The next big step in AI could come from exploring the interface between cognitive science, neuroscience, AI, computer science, and philosophy of mind, as deep convolutional networks have limits in their understanding abilities. However, many machine learning researchers tend to ignore the way the human brain works.
This podcast episode discusses the importance of reinforcement learning in solving intelligence and how it can guide one to find the right direction for research. Theoretical proof such as AIXI, which is a universal intelligence reinforcement learning system in the limit, is also mentioned.
This podcast highlights the significance of placing equal focus on the structure of the environment in order to understand human intelligence and build human-like AI. This is exemplified by DeepMind's self-play mechanism which attributes to the agent's competitive behavior and its effective learning abilities.
Deep learning was initially underestimated but it has proven to be successful when the right mixture of data and computation is applied. The speaker discusses the differences between vision, language, and reinforcement learning.
In this transcript, the co-founder of DeepMind, Mustafa Suleyman, talks about his experience at MIT and the difficulty of getting industry support for AI before 2010. He also discusses DeepMind's founding tenets, including algorithmic advances like deep learning.
The idea of achieving AI by creating a society of different AIs that interact with each other being alive in all over the internet. This was an AI pioneer Marvin Minsky's thought, which was later attempted by a startup WebMind to make a distributed society of AIs living in different places.