Canonical · Relation to Adjacent Work

Hopfield Networks

The cleanest energy landscape there is — and why its tidiness is the thing to watch

A Hopfield network is the textbook case of attractor dynamics: a set of units whose updates minimize an energy function, so that any starting pattern rolls downhill into the nearest of a set of stored minima. Those minima are memories; releasing a corrupted pattern and watching it settle into the clean stored one is content-addressable recall, pattern completion by descent. No system lands on this framework's vocabulary more neatly — energy, basin, minimum, settling, recovery. That neatness is exactly why this page moves carefully. The framework reserves a formal result about Hopfield dynamics for its own future domain-specific work, and this page does not pre-empt it. What it does instead is smaller and honest: it reads the Hopfield attractor against the framework's existing distinctions — the supplied landscape, the sovereignty test, the difference between a bare map and a held structure — and it defers the formal theorem, clearly labeled, to where it belongs.

The phenomenon

Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences. [verify volume/issue/page range and DOI against the primary source before publication]

The network is a set of binary or continuous units connected by symmetric weights. Given those weights, one can write an energy function over the network's states, and the update rule is arranged so that every permitted step lowers the energy. The dynamics therefore descend to a local minimum and halt. The stored patterns are placed at minima by the choice of weights; an input pattern is the starting point of a descent, and the minimum it reaches is the “recalled” memory. Perturb a stored pattern a little and the descent returns it to the same minimum — the pattern-completion behavior. In the framework's terms this is a genuine multi-basin landscape with genuine return-to-identity after a bounded nudge. [verify the energy-descent, symmetric-weight, and content-addressable-recall description against the primary literature]

Where the framework agrees

The dynamics are real attractor dynamics. The stored minima are attractors; the sets of inputs that flow to each are their basins; the return of a corrupted input to its stored pattern is return-to-identity after bounded perturbation — the shape of the return the framework reads elsewhere, and the same effect visible in the live Eco 1.0 simulation. The energy function is a clean scalar potential and the descent is honest gradient-like relaxation. Of all the neighbors treated in this section, none matches the framework's descriptive vocabulary as directly as this one. On the dynamics, there is no dispute.

Where the framework draws its line

The landscape is supplied — so, as with Waddington, it is not sovereign. The energy function is not authored by the network's own activity; it is set by the weight matrix, and the weight matrix is handed in — installed by a storage rule or a training procedure before the network is ever run. This is the same situation as Waddington's ball: a structure descending a landscape it did not make. Measured against the four sovereignty conditions, the network does not self-produce its boundary (Condition 3 — the landscape that defines its basins is external) and does not bear its own maintenance out of its own activity (Condition 4). A Hopfield attractor is not a sovereign attractor. Its stability is real, and it is borrowed. [verify that the weight matrix is fixed prior to recall dynamics in the standard model]
But which non-sovereign thing is it? The framework's catalogue has two non-sovereign readings, and the Hopfield network sits interestingly between them. It is not quite attractlet in the strict sense: an attractlet is a structure held up by an external agent that is actively supplying stability, and once the weights are frozen, nothing outside the running network is actively holding its minima in place — they are baked into a fixed map. That makes it look, instead, like the random-Boolean-network case: a bare deterministic map whose fixed points simply are, for which the “who is paying to keep it there?” question does not really apply during recall, because the descent runs for free once the weights are set. On this reading a Hopfield recall is closer to a bare mathematical fixed point than to an actively-sustained attractlet — neither sovereign nor attractlet, but the third, map-like case the RBN also occupies. [verify the fixed-weights-during-recall assumption; learning-enabled or continuously-driven variants may read differently]

A Hopfield network descends a landscape as cleanly as anything in dynamics — but it did not build the landscape, and once the weights are frozen nothing is actively holding its memories in place. So it is not sovereign, and it is not, strictly, an attractlet either: it is closest to a bare fixed map whose minima simply exist, the same map-like verdict the framework gives a random Boolean network. The tidiness is real; the sovereignty is absent.

What this page does not do: the reserved theorem

A formal result is reserved, and this page does not state it. The framework earmarks a domain-specific theorem about Hopfield dynamics — a bootstrapping-interval result — as future work, on the same footing as its other reserved domain instantiations. That theorem, if and when it is developed, would be the place where the framework says something formal and Hopfield-specific: about when the recall dynamics constitute a bootstrapping interval, about how the framework's interval-bound quantities apply to a settling network, and about what, if anything, changes when the weights are allowed to learn rather than stay frozen. None of that is asserted here. This page is an adjacent-work reading — it locates the Hopfield attractor among the framework's existing distinctions — and it stops at the edge of the reserved result rather than pre-empting it. Treat anything on this page as classification, not as the theorem.
Where the interesting formal question lives. The frozen-weight network is map-like and settled; the genuinely live question — and the natural home of the reserved theorem — is what happens when the landscape is not frozen: a network that updates its own weights from its own activity begins to reshape the very landscape it descends, which is the recursive coupling the framework is built to describe (and the same step that, on the Waddington page, carried the picture from a supplied landscape toward the framework's own ground). Whether a learning Hopfield-type system crosses from map-like recall into something that bootstraps its own structure is exactly the kind of threshold the framework cares about — and exactly what the reserved theorem would have to adjudicate. This page raises the question and leaves it open.
Where this sits, and what it leaves open. This page reads the Hopfield network through the framework's existing distinctions and defers the framework's reserved formal result. Threads deliberately left open:
  • The reserved bootstrapping-interval theorem — the formal, Hopfield-specific result the framework earmarks as future domain work. Not stated here; flagged as reserved and not-yet-canon.
  • Learning versus frozen weights. The map-like verdict above is for fixed weights during recall. A network that adapts its own weights reshapes its own landscape, which is a different and richer case — the one where the interesting formal question lives. [verify how weight-update variants alter the standard picture]
  • Modern and continuous variants. Dense / modern Hopfield networks and their relation to attention mechanisms are a substantial recent development; whether they change the classification is left for separate treatment. [verify]
  • A simulation is deliberately withheld. Consistent with the framework's discipline that structural framing precedes empirical demonstration, no Hopfield simulation is offered here; one would be built only to test a specific claim of the reserved theorem, not to illustrate settling in the abstract.
None of this touches the framework's fixed foundations; the six substrate conditions are closed and are not at issue on this page.
Adjacent-work assessments state where Principia Attractum agrees with and departs from neighboring frameworks. They introduce no constructs and modify no canon; they locate the framework relative to its field. The reading of a frozen-weight Hopfield network as a supplied-landscape, non-sovereign, map-like structure is a downstream application of the framework's existing distinctions. The framework's reserved Hopfield bootstrapping-interval theorem is not stated on this page and remains future, not-yet-ratified work; nothing here should be read as asserting it. Technical characterizations are marked [verify] and should be confirmed against the primary literature before publication.
← Back to Relation to Adjacent Work