One training epoch can erode a network's resemblance to visual cortex

What the work claims

This is a single-author computational study, best read as two papers in one file. The original study trained a small convolutional network with four learning rules (backpropagation, feedback alignment, predictive coding, and a spike-timing-dependent plasticity approximation) on 8,000 CIFAR-10 images for 40 epochs, extracted representational dissimilarity matrices at eight checkpoints, and compared them against fMRI dissimilarity matrices from three human subjects viewing 720 THINGS images across six visual regions.¹ Alignment was the Spearman correlation between model and brain dissimilarity matrices, reported per region with the best layer chosen independently at each epoch.

The v1 claim was that all rules degrade V1 alignment but local rules degrade it least: backpropagation lost 90 percent of its V1 alignment after a single epoch (Spearman r falling from 0.102 to 0.011), while predictive coding and STDP lost only about a quarter. The August 2026 correction, printed at the top of version 2, states that this central comparative finding is not supported. The predictive-coding and STDP implementations overrode eval() with a no-op, leaving batch-normalization layers in training mode during feature extraction; each evaluation batch was then normalized by its own statistics, which removed precisely the activation drift the study was measuring. Repaired, predictive coding degrades V1 alignment more than backpropagation does (epoch-0-to-40 change of minus 0.0403 plus or minus 0.0034, versus minus 0.031 for backpropagation, with all five seeds negative in both affected conditions), and the STDP condition degrades as well (minus 0.0175 plus or minus 0.0033). Backpropagation and feedback alignment results were essentially unchanged (plus 0.0001 and plus 0.0004 at 224-pixel evaluation).

How it works

The machinery is representational similarity analysis. A shared three-block convolutional architecture (32, 64, then 128 filters, batch normalization, ReLU, max pooling, a 512-unit fully connected layer) was trained under each rule: exact gradients for backpropagation, fixed random feedback weights for feedback alignment, ten iterative inference steps minimizing local prediction errors for predictive coding, and Poisson-converted spike trains with an exponential STDP kernel for the STDP approximation. Five random seeds were run per rule, with checkpoints at epochs 0, 1, 2, 5, 10, 20, 30, and 40. Epoch 0 is the untrained control: a randomly initialized network, which prior work had already shown can match or beat trained networks in V1 similarity.

The brain side uses the THINGS-fMRI dataset: blood-oxygen responses from three subjects, six regions of interest from V1 through inferior temporal cortex, dissimilarity matrices built with correlation distance over the 720 shared images. Model and brain matrices are compared with Spearman rank correlation, so alignment reflects geometry, not absolute response levels. Two design details matter for reading the results. First, statistical power is hard-capped: with five seeds, a paired sign-permutation test can reach no lower than one-sided p = 1/32, about 0.031, so the paper reports that floor as evidence of direction consistency rather than tail probability. Second, all absolute correlations are below 0.10, so even large relative changes are small in absolute terms.

The defect itself is instructive. Batch normalization stores running statistics in training and uses them in evaluation; leaving the layer in training mode recomputes mean and variance per batch, which erases the slow drift of activation statistics across epochs that the alignment metric was designed to detect. The apparent preservation of alignment by the two local rules was an artifact of the measurement pipeline, not a property of the learning rules. A companion analysis² adds a second, subtler caveat: most of the apparent backpropagation degradation at 224-pixel evaluation depends on the evaluation resolution, since the networks trained on 32-pixel CIFAR-10 images return to their starting alignment when evaluated at the training resolution (change of minus 0.000 plus or minus 0.005, three of five seeds negative).

Where a skeptic should push

Even the surviving finding deserves pressure. The claim that training universally degrades V1 alignment rests on small absolute correlations, three fMRI subjects, one architecture family, and a resolution mismatch between training and evaluation that a companion paper shows can dominate the effect for the backpropagation condition. The LOC increase reported for backpropagation (delta r = +0.011) was explicitly not significance-tested. And the corrected comparative claims still carry a caveat from the author: in the repaired implementation, the predictive-coding network's first convolutional layer moves only about 2 percent of its initialization norm over 40 epochs and its third layer not at all, so its corrected trajectory may be carried by normalization statistics rather than by the weight-update rule. That is a second measurement confound inside the correction itself.

The fair summary is that this paper now demonstrates less about learning rules and more about the fragility of alignment research: one eval-mode bug inverted a published ranking, one resolution choice manufactured most of another result, and both were caught only because the author re-checked. Readers should credit the retraction and discount the specifics.

A warning shot for organoid benchmarking

Organoid intelligence leans hard on brain-likeness arguments: readouts validated by how well an encoding matches cortical or ANN reference representations, and narratives in which living tissue becomes more brain-like as it is trained. This paper puts both under stress with a concrete mechanism rather than a philosophical objection. If a deterministic eval-mode flag can invert the comparative ranking of four learning rules, any organoid pipeline that normalizes, standardizes, or rescales activity before scoring alignment must show that its preprocessing statistics are frozen and identical across the conditions being compared. If evaluation resolution alone can erase the apparent degradation for backpropagation, any cross-substrate alignment claim that feeds tissue recordings through a different resolution, binning, or embedding pipeline than its reference is measuring the pipeline as much as the tissue.

The deeper implication concerns what training does to a living substrate. The result that survives every correction is that optimization away from initialization moves representations away from the geometry of early sensory cortex, quickly. For organoids this cuts against a common assumption: that closed-loop training makes a culture more cortex-like. There is no theorem saying trained tissue drifts toward native cortical statistics, and this study's surviving finding is the opposite sign for in-silico networks. That reframes the honest architecture of the field. The reservoir configuration, where the biological substrate is left untrained and only the readout adapts, is not a compromise forced by our inability to train tissue; it is the configuration that preserves the native dynamics that constitute the substrate's claimed value. A task-trained organoid may be a better computer and a less brain-like one at the same time, and the field should expect to have to choose, measure both axes separately, and stop letting one metric stand in for the other.

The governance read follows directly. Claims like the organoid learned and it now resembles cortex more closely are cheap to make, and this preprint shows how quietly wrong they can be, because the author here did the expensive thing: audited the evaluation code, published the defect at the top of the abstract, and pointed readers to the superseding analysis. That is the standard of evidence a field selling biological computers will be held to by everyone outside it.

The bottom line

Established after correction: supervised training degrades V1 alignment for every learning rule examined, in this architecture and evaluation setup, and eval-mode handling and evaluation resolution are first-order confounds for any alignment metric. Not established: any clean ranking of learning rules by brain-likeness, and any large absolute effect (all correlations stay below 0.10). What would confirm the surviving claim is a preregistered replication with fixed layer-to-region mappings, frozen normalization, matched train and evaluation resolutions, and more subjects; what would weaken it is evidence that the degradation vanishes once resolution and preprocessing are fully controlled. For organoid intelligence the actionable residue is procedural: freeze your preprocessing, report alignment per checkpoint, and never let training-curve success and brain-likeness share a single number.