New Biologically Constrained Learning Rule

Science General Neuroscience
1

A generalized mathematical framework for the calcium control hypothesis describes weight‐dependent synaptic plasticity

Toviah Moldwin, Li Shay Azran, Idan Segev (2025)
https://doi.org/10.1007/s10827-025-00894-6

Abstract

The brain modifies synaptic strengths to store new information via long-term potentiation (LTP) and long-term depression (LTD). Evidence has mounted that long-term synaptic plasticity is controlled via concentrations of calcium ­([Ca2+]) in postsynaptic dendritic spines. Several mathematical models describe this phenomenon, including those of Shouval, Bear, and Cooper (SBC) (Shouval et al., 2002, 2010) and Graupner and Brunel (GB) (Graupner & Brunel, 2012). Here we suggest a generalized version of the SBC and GB models, the fixed point – learning rate (FPLR) framework, where the synaptic ­[Ca2+] specifies a fixed point toward which the synaptic weight approaches asymptotically at a [­ Ca2+]-dependent rate. The FPLR framework offers a straightforward phenomenological interpretation of calcium-based plasticity: the calcium concentration tells the synaptic weight where it is going and how quickly it goes there. The FPLR framework can flexibly incorporate various experimental findings, including the existence of multiple regions of ­[Ca2+] where no plasticity occurs, or plasticity observed experimentally in cerebellar Purkinje cells, where the directionality of calcium-based synaptic changes is reversed relative to cortical and hippocampal neurons. We also suggest a modeling approach that captures the dependency of late-phase plasticity stabilization on protein synthesis. We demonstrate that due to the asymptotic nature of synaptic changes in the FPLR rule, the plastic changes induced by frequency- and spike-timing-dependent plasticity protocols are weight-dependent. Finally, we show how the FPLR framework can explain the weight-dependence observed in behavioral time scale plasticity (BTSP).

Significance: This work bridges the gap between biology and computers. The authors study biological learning rules in order to formulate a learning rule computer models to use.

3 Likes
2

Hey, I have a few thoughts on this.

Neuromodulatory Control over FPLR Landscapes

The FPLR framework elegantly uses local postsynaptic [Ca^{2+}] concentrations to determine both the target state (fixed point) and the rate of approach. However, local calcium dynamics alone can be rigid under changing task demands.

—-

Global neuromodulatory signals should act as meta-controllers that dynamically warp the [Ca^{2+}] fixed-point curves:

Acetylcholine (ACh): Can temporarily steepen the learning rate curve at lower calcium thresholds, forcing rapid encoding of novel patterns.

GABA (Inhibition): Shifts the zero-plasticity zones, effectively raising the calcium threshold required to induce LTP and preserving system sparsity.

Dopamine: Acts as a gatekeeper for late-phase stabilization, determining whether a calcium-induced target state is worth persisting long-term.

—-

Offline Consolidation vs. Continuous Late-Phase Mechanics

Modeling late-phase plasticity via local protein synthesis kinetics is biologically accurate, but computationally expensive to scale in large AI systems.

Frame late-phase stabilization as a two-phase operational cycle (Online Execution vs. Offline Sleep-Replay). During active phases, FPLR updates are held in a volatile, fast-adapting state. During offline sleep-replay phases, global metaplastizity signals evaluate state lineage and trigger “protein synthesis equivalents” (permanent weight or path updates) while pruning verifiably low-value connections.

—-

Preventing Asymptotic Traps with Lineage Observers

Because FPLR relies on asymptotic convergence toward fixed points, networks using this rule are susceptible to state saturation or topological trapping (where local updates oscillate between nearby fixed points without producing novel representations).

Incorporate a detached meta-level observer (such as a core state lineage observer) to monitor state entropy over time. If the rate of weight change drops below a threshold without reaching a meaningful target state (a “plasticity stall”), the system can inject a global reset signal (a simulated noradrenaline burst) to temporarily displace the calcium fixed points and break the loop.

—-

Translating FPLR to Discrete and Relational Architectures

While the authors formulate FPLR for continuous weight matrices, the fundamental logic (“where it is going and how quickly it goes there”) maps perfectly to discrete, set-theoretic operations.

The FPLR concept can be extended beyond floating-point matrix multiplications into relational and graph-based models. In discrete systems, the “fixed point” represents a target subgraph or set intersection, and the “learning rate” corresponds to the threshold for opening or closing dynamic relational paths.