Papers
Topics
Authors
Recent
Search
2000 character limit reached

Error-Modulated Hebbian Updates

Updated 5 July 2026
  • Error-modulated Hebbian updates are synaptic plasticity rules that factorize learning into local presynaptic-postsynaptic correlations modulated by external error or reward signals.
  • They derive modulators from various objectives—such as contrastive goodness, kernel similarity, and stability measures—allowing local, task-sensitive weight adjustments.
  • Applications in deep networks, spiking systems, and unsupervised representation learning show improvements in performance and stability across diverse learning scenarios.

Error-modulated Hebbian updates are synaptic plasticity rules in which a local Hebbian correlation is multiplied by a third factor carrying error, reward, teaching, or stability information. In canonical form, a synapse i→ji \to j obeys Δwij=η m yjxi\Delta w_{ij}=\eta\,m\,y_j x_i, or Δwij=η m gjxi\Delta w_{ij}=\eta\,m\,g_j x_i when the postsynaptic term includes a local derivative or surrogate factor. This three-factor structure appears across contemporary local-learning research: in Forward-Forward learning with squared-activation goodness, in policy-gradient plasticity for spiking neurons, in layer-wise Information Bottleneck training with working-memory-dependent modulators, in quantum probability-flow objectives, and in microcircuit models where dis-inhibitory inhibition carries the effective error signal (Terres-Escudero et al., 2024, Bartlett et al., 2019, Daruwalla et al., 2021, Ohzeki, 1 Jun 2026, Rossbroich et al., 2023).

1. Canonical three-factor structure

The defining property of an error-modulated Hebbian rule is the factorization of plasticity into a presynaptic term, a postsynaptic term, and a modulatory signal. In its simplest form,

Δwij=η m yjxi,\Delta w_{ij}=\eta\,m\,y_j x_i,

where xix_i is presynaptic activity, yjy_j is postsynaptic activity, mm is a modulatory signal, and η>0\eta>0 is a learning rate. A common variant replaces yjy_j by a local postsynaptic drive gjg_j, such as yjf′(uj)y_j f'(u_j), yielding

Δwij=η m gjxi.\Delta w_{ij}=\eta\,m\,g_j x_i.

The same structure is written in continuous time as

∂wij∂t≐Δtwij=M(t) g(yj) xi,\frac{\partial w_{ij}}{\partial t}\doteq \Delta_t w_{ij}=M(t)\,g(y_j)\,x_i,

with M(t)M(t) as the third factor. In spiking formulations derived from policy gradients, the local Hebbian-like factor often becomes xj(t−1)[si(t)−λi(t)]x_j(t-1)[s_i(t)-\lambda_i(t)], where si(t)s_i(t) is the realized spike, λi(t)\lambda_i(t) is its stochastic firing probability, and the centered term si(t)−λi(t)s_i(t)-\lambda_i(t) plays the role of a postsynaptic surprise variable (Terres-Escudero et al., 2024, Bartlett et al., 2019).

This structure differs from both pure Hebbian learning and backpropagation. Pure Hebbian plasticity uses only two factors,

Δwij∝yjxi,\Delta w_{ij}\propto y_j x_i,

and therefore lacks an explicit task-dependent gate. Backpropagation, by contrast, computes updates from gradients of a global loss propagated through a reverse pathway. Error-modulated Hebbian rules occupy the intermediate regime: they preserve synapse-local update algebra, but the direction and magnitude of plasticity are controlled by an additional signal encoding reward, error, target mismatch, stability margin, or validation.

2. Objective-derived modulators

In many recent formulations, the third factor is not introduced heuristically; it is derived from a local objective. In the Forward-Forward Algorithm (FFA), a layer with pre-activation u=Wx+bu=Wx+b, activation h=f(u)h=f(u), and goodness

G(h)=∥h∥2=∑j=1nhj2G(h)=\|h\|^2=\sum_{j=1}^{n} h_j^2

has local goodness gradient

∂G∂wij=2hjf′(uj)xi,∂G∂bj=2hjf′(uj).\frac{\partial G}{\partial w_{ij}}=2h_j f'(u_j)x_i,\qquad \frac{\partial G}{\partial b_j}=2h_j f'(u_j).

Positive examples increase goodness and negative examples decrease it, so the local update becomes

Δwij=η m 2hjf′(uj)xi,\Delta w_{ij}=\eta\,m\,2h_j f'(u_j)x_i,

where mm is generated by the goodness-to-probability mapping. The paper gives two explicit mappings,

Pσ(G⊕;θ,α)=sigmoid⁡(α(G⊕−θ)),PSym=G⊕G⊕+G⊖,P_\sigma(G_\oplus;\theta,\alpha)=\operatorname{sigmoid}(\alpha(G_\oplus-\theta)), \qquad P_{\text{Sym}}=\frac{G_\oplus}{G_\oplus+G_\ominus},

and the corresponding derivative factors, such as 1−Pσ1-P_\sigma or (1−PSym)/G⊕(1-P_{\text{Sym}})/G_\oplus, become explicit modulators multiplying the local correlation (Terres-Escudero et al., 2024).

