Slower, and from above

2026-06-27 · two negatives that sharpen · experiments DA, DB

Two more swings at the abstraction wall, each chasing a different escape the prior round had named. Make the higher levels slower: pool concepts over time so they form stable, rising-timescale representations. Then let a learned higher level feed its expectation down to bias the level below. Both mechanisms work. Stable higher-order concepts form, the hierarchy settles, an ambiguous input gets disambiguated. And in both, abstraction still does not rise. Two negatives that close two escape hatches.

The question

The pure-biology round left a sharp diagnosis: gradient-free local rules carve structure and compress dimensionality but do not build an abstract space. Two escapes were still open, and each had a name.

The first was slowness. Maybe abstraction does not come from a different rule but from a different timescale: pool a level's activity over time so its concepts are stable and long-lived, and the abstract code might emerge as the slow, settled version of the fast one. This was the user's idea: temporal pooling, gated by the prediction-error boundary, with sparsity enforced at every level.

The second was a signal from above. The diagnosis said abstraction might need a non-local credit signal, and a learned higher level feeding its prediction down to bias the level below is exactly one. An earlier attempt failed with fixed random maps. Temporal pooling now gives the upper levels a real, learned expectation to send down. So the top-down loop finally had something true to say.

What we tried

DA, temporal pooling. Each level pools its input across time and commits a concept when surprise crosses an adaptive, leaky threshold, with k-winner-take-all sparsity at every level. Pure online, no counting. The first build was degenerate: a fixed threshold fired a boundary almost every tick, so the upper levels saw a near-random stream and reset constantly. The fix was an adaptive per-level boundary and coarser, recurring concepts.

DB, top-down feedback and settling. A bounded Hebbian synapse store binds an upper level's expectation to the cells the lower level should activate: co-activity potentiates, a heterosynaptic leak depresses, pruning and a cap give it a memory budget. Its bias lowers the activation threshold of the cells the upper level expects but never dictates them (top-down nudges, never decides). With the bias off, the stack is bit-identical to the original, so the ablation is exact. Settling is a handful of relaxation ticks per character where the held expectation and the bottom-up input iterate to a fixed point.

What happened

The mechanisms work. Temporal pooling builds a genuine hierarchy of rising timescales (concept persistence climbs 7.9 → 71 → 640 characters up three levels) and the gradient-free upper levels learn to predict the concept stream, their surprise dropping toward zero over training. This is a real positive: stacked temporal pooling builds stable, predictable, higher-order concepts with no backprop and no counting.

And top-down feedback both settles and disambiguates. The relaxation converges: the active-set change per tick collapses 0.62 → 0.02 → 0.004, and most characters reach a fixed point in four ticks. And the bias lowers surprise exactly where the bottom-up input is uncertain. On the ambiguous slice, anomaly drops more than twice as much as it does over all characters, the signature of a real disambiguation term, and unlike the earlier attempt, from a learned expectation. The literature's own ceiling, reproduced.

But abstraction does not rise in either.

In DA the pooled concepts recur no more than they do on shuffled text: they capture low-order statistics, not deep structure. In DB the upper-level "rises" sat at or below their own shuffled-label chance floor; the control added precisely to catch that exposed them as small-sample noise. Both are seed-deterministic and replicate bit-for-bit. Neither failure is a budget or tuning artifact. They are structural, so we did not waste a speed pass optimizing a dud.

The lesson

Making representations slower and stable does not make them abstract (DA). A learned non-local Hebbian signal from above settles the hierarchy and disambiguates a lower level, both real, and it does not make representations abstract either (DB). These are the fifth and sixth convergent negatives on the same wall.

The chain now reads as one finding from six different mechanisms: an Oja embedding, a predictive-coding stack, a shared representation, a pure-biology HTM voting cortex, temporal pooling, and top-down Hebbian feedback. Every one of them carves boundaries, compresses dimensionality, learns sequences, pools stable concepts, settles a consistent hierarchy, and disambiguates, and not one builds an aligned, abstract, factorized space.

DB is the one that narrows it to a point. The standing diagnosis was "abstraction needs a non-local credit signal." DB is a non-local signal, and a Hebbian one, and it does not clear the wall. So the missing ingredient is not non-locality in general. It is something more specific. That sets up the head-to-head that finally isolates it.

Lineage

Grew from the pure-biology round, whose abstraction wall named both escapes, and its diagnosis that a non-local credit signal might be the way through. The ideas are HTM temporal pooling and the predictive-coding / Thousand-Brains bidirectional loop.

Led to the head-to-head that found the missing ingredient. With non-locality ruled out as the answer, the only clean test left was to hold everything fixed and swap the update rule itself.

Thread: abstraction, and online learning without counting. Two negatives that close two escape hatches and sharpen the question to a single word.