Predictive representations: building blocks of intelligence
Abstract:
Recent progress in AI has shown that intelligent agents can learn about the world via relatively simple prediction algorithms. But the predictions that modern AI systems currently learn typically focus on predicting things within the present or in the near future. Can neuroscience provide insight into potentially better prediction algorithms that more broadly support intelligence?
Neuroscience evidence suggests that brains learn predictive representations that summarize aspects of the present with the futures that they lead to—that we represent roads by the destinations they afford, rooms by the objects we will use within them, or restaurants by the meals we will experience. The successor representation describes a promising family of theoretical models for instantiating these kinds of predictive representations. I will review some key neuroscience findings and detail how the successor representation can enable building blocks of intelligence that support exploration, knowledge transfer, and efficient planning.