A distinct layer-local derivation appears in HSIC-based Information Bottleneck learning. There the update decomposes as

Δ[Wℓ]ij∝βij ξi,\Delta [W^\ell]_{ij}\propto \beta_{ij}\,\xi_i,

with local Hebbian term

βij=(1−([z0ℓ]i)2)[z0ℓ−1]j\beta_{ij}=(1-([z_0^\ell]_i)^2)[z_0^{\ell-1}]_j

and modulatory term

ξi=∑p=0−(N−1)[kˉ(x0,xp)−γkˉ(y0,yp)]αˉi(zpℓ).\xi_i=\sum_{p=0}^{-(N-1)}\left[\bar k(x_0,x_p)-\gamma \bar k(y_0,y_p)\right]\bar\alpha_i(z_p^\ell).

Here the third factor is neither scalar reward nor explicit backpropagated error; it is a batch-statistic, layer-wise signal built from centered kernel similarities over an effective working-memory window (Daruwalla et al., 2021).

A third construction derives the modulator from a physics-based stability objective. In quantum probability-flow learning, one-spin-flip leakage channels define local gaps

ΔEk(x)=2xk(hk+∑l≠kJklxl),\Delta E_k(\mathbf{x})=2x_k\Bigl(h_k+\sum_{l\ne k}J_{kl}x_l\Bigr),

and the measured survival loss produces instability weights

rk(x)=Γ2e−βΔEk(x)1+Γ2∑ℓe−βΔEℓ(x).r_k(\mathbf{x})=\frac{\Gamma^2 e^{-\beta \Delta E_k(\mathbf{x})}}{1+\Gamma^2\sum_\ell e^{-\beta \Delta E_\ell(\mathbf{x})}}.

The resulting coupling update is

ΔJij∝2βM∑μ=1M[ri(xμ)+rj(xμ)] xμixμj,\Delta J_{ij}\propto \frac{2\beta}{M}\sum_{\mu=1}^{M}[r_i(\mathbf{x}_\mu)+r_j(\mathbf{x}_\mu)]\,x_{\mu i}x_{\mu j},

that is, Hebbian correlation xixjx_i x_j gated by a softmax-like instability signal. In the high-temperature limit β→0\beta\to 0, this reduces, up to a constant learning rate, to the classical Hebbian rule; in the low-temperature limit, learning concentrates on the weakest margin (Ohzeki, 1 Jun 2026).

Framework Local Hebbian factor Modulatory term
FFA 2hjf′(uj)xi2h_j f'(u_j)x_i or ℓjxi\ell_j x_i Derived from Υ(G⊕,G⊖)\Upsilon(G_\oplus,G_\ominus)
HSIC Information Bottleneck βij\beta_{ij} ξi\xi_i from centered kernel statistics
Quantum probability flow xixjx_i x_j ri+rjr_i+r_j or real-time power-law gate

These examples show that the same algebraic form can arise from contrastive goodness maximization, kernelized compression objectives, or local stability margins. This suggests that the term “error” in this literature denotes a broader family of modulators than the conventional supervised residual.

3. Eligibility traces and temporal credit assignment

When rewards or supervised errors are delayed, error-modulated Hebbian rules are typically implemented through eligibility traces. In Bartlett and Baxter’s spiking reinforcement-learning formulation, the network maximizes the long-term average reward

J=η=lim⁡T→∞1TE[∑t=1Tr(t)].J=\eta=\lim_{T\to\infty}\frac{1}{T}\mathbb{E}\Big[\sum_{t=1}^T r(t)\Big].

For a Bernoulli spiking neuron with λi(t)=σ(vi(t))\lambda_i(t)=\sigma(v_i(t)), the local log-likelihood derivative is

∂∂wijlog⁡p(si(t)∣vi(t))=xj(t−1) [si(t)−λi(t)].\frac{\partial}{\partial w_{ij}}\log p(s_i(t)\mid v_i(t)) = x_j(t-1)\,[s_i(t)-\lambda_i(t)].

The synapse accumulates this in a leaky eligibility

eij(t+1)=βeij(t)+xj(t−1)gi(t),gi(t)=si(t)−λi(t),e_{ij}(t+1)=\beta e_{ij}(t)+x_j(t-1)g_i(t),\qquad g_i(t)=s_i(t)-\lambda_i(t),

and plasticity is gated by the broadcast reward or reward-minus-baseline,

Δwij=η ∑t(r(t)−b) eij(t).\Delta w_{ij}=\eta\,\sum_t (r(t)-b)\,e_{ij}(t).

The same paper shows that, under factorized joint policies and standard ergodicity and small-step-size assumptions, each neuron can update its own parameters using only local eligibility and the shared reward, yielding locally optimal performance without explicit inter-neuronal communication (Bartlett et al., 2019).

