NeoHebbian Synapses: Mechanisms & Applications

• NeoHebbian synapses are defined as a three-factor learning mechanism that extends classical Hebbian rules with modulatory signals enabling temporal credit assignment.
• They incorporate eligibility traces to bridge rapid neural activity and slower behavioral timescales, supporting robust reinforcement learning and adaptive network reconfiguration.
• This mechanism underpins energy-efficient neuromorphic hardware implementations and inspires novel algorithms that achieve biological plausibility with computational efficiency.

NeoHebbian synapses are a fundamental construct in contemporary neuroscience and neuromorphic engineering, generalizing classical Hebbian learning to incorporate a third modulatory factor that links local synaptic dynamics to behavioral timescales, credit assignment, and global reward signals. Distinguished by their three-factor update rule, neoHebbian mechanisms underlie a broad class of biological and artificial learning systems, from reinforcement learning models in spiking networks to energy-efficient implementations in ReRAM-based neuromorphic hardware. NeoHebbian synapses enable powerful and biologically plausible online learning not only through activity correlation, but via eligibility traces and modulatory gating that support rapid adaptation, temporal credit assignment, and efficient problem solving (Vladar et al., 2015, Bartlett et al., 2019, Gerstner et al., 2018, Terres-Escudero et al., 2024, Pande et al., 2024).

1. NeoHebbian Synapses: Definitions and Distinctions

Classical Hebbian synaptic plasticity is encapsulated by the principle "neurons that fire together wire together," formalized as a weight update depending directly on the product of pre- and postsynaptic activity, e.g., $\Delta w_{ij} = \eta\,x_j\,y_i$. However, this form is restrictive, assuming that learning and correlation coincide temporally and lacking a mechanism for delayed, context-dependent updates. In contrast, neoHebbian synapses extend this rule to incorporate a third, global or modulatory signal $M(t)$, yielding a prototypical three-factor rule:

$\frac{d}{dt}\,w_{ij}(t) = e_{ij}(t)\,M(t)$

where the eligibility trace $e_{ij}(t)$ stores a transient record of recent Hebbian co-activation, and $M(t)$ represents a reward, punishment, surprise, novelty, or another neuromodulatory signal (Gerstner et al., 2018, Bartlett et al., 2019).

The defining features of neoHebbian synapses are:

• Separation of correlation and conversion: Local coincidences (Hebbian pairing) are flagged via $e_{ij}(t)$, which are later converted into lasting weight changes only if $M(t)$ is nonzero during the eligibility window.
• Temporal bridging: This architecture enables learning from delayed cues, overcoming the rapid decay of neuronal action potentials and allowing behavioral timescale credit assignment (Gerstner et al., 2018).
• Normative foundations: In formal settings, three-factor rules are connected to stochastic gradient ascent on global objectives, as in reinforcement learning and policy-gradient models (Bartlett et al., 2019).

2. Mathematical Formalism and Experimental Validation

The neoHebbian rule is split into eligibility trace dynamics and weight update (conversion by third factor):

Eligibility trace:

$$\frac{d}{dt}\,e_{ij}(t) = \eta\,x_j(t)\,g(y_i(t)) - \frac{e_{ij}(t)}{\tau_e}$$

where $x_j(t)$ is presynaptic input (e.g., glutamate release), $g(y_i(t))$ is (possibly nonlinear) postsynaptic response (voltage, calcium, spike rate), and $\tau_e$ is the eligibility trace decay constant (Gerstner et al., 2018).

Weight update:

$\frac{d}{dt}\,w_{ij}(t) = e_{ij}(t)\,M(t)$

with $M(t)$ a broadcast neuromodulator (e.g., dopamine, acetylcholine) or global error signal.

Experimental Support: Precise optogenetic and electrophysiological protocols have demonstrated eligibility traces on timescales from hundreds of milliseconds (striatal spines; dopamine required within 1 s of induction) up to several seconds (neocortical LTP/LTD gated by noradrenaline/serotonin within 2–10 s), and even minutes (hippocampal CA1 STDP modulated by dopamine or acetylcholine). Biophysical candidates for $e_{ij}(t)$ include CaMKII, cAMP, and protein kinase A, which persist on appropriate timescales and mediate the necessary conversions (Gerstner et al., 2018).

Component	Classical Hebb	NeoHebbian (Three-Factor)
Weight Update	$\eta x_j y_i$	$e_{ij}(t) M(t)$
Temporal Credit Assignment	None	Eligibility trace bridges ms–min scales
Modulatory Signal	None	Dopamine, reward, surprise, error

3. Computational and Evolutionary Models

Theoretical models have extended neoHebbian principles to the level of population dynamics and structural plasticity (Vladar et al., 2015). In such frameworks:

• Hebbian weights control mutation (“switching”) probabilities, e.g., neuron $i$ switching with probability $M_i = 1/(1+e^{Y_i})$, $Y_i = \sum_j \phi_{ij} X_j$.
• The master equation combines selection (amplification of fitter patterns) and mutation (variability driven by synaptic weights):

$$\frac{d\rho_i}{dt} = \rho_i (1-\rho_i) \left(\frac{\partial}{\partial \rho_i} \ln \bar{W}\right) + M_i (2\rho_i - 1)$$

where $\rho_i$ is the fraction of active groups for neuron $i$, and $\bar{W}$ the average fitness.

