 

• It lowers the barriers to entry to creating and testing machine learning models
• It will accelerate the machine learning process in current data science workflows and,
• It allows smaller datasets to perform better by implementing and automating already proven data processing steps learnt from larger datasets. 

Most AutoML algorithms are constructed based on various characteristics and features, however the DeepBlue team took a different approach. A spokesperson from the DeepBlue team said, “We managed to extend AutoML to a variety of data types, allowing pre-processing, engineering and combinations for different types of features. This allows AutoML to be applied to more real-world issues without the intervention of experts. Combined with a fast feature selection method, we managed to extract even higher-order combinations, effectively improving our model performance by such a large margin.” 

AutoML applications

When machines can learn at the push of a button, what are the implications of such an ease-of-use when dealing with AI? Sundar Pichai, the CEO of Google shared his ideas when he pitched Google’s Cloud AutoML functions, saying people could in the near future have ‘Machine Learning-as-a-Service’ (MLaaS) type of products in areas such as autonomous driving, healthcare, robotics and many more. 

In the case of robotics and AI products, lifelong machine learning methods such as AutoML will allow AI products to be truly ‘self-learning’ as they learn to interpret and utilize new data to improve the adaptive ability of the product.  This will add a dimension to next-generation robotics as both hardware and software evolve to be capable of self-actualizing intelligence. 

Hurdles to overcome

However, AutoML is still in its nascent stages and currently has limited capabilities, such as identifying and classifying objects in pictures, and sifting through datasets. Through competitions such as PAKDD and the KDD Cup, as well as research from for-profit companies such as Google and DeepBlue Technology, we can expect more advances in both the research and application aspects of AutoML. 

AutoML promises new possibilities for AI and machine learning in the near future. Not everything in AI is about doing things bigger, but rather as Jeremy Howard from fast.ai succinctly explains: “The truth though is that … the genuine advances consistently come from doings things differently, not doing things bigger.” The victory from the DeepBlue team shows that innovations such as AutoML at the forefront of technological research is only just the first step in changing the future. 

DeepBlue Technology is a Shanghai-based company providing autonomous solutions for traditional retailers using artificial technology (AI).