A temporal-derivative theory of how the neocortex learns

What the work claims

This is a position paper, not a primary experimental report, and it should be weighted accordingly. O'Reilly argues that a sufficient theory of neocortical learning must satisfy three levels: computational, algorithmic, and implementational.¹ At the computational level, he contends that only error backpropagation has been shown to scale to human-level competence, so any biological learning mechanism must approximate it. At the algorithmic level, he proposes that the cortex does so through a temporal derivative: the network first settles into a prediction state, then an outcome state is introduced, and the difference between the two activity patterns provides an implicit error gradient without requiring a separate population of error-coding neurons. At the implementational level, he suggests that the required temporal derivative is realized by a competition between two calcium-activated kinases, CaMKII and DAPK1, which integrate the same calmodulin-driven signal on different timescales.

The paper cites a recent experimental preprint by Jang, Flores, Zito, and O'Reilly as supporting evidence.¹ In that preparation, pre- and postsynaptic pyramidal neurons were driven at 25 Hz or 50 Hz during the first 100 ms of a 200 ms theta cycle and at 25 Hz or 50 Hz during the second 100 ms. After ten repetitions, a rising pattern from 25 Hz to 50 Hz produced long-term potentiation, a falling pattern from 50 Hz to 25 Hz produced long-term depression, and the two flat patterns produced no net change. The paper interprets these results as consistent with a temporal-derivative learning rule.

How it works

The algorithmic core is the contrast between a prediction state and an outcome state. In the simplest three-layer sketch, sensory input drives a hidden layer and a prediction output. Bidirectional excitatory connections let the network settle into a prediction. Then a stronger outcome input drives the output layer to a different activity pattern, and that difference propagates top-down through the same bidirectional connections, changing the hidden-layer activity. The error signal is the plus-minus difference in each neuron's activity over time, not a separate error-carrying pathway.

The paper places this mechanism in corticothalamic circuitry. Layer 6 pyramidal neurons send many relatively weak projections to higher-order thalamic nuclei such as the pulvinar, and these projections are proposed to synthesize the prediction. Layer 5b intrinsic bursting neurons send a small number of strong driver inputs to the same thalamic cells, proposed to provide the outcome. The 5b neurons are said to burst roughly every 100 to 200 ms, corresponding to alpha or theta frequency, so one prediction-outcome cycle fits within about 200 ms.¹ The thalamic relay cells then project back to cortex, carrying the temporal difference that drives local synaptic plasticity.

At the synaptic level, the paper proposes that the temporal derivative is computed as the difference between fast and slow integrals of a shared calcium-calmodulin signal. CaMKII is suggested as the fast integrator and DAPK1 as the slow integrator. If CaMKII dominates during the outcome relative to the prediction, the synapse potentiates; if DAPK1 dominates, it depresses. The paper notes that the learning rule has been implemented in the Axon neural simulation framework using spiking neurons.

Where a skeptic should push

The load-bearing assumption is that backpropagation is the only viable computational-level learning algorithm. O'Reilly treats this as close to settled, but the history of machine learning contains multiple revolutions, and biological evolution is not constrained to rediscover the technique that happens to dominate current artificial networks. Even if backpropagation is the right target, the paper must still show that the temporal-derivative mechanism approximates it accurately in deep networks with realistic activity and noise, not just in toy examples.

The experimental support is also more suggestive than conclusive. The Jang et al. study manipulates firing-rate trajectories over 200 ms windows and observes LTP, LTD, or no change. That pattern is consistent with a temporal-derivative rule, but it is also consistent with other rate-based or timing-dependent plasticity rules, and the paper does not rule out all alternatives. Furthermore, the experiments were performed in vitro on pyramidal neurons, so their relevance to intact cortical circuits, to inhibitory synapses, and to recurrent networks remains open. The paper is explicit that more empirical research is needed, and I read it as a bold, coherent hypothesis rather than an established result.

What prediction-outcome learning means for organoid training

The non-obvious implication is that training organoid intelligence may require structured prediction-outcome sequences, not just global reward or random stimulation. Most current demonstrations of computation in neural cultures and organoids use closed-loop feedback, electrical stimulation patterns, or reinforcement-style reward signals. If the neocortex actually learns by comparing predicted and actual activation states over a theta-scale window, then the most effective training protocol for a living substrate might be one that presents a prediction cue, allows the network to settle, then presents an outcome cue that drives a different activity pattern, and repeats this pairing.

The opportunity is a more biologically grounded training curriculum. Organoids already show spontaneous and stimulus-evoked activity. A temporal-derivative training scheme would attempt to shape that activity by pairing specific precursor patterns with target patterns, letting the network's own bidirectional dynamics compute the required credit assignment. The mechanism from the source supports this: the thalamocortical loop is proposed to generate the two states, and the local synaptic change is driven by their difference. In an organoid, one might approximate the same structure with patterned electrical or optogenetic input that first establishes a predictive state and then imposes an outcome state, with the delay between them tuned to the tissue's intrinsic timescales.

The threat is that organoids may lack the anatomical ingredients the theory requires. The temporal-derivative model depends on strong, focal driver inputs and numerous weak prediction inputs converging on a thalamic-like relay population, then projecting back to cortex. Organoids do not have a thalamus in the same sense, and their laminar organization is rudimentary compared with cortex. If the algorithmic mechanism really needs corticothalamic circuitry, then organoids may be missing a component that neocortical learning cannot do without. The honest response is that this is an open empirical question: either organoids can approximate the mechanism with their own recurrent architecture, or they will need engineered inputs that substitute for thalamic driver and prediction channels.

There is also an ethics angle worth stating carefully. The theory says that the cortex learns by predicting future states, which is one of the capacities often associated with flexible cognition. If organoid training were made substantially more effective by adopting prediction-outcome curricula, the same advance could bring the substrate closer to exhibiting forms of learning that are harder to dismiss as purely reflexive. That does not mean organoids would become conscious, but it does mean the field would benefit from defining which cognitive markers it considers relevant and how it will monitor them, before the substrate becomes more capable.

The bottom line

Established from the source: the paper presents a coherent, three-level theory of neocortical learning as temporal-derivative error-driven predictive learning, supported by one recent in vitro plasticity experiment and by existing corticothalamic anatomy. The theory is not proven; it is a strong working hypothesis. For organoid intelligence, the calibrated conclusion is that training protocols may need to move beyond reward or correlation toward structured prediction-outcome sequences, but that the required circuitry may be partially absent in current organoids. The opportunity would be confirmed by showing that prediction-outcome stimulation trains an organoid or culture to produce a target activity pattern more reliably than unstructured stimulation; the threat would be confirmed by showing that the absence of thalamus-like driver inputs makes such training ineffective.