Rohit Ravi

ANN–BNN

Where learning lives in artificial and animal learners: a prior, an inference made in state, and a write into structure. One conditioning instrument, one ideal observer, and a week of pre-registered runs on the write channel.

Where learning lives in artificial and animal learners: a prior, an inference made in state, and a write into structure. One conditioning instrument, one ideal observer, and a week of pre-registered runs on the write channel.

Paired belief at a single probe cue against delay after the eighth reinforcement, 0.5C to 32C, for a meta-trained transformer. Massed training at C/T 4.5 reads 1.23, 1.40, 2.03 and 2.87; spaced training at C/T 27 reads 1.05, 1.09, 1.21 and 1.42; neither falls below the acquisition asymptote of 1. A hierarchical Bayes observer sits at 1.00 at every delay, and a fixed-parameter Bayes observer collapses to 0.03 at 32C on the massed schedule. Values above 1 are marked as a readout-scale effect.

Where learning lives: prior, inference, write

Every learner is three things: a prior it starts with, an inference it makes in state, and a write that makes the inference permanent. The laws animals and in-context learners share live in the second; the differences live in the first and third. A retention probe on a meta-trained transformer gives the inference-only signature, once its reference is read honestly.

 Three bars of log wait in the amendment R2 diagnostic: the median value the surprise-gated store held at the change point, 1.26 log units; the true post-switch wait, 1.91; and the non-contingent prefix marginal, 4.68. A note records that 51% of the stored observations were written before the switch and that the values were the same on contingent and control streams.

Eight hashed runs on the write channel, and what a gate outcome closes

A week of pre-registered tests asking whether a surprise-gated write channel adds anything an in-context prior lacks. Every prediction file was hashed before scoring; every gate stayed where it was; the predictions were never scored, eight times, and the eighth run explained the other seven.

 Write-only belief at a 0.5C probe after a state reset, for 20 change-point calibration streams at each confidence pseudo-count kappa from 16 down to 0.25. Every median sits between 0.001 and 0.002 against a dose gate of 0.5, the largest single stream reads 0.046, and false acquisitions on the control streams are 0 of 20 at every kappa.

The surprise gate is a biased sampler

A write channel gated on prediction error never opened where the meta-trained prior was good. Moved to streams where the prior is wrong, it wrote, but only chance-short waits, the same on signal and control: half the store filled before the change point, and the stored value sat at 1.26 log units against a true 1.91. The exploratory diagnostic that closed the error-surprise gate as an operationalisation and named its replacement.

 Left: the plasticity metric, per-step parameter displacement, over log training step for a transformer, a leaky GRU and an MLP, each trained on its task (solid) and on shuffled targets (dotted); every curve is largest at step 1 and decays, and the trained and shuffled curves lie on top of each other in all three substrates. Right: a 15 by 7 heat map of mean Spearman correlation between each training metric and each developmental curve; 7 of 105 pairs align in all three substrates, marked with a tick, and 4 of those also align under shuffled targets, marked with a cross.

Every curve looks like development

Fifteen training-trajectory metrics on three small networks against seven reconstructed developmental curves: 7 of 105 pairs align in every substrate, four of them also under shuffled targets, and all seven are a monotone or decaying curve against another. The battery's own shape classifier reported an early inverted U that review traced to a filled-in missing value. A shuffled-target control should be mandatory for any claim of this kind.

 Deficit in held-out loss after a 600-step blur window relative to matched clean training, against blur onset from step 0 to 2400, for a transformer, a leaky GRU and an MLP, three seeds each. Under Adam every point is within about 0.04 of zero with error bars crossing it. Under SGD the transformer reads 0.30 at onset 0, falling to 0.02 at onset 2400, and hollow markers show the same onsets with recovery tripled to 2700 steps: 0.038 and 0.010 for the transformer, 0.019 and 0.014 for the GRU, 0.007 and 0.009 for the MLP.

No critical period in a small network

A 600-step input deficit at five onsets, on three small networks under Adam and SGD. Under Adam the deficit is within seed noise at every onset. Under SGD it is largest at onset 0 and shrinks eightfold when recovery is tripled, from 0.30 to 0.038 on the transformer, still positive in all three seeds and still falling: recovery-limited on the evidence, not shown to be zero. What Achille's convnets add that these runs did not.

 Median reinforcements to a posterior-odds threshold against C/T − 1 from 2 to 107, log-log, for an ideal rate observer at thresholds of 1.4, 10, 30 and 60 nats, with the one-parameter pigeon model 297/(C/T − 1) dashed. Slopes on log(C/T − 1) over C/T 4.5 to 27 are −0.40, −0.53, −0.57 and −0.58 for the observer and −1.00 for the pigeon model; at C/T 4.5 the 60-nat observer reads 83 against the pigeon model's 85. The rat slopes, −0.66 across all groups and −0.82 below C/T 72, are noted.

Sixty nats

An ideal rate observer's trials-to-threshold slope on the animal axis, log(C/T − 1), stops moving above about 30 nats, at −0.57 to −0.58 over C/T 4.5 to 27, against −0.66 for rats across all groups and −1 by construction for the one-parameter pigeon model. At 60 nats it also lands on the pigeon intercept, 83 trials against 85 at C/T 4.5. One fitted parameter, one free prediction.