Marr's levels of analysis
In plain wordsA way of explaining any information-processing system on three levels: what problem it solves and why, what step-by-step method it uses, and how that method is physically built.
A framework, proposed by the vision scientist David Marr, holding that an information-processing system must be understood at three complementary levels: the computational level (what problem the system solves and why: its goal, its input–output mapping and the constraints that make the problem solvable), the algorithmic or representational level (what representations and procedures carry out the computation) and the implementational level (how these are physically realized, in neurons or silicon). Marr argued that understanding at one level does not settle the others, and that vision research had neglected the computational level. The framework is a staple of cognitive sciencecognitive science, n.: The scientific study of how minds work — how we reason, use language, remember and see — combining psychology, neuroscience, linguistics, philosophy, anthropology and artificial intelligence. and is widely invoked in AI: mechanistic interpretabilitymechanistic interpretability, n.: A branch of AI research that tries to reverse-engineer a neural network — working out, piece by piece, which internal parts represent which concepts and how they combine to produce behavior. is often described as recovering the algorithmic level of a trained network whose computational-level task is known but whose algorithm is not.
At Marr's computational level both networks solve the same task; the circuits show they use different algorithms.
Origin
Proposed by David Marr and Tomaso Poggio in an MIT Artificial Intelligence Laboratory memo, 'From Understanding Computation to Understanding Neural Circuitry' (1976), and made famous by Marr's posthumously published book Vision (1982).
Usage note
Poggio later proposed adding a level for learning. Philosophers note that the levels are relative: what counts as implementation in one analysis can itself be described as computation in another.
Related
- cognitive sciencecognitive science, n.: The scientific study of how minds work — how we reason, use language, remember and see — combining psychology, neuroscience, linguistics, philosophy, anthropology and artificial intelligence.
- computational theory of mindcomputational theory of mind, n.: The idea that thinking is a kind of computation: the mind processes information by following rules, much as a computer runs a program, and the brain is the machine doing it.
- mechanistic interpretabilitymechanistic interpretability, n.: A branch of AI research that tries to reverse-engineer a neural network — working out, piece by piece, which internal parts represent which concepts and how they combine to produce behavior.
- multiple realizabilitymultiple realizability, n.: The idea that the same mental state, such as pain, could be produced by very different physical systems — human brains, octopus brains or perhaps computers — just as one program can run on different machines.
- interpretabilityinterpretability, n.: The effort to understand how an AI system arrives at its outputs — what information it uses and how it combines it — instead of treating it as a sealed black box.
Sources
- Marr, D. & Poggio, T. (1976). From Understanding Computation to Understanding Neural Circuitry. MIT Artificial Intelligence Laboratory, AI Memo 357.
- Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. W. H. Freeman.
Mentioned in
cognitive sciencecognitive science, n.: The scientific study of how minds work — how we reason, use language, remember and see — combining psychology, neuroscience, linguistics, philosophy, anthropology and artificial intelligence., computational theory of mindcomputational theory of mind, n.: The idea that thinking is a kind of computation: the mind processes information by following rules, much as a computer runs a program, and the brain is the machine doing it.