What Kind of Computation Is Cognition?

The speaker warmly welcomes Professor Josh Tenenbaum, a computational cognitive scientist at MIT, to speak about the nature of cognition. His achievements and recognition by the MacArthur Foundation are highlighted, and the anticipation for his talk on the computation of cognition is expressed.

Josh Tenenbaum introduces the central question of his talk: what kind of computation is cognition? He discusses the importance of understanding cognition in terms of computation and the challenges of finding meaningful answers. The talk aims to raise foundational questions about modeling the mind and the relationship between cognitive science and metaphysics.

Tenenbaum discusses the current state of artificial intelligence, noting its increasing utility but lack of true general intelligence. He uses the example of self-driving cars to illustrate the challenges and investments in AI, emphasizing the need for fundamental advances in AI to bridge the gap between current technologies and human-like intelligence.

The speaker delves into the specifics of autonomous driving, highlighting the progress made and the underlying AI technologies like deep learning and neural networks. Despite advancements, he points out the limitations and the 'long tail' of problems that remain unsolved, indicating that AI still has a long way to go to achieve true autonomy.

Tenenbaum explores the computational aspects of intelligence, focusing on the 'common sense core' of human cognition. He discusses the intuitive understanding of physics and psychology that humans possess from a young age and the challenge of capturing these aspects computationally.

The speaker uses examples of infants engaging with objects to illustrate intuitive physics and psychology. He discusses how young children demonstrate an understanding of physical objects and goal-driven agents through their actions and reactions, highlighting the innate human capacity for modeling the world.

Tenenbaum presents the famous Heider and Simmel animation as a key example of how humans naturally interpret interactions between agents, even from simple shapes. He discusses the rich understanding of physical dynamics, goals, and emotions that humans exhibit when viewing such animations, emphasizing the computational challenge of replicating this understanding in AI.

The speaker introduces the concept of probabilistic programming as a way to integrate neural networks, symbolic languages, and probabilistic inference to model common sense. He discusses the potential of this computational paradigm to advance AI and our understanding of cognition, including the role of game engines in simulating physical and social interactions.

Tenenbaum discusses research on intuitive physics and how people judge causal responsibility. He describes experiments that use eye-tracking to understand how people's gaze patterns change based on whether they are making predictions or causal judgments, revealing the computational processes underlying human intuition about physical interactions.

The speaker talks about recent work on using probabilistic intuitive physics for problem-solving, particularly in creative tool use. He describes a virtual tools game that demonstrates how people use trial and error to solve problems and how a probabilistic simulation-based model can capture this process, including learning dynamics.

Tenenbaum presents a model called 'Flatland' that simulates multi-agent interactions with physical objects, capturing complex behaviors like collaboration and competition. He emphasizes the potential of these models to understand and predict social interactions and causal relationships, highlighting their success in mimicking human-like interactions.

In the final paragraph, Tenenbaum touches on the relationship between these computational models and the brain, and the potential for understanding how humans learn. He suggests that future learning algorithms may need to be 'program learning programs,' capable of writing their own algorithms, much like how humans construct knowledge.

Tenenbaum concludes his talk by reflecting on his educational experience at Yale and the future of cognitive science. He discusses the potential of algorithms that write algorithms, and the importance of understanding the brain's computational processes. He ends with a call to explore the synthesis of symbolic representations, probabilistic inference, and learning in understanding cognition.