Relation networks 6 devised by DeepMind have achieved super-human performance on a relational reasoning dataset. The language model RoBERTa 4 has also outperformed human on several natural language understanding tasks of the GLUE benchmark 5. For example, deep residual networks (ResNets) 3 have already surpassed human performance on image classification.

In recent years, deep learning 2 has achieved tremendous successes in various AI research areas such as computer vision (CV) and natural language processing (NLP). As we can imagine, devising and obtaining an AGI system would not only accelerate the AI research itself, but also benefit a wide range of AI+ fields including neuroscience, healthcare, and biomedicine.

Despite not being precisely defined, AGI is broadly agreed to have several key features 1 including: (1) matching or exceeding human performance across a broad class of cognitive tasks (e.g., perception, reading comprehension, and reasoning) in a variety of contexts and environments (2) possessing the ability to handle problems quite different from those anticipated by its creators and (3) being able to generalize/transfer the learned knowledge from one context to others. Since this form of human-level artificial intelligence (AI) is too far from reality hitherto, researchers change to set a less ambitious goal of achieving artificial general intelligence (AGI). Science fictions and sci-fi films, that describe highly intelligent computer minds, robots, or even human-shaped ones, can be said to understand or have primitive cognitive abilities analogous to human intelligence.

Abridge example sentence