Mar 22, 2022|
A Major Achievement: Super Deep Learning Model by JD Explore Academy
by Doris Liu
A super deep learning model, ViTAEv2, which was jointly proposed by JD Explore Academy and the University of Sydney, Australia, was recently ranked top in accuracy of image classification on ImageNet ReaL without using extra training data.
With larger scale, better results and better adaptability to various vision tasks, the model with 600 million parameters set a world record of 91.2 percent accuracy in February in the field of image classification technology by not using extra training data.
The ViTAEv2 model adopts a “pre-training and fine-tuning” paradigm to make a breakthrough in model architecture and training paradigm, which fully uses the effectiveness of inductive bias in large-scale models, as well as pre-training and migration learning algorithms that fit the structure of the model to achieve the target results.
JD Explore Academy also developed the few-shot learning capability of the large-scale ViTAEv2 model by fine-tuning the large-scale ViTAEv2 model with 1, 10 and 100 percent of the data respectively, and the result showed that when fine-tuned with only a small amount of data, namely 10 percent, the large-scale model significantly outperformed the smaller scale model which was using all the data. “It further confirmed that the large-scale model has a strong ability in few-shot learning, which indicates that the super deep model has strong representation ability, learning ability and sample efficiency,” said a deep learning scientist from JD Explore Academy.
The ViTAEv2 model’s ability to help solve challenging tasks with low or even zero resources was fully validated, as well as its ability to reduce data annotation costs, accelerate algorithm development cycles, simplify model deployment, empower and facilitate the R&D and implementation of next-generation automated learning technologies.
As the largest public dataset for image classification, the ImageNet dataset has attracted the attention and participation of top international technology companies such as Google, Microsoft and Facebook, as well as leading research universities such as Stanford University, Massachusetts Institute of Technology and National University of Singapore especially in terms of the accuracy ranking list. Its data metrics are widely recognized as an important criterion to measure the level of computer vision technology.
The performance of the ViTAEv2 model helped JD Explore Academy take its computer vision models to a new level, and is expected to continue to push forward the development of a range of vision tasks, such as semantic segmentation, object detection, pose estimation, and video object segmentation.
Apr 22, 2024|
JD.com Debuts AI Digital Representative of Founder Richard Liu During Livestream, Drawing 20 Million Views in Under One Hour
During a livestream on April 16th, JD.com introduced an AI digital representative of its founder, Richard Liu (Qiangdong Liu), nicknamed “procurement and sales manager Brother Dong (采销东哥).” Liu appeared in both JD Home Appliances and JD Supermarket’s livestreaming rooms, attracting over 20 million views within the first hour and generating RMB 50 million in sales throughout the entire livestream.
The digital representative almost perfectly replicates Liu’s facial expressions, body language, gestures, and voice, capturing even the subtlest movements of his fingers. Behind this technology is JD Cloud’s ChatRhino, an advanced large language model (LLM) that brings the avatar to life.
Behind the Technology
The LLM precisely mirrors Liu’s distinctive expressions and movements, naturally weaving in his occasional finger twiddles while he speaks, accentuating points with hand gestures, and adding nods at just the right moments for emphasis, creating a natural and authentic representation. Perfecting the voice proved to be a greater challenge. Those who have heard Liu speak are well-acquainted with the distinct inflections of his Suqian accent—a characteristic touch from his small-town roots in Jiangsu province. Observers will note the brisk pace of his speech, lighter enunciation, and his tendency to meld words in a seamless flow. His “sh” sounds carry a nasal quality, and he often addresses people as “brothers” to boost morale.
While programming the standard Mandarin dialect is relatively straightforward, teaching the AI to adopt the nuances of the Suqian dialect and Liu’s unique vocal patterns required a precise level of customization from the ChatRhino LLM. For example, the LLM had to provide accurate judgments on when to introduce the nasal tonality or the specific moments to merge words, ensuring the AI’s linguistic output mirrored Liu’s signature speaking style authentically.
Beyond replicating vocal characteristics, the AI’s voice must also be suitable for the casual, dynamic atmosphere of e-commerce livestreaming, such as adopting colloquialisms like “xiong di men” — a phrase comparable to “folks” or “guys” in English — to create a friendly rapport with the audience. JD Cloud’s ChatRhino team fine-tuned the model to focus on sociable nuances. The initial dataset for the LLM consisted predominantly of Liu’s formal speeches—spirited and compelling, yet too ceremonious for the interactive world of livestreaming. To cultivate a more congenial and approachable persona, the team utilized Liu’s recent casual conversations as the primary source material, rich with personal anecdotes and travel tales. This approach shaped the AI’s voice into one that is that is not only natural and approachable but also genuinely engaging for the end user.
