# Ilker Yildirim (Yale) Source: https://www.youtube.com/watch?v=a7RJHfyZD0k Recap page: https://rapidrecap.app/video/a7RJHfyZD0k Generated: 2026-08-05T01:23:37.1+00:00 --- ## The Gist Ilker Yildirim from Yale University presents two algorithmic motifs of perception in the mind and brain, focusing on task-performant statistical heuristics and structure-preserving representations to explain how the brain efficiently achieves visual cognition. ## Quick Overview Perception is transformed into cognition through efficient neural algorithms that handle the complex computational challenges of the mind and brain. Ilker Yildirim discusses how task-performant statistical heuristics and structure-preserving representations provide the foundation for visual cognition. Through empirical studies and computational models, the lecture demonstrates how neural networks and dynamical systems account for human-level efficiency and visual processing. **Key Points:** - Ilker Yildirim is an Assistant Professor of Psychology at Yale University with appointments in the Foundations of Data Science Institute and the Wu Tsai Institute. - The lecture explores two primary algorithmic motifs of perception in the mind and brain, connecting raw sensory inputs to higher-level cognition. - Task-performant statistical heuristics utilize matched filters and statistical patterns to explain seemingly sophisticated behaviors in biological systems. - Structure-preserving representations map external physical and geometric structures into internal neural states to support general-purpose reasoning and planning. - Dynamical structure-preserving manifolds combine cognitive science representations with neuroscience dynamical systems to model physical scene prediction. - Single-cell recordings in the macaque dorsomedial frontal cortex reveal single-state sufficiency mechanisms that enable rapid trajectory predictions. - Multigranular optimization models explain phenomena like inattentional blindness by balancing task-relevant computations against computational costs on the fly. - Multilevel computational theories cut across levels of analysis to provide more interpretable and understandable models that predict neural data. ![Screenshot at 07:20: The presentation outlines the two main algorithmic motifs of visual cognition involving physical scene prediction and goal-conditioned abstractions.](https://ss.rapidrecap.app/screens/a7RJHfyZD0k/00-07-20.jpg) **Context:** This academic talk was delivered by Ilker Yildirim at Yale University on April 30, 2026, exploring computational cognitive science and the intersection of artificial intelligence, neuroscience, and psychology. ## Detailed Analysis Ilker Yildirim details the computational mechanisms underlying visual cognition by examining how brains transition from raw sensory inputs to flexible thought. He introduces two main algorithmic motifs: task-performant statistical heuristics and structure-preserving representations. By analyzing physical scene predictions in the macaque dorsomedial frontal cortex and evaluating multigranular optimization models, Yildirim demonstrates how neural circuits balance efficiency and accuracy. The talk bridges symbolic, connectionist, and dynamical system approaches to offer a comprehensive framework for understanding perception. ### From Perception to Cognition Visual perception transforms raw sensory inputs into flexible, general-purpose cognitive representations. - Human intelligence achieves incredible flexibility by turning physical sensory inputs into objects, scenes, dynamics, and goal-driven attentional processes. - Visual cognition can sometimes produce quirks, such as the invisible gorilla effect, where observers miss salient stimuli while attending to a specific goal. - Reverse engineering these perceptual processes requires looking at the entire scope from object recognition and scene composition to planning and reasoning. ![Screenshot at 02:40: A diagram mapping the full scope of visual cognition from objects and scenes to dynamics, planning, and attention.](https://ss.rapidrecap.app/screens/a7RJHfyZD0k/00-02-40.jpg) ### Two Algorithmic Motifs of Perception The mind utilizes specific algorithmic motifs to process visual information efficiently. - The first motif relies on task-performant statistical heuristics and matched filters to exploit statistical regularities in the environment. - The second motif involves structure-preserving representations, which map planar geometry and physical structures into algebraic and vector representations. - These motifs form the foundation for bridging cognitive science models with neural circuitry in the brain. ![Screenshot at 05:19: A demonstration of a structure-preserving map transforming planar geometry into algebraic representations.](https://ss.rapidrecap.app/screens/a7RJHfyZD0k/00-05-19.jpg) ### A Neural Algorithm of Physical Scene Prediction Neural algorithms in the frontal cortex implement physics-based predictions for moving objects. - Single-cell recordings from the macaque dorsomedial frontal cortex provide data on how neural populations represent physical trajectories. - Dynamical structure-preserving manifolds embed physics-based representations of gameboards into reservoir computers without requiring traditional stochastic training. - These neural populations rapidly encode future trajectories and collision points within hundreds of milliseconds. ![Screenshot at 12:12: Equations and face space plots representing dynamical structure-preserving manifolds for physics-based gameboard representations.](https://ss.rapidrecap.app/screens/a7RJHfyZD0k/00-12-12.jpg) ### Goal-Conditioned Abstractions and Inattentional Blindness Multigranular optimization models explain how attention manages high cognitive loads efficiently. - Inattentional blindness occurs because perceptual systems optimize for task-relevant computations while discarding wasted processing. - The multigranular optimization model dynamically refines schemas on the fly to maximize goal-relevant information and minimize computational waste. - This framework successfully replicates human noticing rates in complex visual tracking tasks where alternative models fail. ![Screenshot at 28:03: A diagram illustrating the multigranular optimization model and its evaluation against alternative cognitive accounts.](https://ss.rapidrecap.app/screens/a7RJHfyZD0k/00-28-03.jpg)