• Structural Synaptic Plasticity (SSP) operates on a slower timescale, adding or removing synapses based on co-activity probability and synaptic mutual information, with deletion favoring weak or noisy (uninformative) links. Network topology adapts via a Metropolis acceptance criterion, driving systems to optimal task-specific connectivity while penalizing excessive wiring cost.

This synergy of weight (Hebbian/Oja), parallel selection, and structural plasticity leads to rapid convergence and adaptive circuit reconfiguration—a process that is Darwinian in spirit and enables escape from local optima through structural macromutations (Vladar et al., 2015).

4. NeoHebbian Rules in Reinforcement Learning and Algorithmic Synapses

NeoHebbian synapses can be derived from normative, gradient-based approaches to reinforcement learning. By treating each synapse as an agent and applying policy-gradient principles:

• The eligibility trace is accumulated locally as

$$z_{j,t+1} = \beta z_{j,t} + (u_t - \sigma(v_t)) u_{t-1}^j$$

where $u_t$ and $u_{t-1}^j$ are post- and pre-synaptic spikes, respectively.

• Weight updates occur only upon arrival of the global reward:

$w_{j,t+1} = w_{j,t} + \gamma r_{t+1} z_{j,t+1}$

• This algorithm performs unbiased stochastic gradient ascent on expected long-term reward under mixing and slow learning (Bartlett et al., 2019).

NeoHebbian synapses are thus a formal mechanism for distributed credit assignment in agent-based networks, supporting parallel and fully local learning without backpropagation.

5. Relationship to Modern Learning Algorithms and Neuromorphic Hardware

Recent work establishes a formal equivalence between neoHebbian updates and the learning rules underlying the Forward-Forward Algorithm (FFA), a fully local alternative to backpropagation. When using a squared-Euclidean "goodness function", FFA gradient updates reduce to a three-factor rule:

$\Delta w_{ij} = M(t)\, \ell_j\, x_i$

where $x_i$ is presynaptic input, $\ell_j$ postsynaptic trace, and $M(t)$ encodes a local contrastive or reward-based signal. This mapping holds for both analog and spiking network variants (Terres-Escudero et al., 2024). Empirically, FFA-trained analog and Hebbian spiking networks achieve similar accuracy and representational sparsity, suggesting that three-factor learning can close the gap in energy efficiency and locality between artificial and biological computation.

Physically, implementation of neoHebbian synapses in neuromorphic hardware is feasible via ReRAM devices with integrated nanoheaters: the coupling weight is stored as conductance, while an "eligibility trace" is encoded in local temperature, modulated in real-time by voltage pulses reflective of pre- and post-synaptic activity (and global modulatory signal). This architecture yields robust learning performance in tasks such as RL maze navigation and temporal signal classification, is resilient to device variability and thermal crosstalk, and achieves energy per synapse update on the order of 5 pJ (Pande et al., 2024).

6. Biological and Hardware Constraints, Limitations, and Perspectives

The biochemical underpinnings of eligibility traces have been linked to spine-localized molecules including CaMKII, cAMP/PKA, and receptor-activated cascades. Experimental eligibility trace windows span milliseconds to minutes, enabling temporal credit assignment congruent with behavioral contingencies (Gerstner et al., 2018). NeoHebbian learning is distinguished by:

• Locality: All computations except the modulatory third factor are synapse-local.
• Temporal Dynamics: Trace decay constants are matched to behavioral or task time constants; reward signals often arise from neuromodulatory bursts (dopamine, noradrenaline, serotonin, acetylcholine).
• Hardware Instantiation: Two distinct state variables—weight (persistent, e.g., ReRAM conductance) and eligibility (transient, e.g., temperature)—allow strict separation of fast online update and slower consolidation, aligning with the two-stage logic of three-factor learning (Pande et al., 2024).

Limitations of neoHebbian designs include sensitivity to decoherence of eligibility signals, the need for precise decay constants, and, in hardware, thermal or device variability. Nevertheless, empirical and system-level simulations demonstrate high robustness and minimal accuracy degradation under these non-idealities.

7. Significance, Impact, and Ongoing Challenges

NeoHebbian synapses unify theories of synaptic plasticity, reinforcement learning, and structural circuit evolution. Their three-factor architecture solves the credit assignment problem over behavioral timescales, supports adaptation through both weight and connectivity reconfiguration, and underlies the rapid, efficient problem solving observed in neural and artificial networks. The convergence of biological experiment, computational neuroscience, and neuromorphic hardware design affirms the central place of neoHebbian mechanisms in advancing both our theoretical understanding of learning and the development of energy-efficient, scalable artificial cognitive systems (Vladar et al., 2015, Gerstner et al., 2018, Pande et al., 2024). Remaining challenges include delineating the molecular identity and dynamics of eligibility traces, optimizing hardware instantiations for synaptic architecture, and extending these principles to continuous, hierarchical, and multi-agent architectures under real-world constraints.