Link to the code: brain-emulation GitHub repository

The Wrong Learning Rule: How BTSP Rewrites What Brain Emulation Must Replicate

For three decades, a particular learning rule has dominated computational neuroscience: spike-timing-dependent plasticity, or STDP. The rule is elegant. If neuron A fires just before neuron B, the synapse connecting them strengthens. If B fires before A, the synapse weakens. The timing window is milliseconds. The rule connects Hebbian learning theory to measurable biology and has been implemented in spiking neural network simulators, neuromorphic chips, and whole brain emulation frameworks worldwide.

A 2026 paper in Nature Neuroscience now presents evidence that this rule, while real, is not the primary mechanism driving memory formation in the hippocampus, the brain structure most directly responsible for encoding new episodic memories. Behavioral timescale synaptic plasticity, or BTSP, operates on a fundamentally different timescale and a fundamentally different mechanism. The implications for whole brain emulation are direct: if WBE frameworks have modeled hippocampal memory encoding with STDP, they have modeled the wrong rule, and the emulated hippocampus would not form or retrieve memories the way a biological hippocampus does.

What STDP Is and Why It Was Assumed

Spike-timing-dependent plasticity was characterized experimentally in the late 1990s and quickly became the dominant framework for understanding how synaptic weights change during learning. The mechanism is local and temporally precise: the synapse monitors the relative timing of pre- and postsynaptic spikes and adjusts its strength accordingly. This mechanism is measurable, computationally tractable, and consistent with Hebb’s original 1949 postulate that neurons that fire together wire together.

STDP has been confirmed as a real phenomenon in multiple brain regions and in multiple species. The question raised by the 2026 research is not whether STDP exists but whether it is the mechanism driving hippocampal place field formation and episodic memory encoding, the functions most relevant to preserving a person’s memories in a brain emulation.

BTSP: What the New Data Shows

Behavioral timescale synaptic plasticity operates on a different timescale entirely. Where STDP responds to pre-post spike timing differences of milliseconds, BTSP operates across seconds, the timescale of behavior rather than the timescale of individual spikes.

The core experimental finding involves dendritic plateau potentials in hippocampal CA1 pyramidal neurons. These are sustained depolarizations in the apical dendrite, lasting hundreds of milliseconds, generated by coincident excitatory inputs. The 2026 Nature Neuroscience paper shows that a single dendritic plateau potential is sufficient to produce bidirectional synaptic weight changes: inputs active just before the plateau potential are strengthened; inputs active just after are weakened.

This is a fundamentally different computational event than a spike. Plateau potentials integrate inputs over a behavioral timescale. They do not require millisecond-precision spike coincidence. They link synaptic changes to behavioral episodes rather than to individual action potentials.

The research documents how hippocampal place fields, the selective firing of specific neurons when an animal occupies a particular location in space, shift and form through BTSP rather than STDP. Place fields are the most studied example of episodic memory encoding in the hippocampus. If their formation depends on BTSP, then the mechanism generating location-specific memories is a plateau potential-driven process operating on second-long timescales, not a spike coincidence detector operating on millisecond timescales.

What This Means for Brain Emulation

The practical consequence depends on how existing WBE models implement hippocampal circuits. The most common approach uses STDP as the synaptic learning rule in hippocampal simulations, partly because STDP is well-characterized and mathematically tractable, and partly because the evidence for alternative rules was not as strong when those frameworks were designed.

If BTSP is the dominant mechanism for place field formation and episodic memory encoding, a simulated hippocampus running STDP will not form place fields correctly, will not encode new episodic memories through the right computational process, and will not retrieve those memories with the fidelity the biological system achieves.

The amyloid protein discovery from Stowers Institute earlier in 2026 showed that neurons deliberately form amyloid structures to store long-lasting memories, a finding that already added a molecular layer to what WBE must preserve beyond synaptic weights. BTSP adds a second layer: the temporal structure of the learning rule itself. Synaptic weights stored in an emulated brain may have been established by the wrong mechanism, meaning they encode a distorted version of the original experience.

The Dendritic Computation Problem

The BTSP finding connects to a broader problem in computational neuroscience that has become more urgent in the context of brain emulation: neurons are not point integrators. The standard leaky integrate-and-fire neuron model, and many spiking neuron models derived from it, treat the neuron as a single compartment that sums its inputs and fires when a threshold is reached.

This simplification has been known to be incomplete for decades. Dendrites perform local computations. Apical and basal dendritic compartments receive different inputs and respond to different signals. The discovery that motor cortex dendrites follow compartment-specific plasticity rules is a parallel finding to BTSP: both point to the same conclusion that single-compartment neuron models miss computations that happen below the level of the soma.

BTSP specifically involves the apical dendritic compartment of CA1 pyramidal neurons. The plateau potential that drives BTSP occurs in the apical dendrite, not at the soma. A neuron model without an apical dendritic compartment cannot generate plateau potentials and therefore cannot implement BTSP. The Izhikevich model and other reduced-form neuron models commonly used in large-scale brain simulations lack this compartment.

Implications for Large-Scale Simulation Frameworks

The Allen Institute mouse cortex simulation used biophysically detailed neuron models for its 9 million neuron simulation, which included some dendritic structure. The JUPITER exascale SNN simulation used simplified spiking neurons at 20 billion neuron scale. Both represent trade-offs between biological fidelity and computational tractability.

The BTSP finding does not invalidate these simulations. They were not designed to model episodic memory encoding. But it sets a constraint on what a simulation must implement before it can claim to emulate hippocampal memory function. A brain emulation that does not support plateau potential-driven BTSP in its hippocampal circuits is, by this evidence, not emulating hippocampal memory.

Implementing BTSP in large-scale simulations requires two-compartment neuron models with distinct apical and basal dynamics, and synaptic update rules that respond to dendritic plateau events rather than spike coincidences. The computational cost increases, but the alternative is a simulation that cannot faithfully encode and retrieve episodic memories.

What Must Be Preserved

The building brains on a computer framework outlined three non-negotiable requirements for whole brain emulation: structural fidelity, functional replication, and dynamic adaptability. BTSP makes the second requirement more demanding than previously understood.

Functional replication of the hippocampus requires not just the correct connectivity between neurons but the correct learning rule operating in each circuit. A structurally accurate connectome of the hippocampus, connected by synapses with the correct initial weights, but governed by STDP rather than BTSP, would not replicate hippocampal function. It would store and retrieve patterns differently, encoding memories through a mechanism the biological system does not use.

Technology Readiness Level: Not applicable (basic neuroscience). The findings are validated in mouse hippocampus in vivo. Human hippocampal BTSP has not been directly measured, but the cellular mechanisms are evolutionarily conserved across mammals.

Official Sources