As a fully self-developed technology product of JD Cloud, the ChatRhino Digital Representative solution has already served over 4,000 brand merchants, contributing to a 30% increase in order conversion rate during idle hours and helping merchants reduce livestreaming costs, enhance operational efficiency, and optimize user experience. Equipped with 50,000 hours of voice data, this technology enables digital representatives to intelligently adapt to various livestreaming styles dynamically.
The AI is also adept at managing 70% of routine live chat inquiries autonomously without any manual intervention, while also optimizing its responses with a self-improving algorithm. For instance, digital representatives can autonomously fields questions using product details and scripts, and post-livestream, and assimilate new information. Mirroring the acumen of seasoned sales professionals, it boasts a 90% accuracy rate in offering tailored product recommendations when solicited by customers.
The recent introduction of a cutting-edge livestream control console serves as the “brain” for the digital representative solution. It diligently monitors real-time inventory, adjusting the showcased products in the livestream and tailoring the script to ensure relevance—sidestepping items that are out of stock and spotlighting those in high demand. It also gauges and fine-tunes the frequency and style of interactions based on live audience engagement. With this advanced platform, businesses managing multiple livestreams can effortlessly operate thousands of sessions concurrently via cloud services, bypassing the need for intricate local setups. This innovation yields a substantial cost reduction of 30%.
Human-Machine Synergy
The intention behind digital avatars is not to outshine or compete with human talent but to harness the potential of technology to work in concert with livestreamers, maximizing overall growth. Essentially, the digital representative solution serves as a tool for brands and merchants to leverage long-tail traffic once real livestreamers conclude their broadcasts. JD’s analytics indicate this strategy can boost order conversion rates by 30% during off-peak hours, such as late into the night.
Digital representatives also excel in specific product categories that require detailed explanations, such as health products like nursing beds, electric wheelchairs, and oxygen concentrators. The solution can provide constant, round-the-clock guidance, adding substantial value to livestreams when traffic is typically lower.
Last week, JD.com announced an investment of one billion yuan into video content creation. This investment underscores JD’s commitment to leveraging the power of short videos and livestreams to transform the online shopping experience for its users and create new growth avenues for brand merchants.
Apr 20, 2024|
JD.com Unveils Ambitious Plans for 618 Grand Promotion: Aiming to More Than Double High-Performing Merchants
JD.com is gearing up for this year’s 618 Grand Promotion with an ambitious goal: to more than double the number of merchants achieving sales over one million yuan on its platform. At the recent Conference of Merchants and Ecosystem Partners on April 18, JD.com introduced a range of supportive initiatives aimed at achieving this ambitious target.
Sandy Xu, CEO of JD.com, noted at the conference that the company will dedicate significant resources to bolster traffic support, incorporate advanced AI technology, and enhance service capabilities. These efforts are particularly focused on empowering third-party merchants, including small and medium-sized enterprises in its marketplace.
Key initiatives include optimizing the platform’s traffic distribution ecosystem to prioritize highly-rated stores and products and competitively priced products through enhanced search and recommendation mechanisms. Additionally, JD.com will continue to offer merchants free access to high-traffic channels, which include significant incentives such as the popular 10-billion-yuan discount and free shipping programs.
To engage more content creators, JD.com recently announced that it will allocate over one billion yuan in cash incentives and traffic resources. This investment aims to cultivate a vibrant ecosystem that supports merchants’ traffic, business growth, and innovation.
At the core of JD.com’s strategy is technological innovation, which involves integrating AI-driven solutions into merchant operations. This includes the deployment of JD’s proprietary large language model, ChatRhino, and an AI-empowered merchant operation platform that offers are variety of free AI applications, and more. These technologies streamline various operations scenarios, enhance efficiency, and reduce costs for merchants.
For instance, On April 16th, JD.com unveiled an AI digital avatar of its founder, Richard Liu, as a sales host on its livestreaming platform. This event attracted 20 million viewers in just one hour and generated sales exceeding RMB 50 million ($6.9 million). This AI-driven virtual host technology, developed by JD Technology, is now supporting over 4,000 brands on the platform. Additionally, merchants such as a sashimi knife seller can utilize JD.com’s AIGC content generation tool to quickly produce high-quality demonstration images, bypassing traditional costs associated with product photography.
JD.com also continues to enhance its service capabilities, particularly for small and medium-sized merchants. This includes personalized business guidance, comprehensive training programs, and upgraded after-sales services, ensuring a seamless shopping experience for customers.
As of March 2024, JD.com has seen a significant increase in active third-party merchants on its platform, surpassing one million. This growth is a testament to the platform’s effectiveness and its appeal to both new and established merchants.
A Volunteer–Turned-Entrepreneur Pledges to Make a Difference with JD Pet
