Levels of Explanation

Mind and brain can be described at many scales: molecules, cells, circuits, algorithms, representations, beliefs, actions, and social practices. Levels of explanation are ways of organising these descriptions according to the questions, variables, and mechanisms they address. A neural account does not automatically compete with a psychological one, and a useful higher-level explanation is not automatically irreducible.

The central methodological problem is integration. Which higher-level kinds map onto mechanisms? When does lower-level detail explain a capacity rather than merely correlate with it? When are two explanations complementary, and when do they make incompatible claims about the same dependence?

Marr's three levels

David Marr distinguished:

1. Computational level: what problem does the system solve, and why is that problem appropriate?
2. Algorithmic/representational level: which representations and procedures solve it?
3. Implementational level: how are those procedures physically realised?

For vision, one may ask why depth must be recovered, which representations and algorithms estimate it, and which neural circuits implement them. Answering only one leaves the others open.

The levels constrain each other. Hardware limits feasible algorithms; algorithmic structure predicts lesions and timing; ecological tasks determine which output counts as success. Marr's framework is not a licence to study each level in isolation.

Personal and subpersonal explanation

Personal-level explanations attribute perceptions, beliefs, intentions, and reasons to an agent. Subpersonal explanations describe processes by which the agent's capacities are implemented: feature extraction, prediction errors, memory consolidation, and motor control.

Confusing levels produces category mistakes. A neural population does not literally believe a proposition merely because its activity carries information; a person does not calculate every equation used in a model of their visual system.

Subpersonal explanations can correct personal-level assumptions, and personal-level tasks help individuate mechanisms. The distinction concerns explanatory role, not two worlds or a ban on cross-level relations.

Mechanisms and decomposition

A mechanism explains a capacity through organised entities and activities. Researchers decompose memory, attention, or control into components and localise operations through intervention and dissociation.

Successful decomposition is neither pure reduction nor mere correlation. It shows how a higher-level capacity is produced while preserving the capacity as the phenomenon being explained.

Mechanisms can be nested. A circuit operation is realised by cells, which are realised through molecular processes. The useful stopping point depends on the contrast: a drug's effect may require molecular detail, while a planning error may be explained at an algorithmic level.

Multiple realisation

A higher-level state is multiply realised when different lower-level structures can perform the same role. This supports autonomy because generalisations can range over different species, individuals, or machines.

Multiple realisation is empirical and grain-relative. Two systems may differ molecularly while share network dynamics, or differ in dynamics while match coarse input-output function. The right level must be specified rather than assumed.

The phenomenon blocks simple one-type/one-neural-type reductions. It does not block mechanistic explanation or token physicalism. The relation is developed metaphysically in Non-Reductive Physicalism.

Reduction

Reduction can mean:

• ontological: no additional entities are required;
• type: higher-level kinds are identical with lower-level kinds;
• theoretical: higher-level laws are derived through bridge principles;
• explanatory: lower-level mechanisms make the higher phenomenon intelligible;
• methodological: inquiry should prioritise lower levels.

These claims do not travel together. A capacity can be ontologically physical, mechanistically explained, and still require higher-level models for prediction. Failure of type reduction does not establish non-physical properties.

Likewise, successful neural prediction is not explanatory reduction unless it identifies why the relevant organisation produces the capacity.

Autonomy

A level is explanatorily autonomous when its generalisations remain stable across lower-level variation and support interventions. Software, population biology, and economics provide familiar examples.

Psychological kinds can be autonomous because they group states by content and role. A belief that danger is near predicts action across diverse neural implementations.

Autonomy is not insulation. Higher-level explanations should integrate with known mechanisms and respect physical constraints. A level that cannot connect with any intervention or lower-level process risks becoming merely interpretive.

When explanations complement

Two explanations complement when they answer different contrasts. "Why did she leave?" may be answered by her belief about danger, by a decision mechanism, or by motor execution. Each can be true.

The reason explanation identifies rational content; the mechanism explains production; the neural implementation explains physical realisation. No duplication occurs if the levels describe one process.

Complementarity requires consistency. A psychological explanation claiming flexible representation conflicts with a mechanism showing a fixed reflex, because both address the organisation responsible for variation.

When explanations compete

Explanations compete when they assign incompatible dependencies or deny the relevance of variables used by the other. A memory theory based on stored traces conflicts with one claiming current reconstruction alone if an intervention can distinguish their predictions.

Cross-level claims also compete over constitution. If a theory says global broadcast is consciousness and another says it only causes report, they disagree despite sharing a correlation.

The test is not vocabulary overlap but counterfactual structure: would the explanations predict different outcomes under a controlled change?

Top-down and bottom-up constraints

Bottom-up constraints arise from implementation. A proposed algorithm must fit timing, capacity, connectivity, and lesion evidence.

Top-down constraints arise from the task and organised system. Knowing what a circuit must compute guides search for mechanisms; organism-level goals constrain which component effects count as functional.

Top-down explanation need not mean a new force pushing components. Organisational boundary conditions and control can be realised through lower-level interactions. The causal question is treated in Higher-Level Causation.

Neural decoding

A decoder can predict a stimulus or report from neural activity. This establishes information availability to the analyst. It does not show the brain uses the same code, that the decoded variable causes behaviour, or that the subject consciously represents it.

Encoding asks how a variable affects activity; decoding asks what can be recovered. Both can succeed for correlated but causally irrelevant features.

Mechanistic interpretation needs intervention, generalisation, and a consumer process that uses the information. Otherwise "representation" may describe the experimenter's model rather than the system's cognition.

Network and region explanations

Mental capacities rarely reside in one region. A local lesion can be necessary because it disrupts a distributed network; activation in one region can reflect input, output, or control rather than the target process.

Network explanations map interactions, dynamics, and connectivity. They should not replace one simplistic localisation with the claim that "the whole brain" does everything. Components can have differentiated causal roles within distributed systems.

The appropriate unit is determined by intervention and mechanism, not the visual scale of an image.

Cross-species and artificial comparison

Comparative explanation requires functional homology at an identified level. Similar behaviour may arise through different algorithms; similar circuits may serve different ecological tasks.

Artificial systems can realise algorithmic structures without biological implementation. This supports comparison for capacities defined computationally and weakens it for theories requiring metabolism or specific intrinsic dynamics.

Claims should name the preserved level: task equivalence, algorithmic similarity, dynamical similarity, or material homology. "Brain-like" and "human-level" are too vague.

A working integration protocol

For a cognitive explanation:

1. State the phenomenon at the personal or behavioural level.
2. Specify the computation or functional organisation proposed.
3. Identify an implementational mechanism.
4. Derive predictions from relations among levels.
5. Intervene at one level and test effects at others.
6. Check whether alternative organisations predict the same result.

This protocol makes levels mutually constraining. It avoids both premature reduction and a protected higher-level vocabulary answerable to no mechanism.

Assessment

Levels are explanatory abstractions grounded in real organisation, not a fixed cosmic staircase. Different tasks and timescales support different decompositions. Reduction, realisation, mechanism, and autonomy should be judged separately.

Mind science progresses when it links personal capacities to algorithms and mechanisms without confusing those links with identity or elimination. The same discipline governs the science of consciousness, where correlations are plentiful and constitutive claims remain disputed.

Selected references

• Bechtel, William. Mental Mechanisms (2008).
• Craver, Carl F. Explaining the Brain (2007).
• Fodor, Jerry A. "Special Sciences" (1974).
• Marr, David. Vision (1982).
• Woodward, James. Making Things Happen (2003).