Intel's Bob Rogers explains the possibilities that emerge as AI progresses beyond standard machine learning. DeepMind's self-taught Go champion is just the beginning.
DeepMind, the division of the Alphabet conglomerate that is devoted to artificial intelligence, recently announced that its Go-playing AI, called Alpha Go, had evolved into a new iteration it calls AlphaGo Zero. The reason for the zero is that the new version is capable of teaching itself how to win the game from scratch.
“Zero is even more powerful and is arguably the strongest Go player in history,” according to the DeepMind announcement. The AI not only can beat the best human players but can even defeat “the previously published champion-defeating version of AlphaGo by 100 games to 0.”
The key difference between this version and earlier AlphaGo iterations is that Zero’s predecessors were taught to play Go, having been “trained on thousands of human” experiences in the game. In contrast, AlphaGo Zero “learns to play simply by playing games against itself, starting from completely random play,” DeepMind states.
By doing that through a million games, the AI consistently improved. A gif that graphs AlphaGo Zero’s learning progress shows its game improvement over time as it fine-tuned its ability to predict moves. This self-teaching is called reinforcement learning, and it could hold great potential for AI applications.
To gain some expert insight into what this means for future AI applications, I spoke with Bob Rogers, Intel’s chief data scientist. Rogers explained the possibilities that emerge as AI progresses beyond standard machine learning with reinforcement learning, as well as the limitations that are still in place.
Machine learning uses “data to create a model,” Rogers noted, but the data has to be selected and fed in by humans who determine the parameters and rules for classification. That requires people with complete domain expertise as well as programming ability — resources that are not on hand for every organization.
By shifting to deep learning, AI can learn from examples and infer rules for itself. Rogers described the process as the “part of machine learning in which neurons pass information from one layer to another” and in doing so “create complex and subtle interpretations of data.” Instead of having humans set up the algorithms ahead of time toward where they believe “the most important data is,” he said, deep learning can learn to extract its own rules from the examples it sees.
NEXT PAGE: The Next Iteration of AI