Marr's Computational Theory of Mind in Color Vision and Face Perception | MIT OpenCourseWare

Overview

In this MIT OpenCourseWare lecture, Marr’s computational theory of mind is introduced and applied to color vision and face perception. The instructor emphasizes the inputs and outputs of mental processes, and highlights how ill posed problems necessitate additional information and assumptions to infer object properties from sensory data.

• The Marr framework centers on what is computed and why, not merely hardware.
• Color vision is treated as ill posed, requiring knowledge about illumination to infer reflectance.
• Face perception is examined through psychophysics and functional MRI to reveal how the brain encodes faces and processes social cues.

Overview

The lecture introduces David Marr’s computational theory of mind, arguing that understanding minds requires specifying the computations the brain performs, the representations it uses, and the hardware that implements them. The presenter uses color vision as a concrete example to show how the mind turns sensory input into useful percepts by solving a sequence of informational problems. The ultimate goal is a precise computational theory of mind that could, in principle, be implemented in software or hardware.

Levels of Analysis in Marr’s Framework

The talk emphasizes Marr’s three levels of analysis: the computational theory level (what is computed and why), the algorithmic/representation level (how the computation is carried out and what is represented), and the hardware level (how the brain implements these computations). The instructor stresses that understanding perception requires moving beyond anatomy to contemplate the problems being solved and the information available to the system.

Color Vision as a Case Study

Color vision is analyzed as a problem of determining a surface reflectance from light that arrives at the eye. The input is luminance L, a product of surface properties and illumination, while the desired output is reflectance R. This is an ill posed problem because multiple combinations of surface properties and illumination can produce the same retinal signal. To illustrate how the brain tackles such problems, the talk discusses the need to incorporate context, prior knowledge, and world statistics to constrain inferences about color.

The Ill posed Problem and Its Implications

The lecturer emphasizes that many perceptual inferences are made under ill posed conditions, requiring assumptions about lighting, lighting statistics, and object familiarity. The color case also introduces the concept of inverse optics, where the brain infers what caused a retinal image, a problem often underdetermined by available data alone.

Psychophysics as a Window into the Mind

Psychophysics is presented as a practical method to probe the computations and representations of the mind. An example involves color perception tests where observers are asked to identify the color of a car given different illumination patches. The demonstration shows that observers can deduce the illuminant color and adjust their perceptual interpretation accordingly, revealing the brain’s use of contextual information to solve ill posed problems.

From Perception to Brain Implementation

The talk briefly outlines how color vision is studied at multiple levels, including hardware (neurons and cortical areas) and behavior. It notes that while there is extensive research on the brain basis of color vision, a full understanding requires integrating behavioral data, neural measurements, and computational models.

Faces as a Neuroscience Benchmark

Turned to face perception, the lecture introduces the social importance of faces and the existence of prosopagnosia, a condition where individuals have difficulty recognizing faces while retaining other perceptual abilities. The instructor uses a real-world example of a person with prosopagnosia to illustrate how specialized processes for faces contribute to social functioning and how studying these processes can inform our theories about mind and brain.

Functional MRI and the Human Brain

The session shifts to functional MRI as a premier method for imaging human brain activity noninvasively. The BOLD signal is explained as a proxy for neural activity, with caveats about temporal and spatial resolution. The classic finding of a face-selective region in the brain is introduced, along with the importance of rigorous controls to rule out alternative explanations for observed activations.

Closing and What’s Next

The lecture ends with a preview of later sessions on cognitive neuroscience methods and their application to brain function, using face perception as a case study to illustrate how perceptual theories, behavioral experiments, and neuroimaging come together to advance our understanding of the mind.

Key Takeaways

• Marr’s computational framework focuses on the problem to be solved, inputs, outputs, and the constraints that shape computation.
• Ill posed problems are common in perception, necessitating priors and contextual information for stable inferences.
• Color vision highlights how the brain infers object properties from ambiguous sensory data.
• Face perception illustrates domain-specific processing and how to test brain organization with psychophysics and MRI.