DeepMind is perhaps amongst the only few players in the space to have mastered the art of reinforcement learning – a computational technique that had received surprisingly little or no attention in the advancement of artificial intelligence – an idea of a learning system that learns through a process of trial and error.
So much so to an extent where it has successfully applied reinforcement learning in many domains, including systems like AlphaZero, which went on to master the games of chess, Go and Shogi. It even put some of the world champions to the ground.
“Even if I become the number one, there is an entity that cannot be defeated,” said Lee Se-dol, the South Korean Go champion. After this, he even went on to announce his retirement plans from professional play indefinitely. In 2016, Lee lost 4-1 (four matches to one) to DeepMind’s AlphaGo.
Alphabet-owned artificial intelligence lab DeepMind didn’t just stop at board games; it went on to tackle significant real-world problems like reducing energy use, medical research projects (prevention vision loss, radiotherapy, etc.), several use cases for Google products like Google Photos and Google Translate, and more.
DeepMind was acquired by Google parent Alphabet in 2014. It believes that AI systems underpinned by reinforcement learning could grow such that it could break the theoretical barrier to ‘artificial general intelligence,’ or AGI, without any new technological developments.
DeepMind’s AlphaFold, the latest proprietary algorithm, predicted the structure of proteins in a time-efficient way. Also, DeepMind MuZero matches the performance of AlphaZero on Go, chess and Shogi, along with mastering a range of visually complex Atari games. Here is a complete timeline of some of the key highlights of DeepMind in the field of reinforcement learning:
Besides reinforcement learning, DeepMind also looks at other fundamental areas like 'symbolic AI' and 'population-based training.'
DeepMind is Not Alone
From 2014 onwards, the trend around reinforcement learning seems to be growing exponentially. While reinforcement learning has been behind some of its most well-known research breakthroughs, DeepMind is not alone in space. Other players include Microsoft, Google, Facebook and Amazon.
A deeper look into the trends showed us that Microsoft is sitting on a gold mine of resources around reinforcement learning, followed by DeepMind, Facebook and Amazon. Surprisingly, Google had the least amount of resources, where the number was close to 35. The graph below illustrates this purely based on work and experiments published on its website around reinforcement learning.
Towards Artificial General Intelligence (AGI)
For more than a decade now, DeepMind has been championing reinforcement learning. From AlphaGo, AlphaGo Zero to MuZero and the recent AlphaFold and RGB-Stacking, it has created history with reinforcement learning.
DeepMind had recently proposed a hypothesis, where it had argued that ‘Reward is enough’ to reach ‘artificial general intelligence,’ or ‘human-level AI.’ According to DeepMind, the maximisation of total reward might be sufficient to understand intelligence and its associated abilities.
The company said that it has explored in greater depth several abilities that may at first glance seem hard to comprehend through reward maximisation alone – be it language, knowledge, learning, perception, social intelligence, generalisation, limitation, or general intelligence, and noted that reward maximisation could provide a basis for understanding each ability.
All in all, DeepMind believes that intelligence could emerge in practice from sufficiently powered reinforcement learning agents that learn to maximise future rewards. If this hypothesis holds true, it could provide a direct pathway towards understanding and constructing an AGI or human-level AI – a very close goal for DeepMind.