Trends and Challenges in Artificial Intelligence: Data Security, Deployment Bottlenecks, and Transfer Learning

Yang Qiang, a lecture professor in the Computer Science Department of Hong Kong University of Science and Technology, specializes in machine learning and transfer learning, and has held leadership roles at Fourth Paradigm, WeBank AI, and Huawei Noah’s Ark Lab.

In 2018, artificial intelligence (AI) remained a hot public topic. After two earlier valleys caused by lack of algorithms, computing power, and data, AI has entered a third golden period driven by the big‑data environment, becoming a favorite of industry and the public.

1. Achievements and Shortcomings of AI in China

China has made notable progress in face recognition, speech recognition, and automatic image analysis, reaching or approaching world‑leading levels. The number of AI research papers submitted and accepted has grown significantly. However, major gaps remain: most research follows existing trends rather than pioneering new directions, and originality is limited. The industrial‑Internet applications of AI lag behind global standards, especially in sectors such as healthcare and education where data scarcity hampers progress.

2. Data Security and Deployment Bottlenecks

Public awareness of AI grew alongside media coverage, but concerns about data breaches have intensified. Regulations such as the EU General Data Protection Regulation (GDPR), China’s Cybersecurity Law and Civil Code, and similar initiatives worldwide impose strict requirements on data handling. Traditional AI pipelines that collect, process, and use data may violate these laws. Federated learning, which exchanges encrypted model parameters without moving raw data, is presented as a promising solution to protect privacy while enabling collaborative model training.

Deployment bottlenecks also stem from data islands: many industries store data in isolated silos, providing only “small data” that limits the effectiveness of models designed for large‑scale datasets. Moreover, AI is a solution rather than a ready‑to‑use product; successful deployment requires high‑quality training data, continuous feedback, and integration across multiple data owners. Building vertical ecosystems or open platforms for AI and big data is essential for practical adoption.

3. Transfer Learning as a Key Direction

Transfer learning, a research area with more than 20 years of history, aims to enable AI systems to apply knowledge from one domain to another with limited data. In 2018, it attracted significant industry attention. Notable works include Google’s large‑scale language model that can be fine‑tuned on small downstream tasks, and Facebook’s image‑recognition transfer experiments. Companies such as DeepMind and Google’s AutoML are investing heavily in this paradigm, viewing it as “learning how to learn.”

Fourth Paradigm leverages transfer learning for financial services: a model trained on abundant small‑loan data is adapted via transfer learning to the scarce large‑loan scenario, enabling effective user profiling with limited data.

4. Misconceptions About AI

The author refutes two prevalent myths. First, the notion of an “AI Cold War” between China and the United States is misleading; cooperation is fundamental to AI progress, and China still trails the U.S. in many aspects. Second, the fear that AI will cause massive unemployment or surpass human intelligence is exaggerated. AI will augment specific workflows (e.g., logistics) but will not replace entire professions; technological advances improve efficiency rather than eliminate human roles.

5. Outlook for 2019

Looking ahead, AI is expected to accelerate further, with continued research on data security, practical deployment, and transfer learning shaping the field. The ultimate goal is to use AI to improve society, enhance convenience, promote equity, and increase work efficiency.