Murray Shanahan, veteran AI scientist at Imperial College London and principal scientist at DeepMind, discusses the importance of embodied interaction in AI progress, potential breakthroughs in the field, and how salary inflation for commercial AI engineers may hinder research.
Embodied interaction is crucial for the progress of AI and achieving artificial general intelligence (AGI).
Transformers, with their ability to capture contextual information and adapt attention, have shown exceptional performance in various domains of AI.
Deep dives
Progress in AI and the Challenges of Ethical AI Implementation
Over the past decade, artificial intelligence (AI) has made significant progress, with deep learning and neural networks driving advancements in various fields. However, a major challenge has been ensuring the ethical and fair behavior of AI systems. Common sense, a fundamental aspect of AI, remains one of the biggest obstacles in achieving artificial general intelligence (AGI). While commercial applications have seen success in areas like face recognition and image classification, common sense understanding is still lacking. The integration of deep learning and reinforcement learning has revolutionized AI, particularly in areas like game playing and language translation. However, the progress towards AGI and the development of AI that exhibits human-like common sense require further breakthroughs and a focus on embodiment and interaction with the physical world. Exploration and research in new directions outside the deep learning paradigm are essential for achieving these goals.
Transformers and the Unreasonable Effectiveness of Data
Transformers, a type of artificial neural network, have emerged as a powerful tool for learning and generalizing from data. Unlike traditional neural networks, transformers prioritize learning from scratch rather than relying on hand-coded rules. They have shown exceptional performance in various domains, including language translation, image and video processing, and protein folding prediction. Transformers excel at capturing contextual information and can adapt their attention to different parts of the input data, making them highly effective at a range of sequence-based tasks. The success of transformers demonstrates the importance of large-scale datasets and increased compute capacity in achieving breakthroughs in artificial intelligence.
Reinforcement Learning and its Application
Reinforcement learning, a trial-and-error learning approach, has proven instrumental in AI achievements. DeepMind's DQN system, combining deep neural networks with reinforcement learning, showcased impressive game-playing abilities on classic Atari games. Reinforcement learning has since expanded into various industrial applications, including autonomous driving. However, challenges remain, such as ensuring machines possess common sense and bridging the gap between human-level general intelligence and commercial applications. Further progress is needed in understanding how to balance exploitation and exploration in AI systems and incorporating expertise from fields like animal cognition and developmental psychology.
The Path to Artificial General Intelligence
While AI has made remarkable progress over the past decade, achieving artificial general intelligence is still a substantial challenge. Common sense understanding and embodiment are crucial factors in advancing AI towards human-level capabilities. The embodiment of AI agents through virtual simulation presents opportunities for learning foundational common sense. Current developments in AI, while impressive, do not yet warrant descriptions of consciousness or human-like general intelligence. Continued exploration, interdisciplinary collaboration, and a focus on tackling fundamental challenges, such as reasoning, generalization, and the understanding of everyday physical objects, are necessary for progress towards artificial general intelligence.
Artificial Intelligence (AI) is on every business leader’s agenda. How do you ensure the AI systems you deploy are harmless and trustworthy? This month, Azeem picks some of his favorite conversations with leading AI safety experts to help you break through the noise.
Today’s pick is Azeem’s 2021 conversation with veteran AI scientist Murray Shanahan, professor of cognitive robotics at Imperial College London and principal scientist at DeepMind.
They discuss:
Why some aspects of AI progress depend on embodied interaction.
Understanding from where the major breakthroughs in the field may come.
Why the salary inflation for commercial AI engineers might hinder research.
