Nice to have AI tech when you have 400 million designs to be done in 10 days. Alibaba has developed a new AI able to learn from human designers how to produce its own original designs. Dubbed “Lu Ban” (named after a legendary Chinese craftsman), this innovative new program is able to design up to 8,000 high quality and fully unique banner adverts every second. Not only that but, according to Le Sheng, the Chief Technician in charge of Lu Ban’s development, the program has evolved to the point where it is producing a design style beyond what it has been taught, saying that “Lu Ban spent several weeks studying our design style and has now begun to produce designs that humans have never taught it”. Lu Ban was created as a direct response to the growing needs of China’s biggest e-commerce platform (Taobao.com). Every year, Alibaba designers are under immense pressure to produce an endless supply of banner adverts for the millions of products available on the website, while remaining cost effective. The sheer gigantic number of banners recurred is partly due to the fact that every user sees different banners designed specially to match his/her preference. This is especially true on November 11th, otherwise known as Single’s Day, China’s biggest online shopping day of the year, with billions of dollars’ worth of transactions occurring in just 24 hours. Although initially debuted in 2016, Lu Ban’s AI was significantly improved upon for Single’s Day 2017, managing to design an incredible 400 million banners during the sales period at a rate of 40 million a day. This was up from a total of 170 million the previous year. Every banner was designed differently according to individual user’s preference. How does AI designer work? The secret to the AI’s success comes from a three-stage learning development program that saw Lu Ban studying the work of human designers and reviewing millions of design drafts and data. Three-stage learning development program Ultimately, Lu Ban’s output is from the pairing of suitable design structures with appropriate design elements to produce a full banner image. In order to produce the best design structures, Lu Ban first had to develop a Style Learning Module. Through a neural network learning process, Lu Ban studied the composition of different design elements: backgrounds, and masking as well design techniques and style. In doing so, Lu Ban was able to develop a basic understanding of how banners should be constructed and to recall and apply the different procedures required for a variety of complex design processes. Following this initial stage, a design framework was developed. This forms a series of basic design knowledge and rules from which Lu Ban is able to generate a set of designs. It is the equivalent of a designer having a general concept or vision of a design in their mind, before they put pen to paper. At the same time as Lu Ban’s Style Learning Module was being developed, the team also began building the Element Center — a vast library of individual design elements, such as images and backgrounds, from which Lu Ban can select appropriate components to apply to a specific and basic structure. Within the library, each design element is classified according to various visual characteristics and types, such as the subject, color, and background. With both the Style Learning Module and the Element Center working, it is the role of the Actuator to finally start producing designs. The main function of the Actuator is to select the best design structure generated from the Style Learning Module according to the requirements of the design request, and then select the most appropriate elements from the Element Center. This allows Lu Ban to plan a plurality of optimal generation paths and deliver the banner design. When this process is complete, Lu Ban produces multiple designs which are delivered to the Evaluation Network. The purpose of the Evaluation Network is to input a large amount of design pictures and scoring data, and after training, let the machine learn to judge the design quality. AI designer creates suitable flower backgrounds for clothes on display. Secrets behind Lu Ban’s success Lu Ban has two teachers training the AI every day to improve the output. The first designer is responsible for helping Lu Ban to continually learn new design templates and methods so that Lu Ban can continue to evolve. The second teacher evaluates the results of Lu Ban’s design, providing the AI with feedback on which designs are better. In this way, Lu Ban can incorporate this feedback on to future decisions when selecting the appropriate templates and design elements. The designer’s core responsibility is to turn design into data. At present, Lu Ban has learned a million design drafts and evolved hundreds of millions of poster design capabilities. To do this, Alibaba’s designers and algorithm engineers first had to study how visual design specialists abstract design into a data model of style, technique, templates, and elements. This allows them to turn years of visual design experience into machine-learnable data. After the data model is defined, the data is captured and labeled, and the data set is classified and managed. This raises several challenges such as what data should be used to verify this model and how can we assess the effectiveness of the model? These data problems require a clear data link design. Lastly, an algorithm framework needs to be developed through discussions between product designers and algorithm scientists to turn business scenarios and data problems into algorithms. Of course, the development of Lu Ban has not been without its challenges. Not least of which is the fact that this simply has not been done before and so there is no real precedent to follow and there are few off-the-shelf technologies or frameworks available for reference. Additionally, there is a lack of annotation data. All of today’s AI is based on large-scale structured annotation data, however in this area there is currently not a lot of data and that which does exist is far from being standardized and structured. Lastly, the very aim of the project, to create an AI with an aesthetic sense, is wildly ambitious. Design is a very uncertain and subjective thing. To evaluate the beauty or visual appeal of something is an extremely human trait. Design is an art and therefore traditionally at odds with the science of AI. The progress taken in this project nonetheless hint that this divide may yet be bridged. Although Lu Ban is far from being the next Da Vinci or Van Gogh, this is definitely a very interesting step in the progress of AI. (Original article by Wu Chunsong 吴春松) Alibaba Tech First hand, detailed, and in-depth information about Alibaba’s latest technology → Facebook: . Twitter: . “Alibaba Tech” “AlibabaTech”

Facebook

Plan A

Twitter

City Brain: Cloud Computing and Data Technology Help Our City Think Deeply

400M Works: AI Designer’s Shopping Extravaganza

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

A Brief History of Flink: Tracing the Big Data Engine’s Open-source Development

29/06/2018: Biggest Stories in the Cryptosphere

27/04/2018: Biggest Stories in the Cryptosphere

2 Methods for Installing Serverless Kubernetes in Alibaba Cloud

A Brief Overview of Alibaba

Alibaba’s “Homecoming” IPO in Hong Kong

A Brief History of Flink: Tracing the Big Data Engine’s Open-source Development

29/06/2018: Biggest Stories in the Cryptosphere

27/04/2018: Biggest Stories in the Cryptosphere

2 Methods for Installing Serverless Kubernetes in Alibaba Cloud

A Brief Overview of Alibaba

Alibaba’s “Homecoming” IPO in Hong Kong

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps