Deep learning is the rapid reconstruction of hard-to-vary explanatory networks inside problem-situations
Status: surviving, 22 May 2026
Deep learning is the rapid reconstruction of hard-to-vary explanatory networks inside problem-situations.
The fundamental unit of learning is not information, fact, concept, or subject, but the hard-to-vary explanation: an account whose parts are constrained by the problem it solves. These explanations form networks with two kinds of edges: explanatory edges, which show what entails, supports, depends on, or constrains what; and critical edges, which show what would refute, weaken, expose, or delimit what.
These networks do not live primarily inside “subjects” or “domains.” Subjects are often sociological artifacts: university departments, textbook categories, professional boundaries, inherited labels. Problem-situations are functional artifacts: they organize reality around something not yet understood, something that does not work, something that resists explanation, or something that needs to be made possible. A problem-situation assigns weight to everything the learner encounters. It tells the learner what matters, what is peripheral, what is misleading, and what kind of explanation would count as progress.
But before a learner can reconstruct the explanatory network, they must absorb the epistemic grammar of the tradition they are entering: how claims are made, defended, criticized, formalized, tested, connected to evidence, and related to other claims. Mathematics speaks in definitions, proofs, counterexamples, invariants, and abstraction. Physics speaks in models, measurement, symmetry, idealization, and experiment. History speaks in sources, chronology, rival accounts, causation, and contingency. Without the grammar, the learner cannot even parse what kind of thing an explanation is supposed to be.
So the learning stack is:
Grammar lets you parse.
Explanations are what you parse.
Problem-situations decide what is worth parsing.
But in practice, the stack is recursive. Problem-finding often precedes problem-solving, because the received framing may itself be part of the confusion. The elite learner does not simply enter a field and ask, “What information must I absorb?” They ask, “What problem-situation makes these explanations necessary, what grammar governs their expression and criticism, and what network of hard-to-vary explanations must I reconstruct?”
Deep learning is therefore the process by which a learner enters a problem-situation, acquires the epistemic grammar needed to parse its tradition, reconstructs the explanatory and critical network within it, and reaches the point where they can regenerate, criticize, vary, and apply the explanations without depending on their original presentation.