A recent (Dec. 5th) paper  of Google DeepMind is discussed. The DeepMind team reported a new algorithm termed AlphaZero which represents a seminal breakthrough in AI research. AlphaZero is able to learn how to play a large class of board games just via self-play. Only the rules of the game need to be implemented by hand. The algorithm was applied to three distinct games, namely Chess, Go, and Shogi. For all three games, super-human performance was reached after training for a couple of hours (!). This represents a giant leap towards artificial general intelligence.
The presentation will be introductory. We introduce deep neural networks, review previous achievements in AI research, and discuss reinforcement learning (RL), a technique on which AlphaZero is based on. Two specific RL algorithms are discussed, DeepMind's DQN and AlphaZero, and their impressive performances are appreciated. We briefly look into specific chess games played by AlphaZero which highlights how AI can help to get new insights into complex systems. Closing, we speculate about the future of AI and AI safety.
 D. Silver et al. "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm." arXiv:1712.01815 (2017)
SCB Programme seminar
Data i hora d'inici: 07/03/2018, 9.30h
Data i hora de fi: 07/03/2018, 10.30h
Organitzador: IRB - Institut de Recerca Biomèdica
Lloc: Edifici Clúster Laboratoris, Aula Félix Serratosa
Host: Adam Hospital - IRB Barcelona