Levels of Explanation

In brief

David Marr (1945 to 1980) argued that when you try to understand any system that processes information, a brain, a calculator, an eye, you are really facing three separate questions, and the great error is to run them together. The first asks what the system is doing and why: what problem does it solve? The second asks how: by what representations and what procedure? The third asks with what: in what physical stuff is the procedure realized? Marr called these the computational, algorithmic, and implementational levels. A description at one level does not translate cleanly into another, and a question posed at the wrong level cannot be answered at all. Explaining the brain as meat, the mind as software, or the task as pure logic each captures something real and none captures the others. The related idea of emergence presses further: the higher levels may have explanatory autonomy, doing genuine work no account of the parts can replace.

The full treatment

The problem it answers

By the 1970s the study of the mind was split between camps that talked past each other. Neuroscientists mapped cells and circuits; psychologists posited mental representations; logicians and early AI researchers wrote programs. Each treated its own level as the real one. Marr, working at the Massachusetts Institute of Technology on how the brain sees, came to believe this was why the field kept stalling. Knowing which neuron fires tells you almost nothing about what computation it helps perform, just as knowing the physics of transistors tells you nothing about what a chess program is trying to do. The levels framework is a diagnosis of category error: it names the different kinds of question so a researcher can tell which one they are actually asking.

How it works

Marr's three levels, laid out in Vision (1982) and an earlier 1976 memo with Tomaso Poggio, run from the abstract to the concrete.

The computational level (also called the theory of the computation) asks what the system computes and why: what is the goal, what logic makes achieving it possible, and what constraints does the world impose? For vision, the question is how a two-dimensional pattern of light on the retina can be turned into knowledge of a three-dimensional world of surfaces and objects. This is the level of the problem, stated apart from any method.

The algorithmic (or representational) level asks how the computation is carried out: what representations does the system use for its input and output, and what procedure transforms one into the other? The same problem admits many algorithms. Addition can be done by counting, by lookup table, or by carrying digits. Marr's own vision theory proposed a sequence: a "primal sketch" of edges and blobs, then a "2.5-dimensional sketch" of surface orientation and depth, then a full three-dimensional model.

The implementational level asks how the representations and algorithm are physically realized: in neurons and synapses, in silicon, in gears. Marr's famous claim is that this level is largely independent of the other two. The same algorithm can run on wetware or hardware; the same computation can be reached by different algorithms.

What it claims

The strong claim goes past saying these are three useful perspectives. Marr held that they are logically distinct and only loosely coupled, so an explanation pitched at one level cannot be derived from, or reduced to, one at the level below. You will not read the purpose of vision off a wiring diagram, and you will not predict the wiring from the purpose. For Marr the computational level was the most neglected and most important: without knowing what problem is being solved, data about mechanism is uninterpretable. This inverts the reductionist instinct that understanding must be built upward from the smallest parts.

The related idea: emergence

Marr's levels describe how to analyze a system. Emergence makes an ontological claim about why the higher levels can stand on their own. In "More Is Different" (1972), the physicist Philip Anderson argued that at each scale of complexity genuinely new properties appear, so that "psychology is not applied biology, nor is biology applied chemistry." The philosophers Hilary Putnam and Jerry Fodor gave the point a sharp form with multiple realizability: a mental state like pain, or a computational state like "storing a value," can be realized in brains, in circuits, in anything with the right functional organization, so its scientific description cannot be identified with any one physical one. Reduce it away and you lose the very generalization that made it useful. This is the deeper reason Marr's implementational independence holds, and why the study of wholes (see emergence) does not collapse into lower-level physics awaiting patience.

Lineage

The distinction between a computation and its physical substrate was already implicit in Alan Turing's abstract machine. Marr's immediate context was the disappointment of 1970s AI, whose vision programs worked in toy worlds and failed in real ones; he argued they failed because they skipped the computational level and jumped straight to algorithms. The autonomy of the higher levels was armed philosophically by Putnam's functionalism (1967) and Fodor's defense of the "special sciences" (1974), and it rhymes with the older tradition of emergence running back through John Stuart Mill. Within psychology it is a cousin of the decompositional strategy behind localization of function (see localization-of-function): both explain the mind by carving it into parts, but Marr's levels carve by kind of explanation rather than by region.

The strongest case for it

The framework's power is that it dissolves arguments that were never really disagreements. When a neuroscientist and a psychologist quarrel over whether memory "is" a synaptic change or a retrieval process, Marr shows both are right about different levels. It also gives research direction: it explains why decades of recording single neurons yielded so little understanding of perception, the computational theory being missing, and why stating the problem cleanly (as Marr did for stereo vision, with Poggio) can drive progress at every level below. The map-is-not-the-territory caution applies here too (see the-map-is-not-the-territory): a description of the hardware is not the same object as a description of the task. And multiple realizability makes the autonomy more than a convenience. If minds can run on different stuff, a theory of mind pitched at the level of stuff looks in the wrong place.

The strongest case against it

The sharpest objection is that the levels are not as separable as Marr claimed. In real biological systems the implementation constrains the algorithm heavily: brains are slow, parallel, noisy, and energy-limited, and these facts rule out whole classes of algorithm, so you cannot fix the computational theory first and treat the hardware as an afterthought. This became vivid with the rise of connectionism in the 1980s, where a network's architecture and its function are hard to prise apart, and Marr's tidy top-down ordering does not describe how such systems are understood. From the dynamical-systems and embodied-cognition traditions, Tim van Gelder and Andy Clark argued that treating cognition as computation over representations, the very frame Marr assumes, may itself be the mistake, and that some cognitive systems are better described as coupled dynamical processes with no clean representational level. A reductionist reply, in the spirit of Steven Weinberg, holds that the independence of levels reflects our practical limits, not the world's structure: in principle the arrows of explanation still point down. Jaegwon Kim pressed the version that bites hardest on autonomy claims, arguing that if every physical event already has a sufficient physical cause, no causal work is left for an emergent higher level to do without breaking physics. No consensus reply to his exclusion argument exists.

Where it stands now

Marr's three levels are standard equipment in cognitive science, taught in nearly every introduction to the field, and Vision remains a touchstone even for researchers who reject its specifics. The consensus is a qualified acceptance: the distinction between asking what a system computes, how, and in what medium is indispensable, but the strict top-down independence Marr sometimes suggested is not tenable, and the levels interact in both directions. On the emergence side, the pattern tracks the broader debate: the explanatory autonomy of higher levels (weak emergence) is broadly accepted, while strong metaphysical autonomy remains contested. What endures is Marr's discipline, forced on anyone who studies minds or machines: before you explain how something works, be sure you know which question you are answering.

Test yourself

Think of something you understand at one level but not others: how a search engine ranks results, how you recognize a friend's face, how a market sets a price. Ask which of Marr's three questions you can actually answer. Do you know what problem is being solved and why (the computational level)? The procedure and representations it uses (the algorithmic level)? The physical machinery that runs it (the implementational level)? Most confident-sounding explanations answer one of the three and quietly assume they have answered all of them.