MOHQA uses a similar temporal mechanism, but with sparse STDP-inspired correlations. Its eligibility dynamics are

E˙(t)=−E(t)TE+O(t),\dot E(t)=-\frac{E(t)}{T_E}+O(t),

where O(t)∈{−1,0,+1}O(t)\in\{-1,0,+1\} depends on top-percentile or bottom-percentile pre-post correlations between DQN-derived features and a one-hot MOHN action head. The synaptic update is

ΔW(t)=(r(t)+b) E(t),\Delta W(t)=(r(t)+b)\,E(t),

followed by clipping to [−1,1][-1,1]. This architecture was designed for confounding POMDPs in which TD errors are inaccurate; the trace bridges temporal delays between salient feature–action events and sparse reward (Ladosz et al., 2019).

A more explicit gradient-free temporal rule appears in noise-based reward-modulated learning. There, noisy neurons induce a directional-derivative factor

ρ(t)=log⁡πW~(at∣St)−log⁡πW(at∣St),\rho(t)=\log \pi_{\tilde W}(a_t\mid S_t)-\log \pi_W(a_t\mid S_t),

and the inter-reward eligibility becomes

eijl(τ)=∑t=τ0τξˉjl(t) pt x~il−1(t),pt=ρ(t).e^l_{ij}(\tau)=\sum_{t=\tau_0}^{\tau}\bar \xi^l_j(t)\,p_t\,\tilde x^{l-1}_i(t), \qquad p_t=\rho(t).

At reward time τ\tau,

Δwijl(τ)=η δrτ eijl(τ),\Delta w^l_{ij}(\tau)=\eta\,\delta r_\tau\,e^l_{ij}(\tau),

with δrt=rt−rˉt\delta r_t=r_t-\bar r_t as reward prediction error relative to a running average. The update remains synaptically local because it uses presynaptic activity, postsynaptic perturbation, and two global scalars, ptp_t and δrt\delta r_t (Fernández et al., 31 Mar 2025).

Online supervised spiking decoders extend the same logic to dense frame-wise errors. In the BCI setting, the instantaneous three-factor update is

ΔWhebb(ℓ)(t)=(e~t(ℓ)⊙dt(ℓ)) (pret(ℓ))⊤,\Delta W^{(\ell)}_{\mathrm{hebb}}(t)= \big(\tilde{\mathbf e}^{(\ell)}_t\odot d^{(\ell)}_t\big)\, \big(\mathrm{pre}^{(\ell)}_t\big)^\top,

with dual-timescale traces

Efast(ℓ)(t)=λfastEfast(ℓ)(t−1)+ΔWhebb(ℓ)(t),E^{(\ell)}_{\mathrm{fast}}(t)=\lambda_{\mathrm{fast}}E^{(\ell)}_{\mathrm{fast}}(t-1)+\Delta W^{(\ell)}_{\mathrm{hebb}}(t),

Eslow(ℓ)(t)=λslowEslow(ℓ)(t−1)+ΔWhebb(ℓ)(t).E^{(\ell)}_{\mathrm{slow}}(t)=\lambda_{\mathrm{slow}}E^{(\ell)}_{\mathrm{slow}}(t-1)+\Delta W^{(\ell)}_{\mathrm{hebb}}(t).

This formulation avoids backpropagation through time, uses O(1)O(1) memory in sequence length, and couples rapid adaptation to slower consolidation (Nallani et al., 17 Sep 2025).

4. Synaptic specificity, circuit realization, and biological interpretation

A central biological issue is whether the third factor is merely global or whether it can be synapse-specific enough to preserve higher-order learning. The proofreading account of complex Hebbian learning argues that higher-order-correlation learning fails abruptly when synaptic updates are insufficiently connection-specific. In a one-unit ICA model with crosstalk matrix EE, the nonlinear Hebbian rule

Δw=±k E x f(y)\Delta \mathbf w=\pm k\,E\,\mathbf x\,f(y)

collapses above a modest threshold to a second-order, PCA-like solution. In a three-input example with one Laplacian and two Gaussian sources, the empirical threshold was bc≈0.0425b_c\approx 0.0425 with Q≈0.88Q\approx 0.88, and across four mixing matrices the average was 0.04±0.030.04\pm 0.03. To prevent this “error catastrophe,” the paper proposes a neocortical proofreading mechanism in which a validation factor gates plasticity: Δwi=η xi y Ei(t),\Delta w_i=\eta\,x_i\,y\,E_i(t), or in STDP form,

Δwi=η∑(tpre,tpost)K(tpost−tpre) Ei(tpre,tpost).\Delta w_i=\eta \sum_{(t_{\text{pre}},t_{\text{post}})} K(t_{\text{post}}-t_{\text{pre}})\,E_i(t_{\text{pre}},t_{\text{post}}).

Here EiE_i is generated by a circuit involving thalamic relay cells, layer-4 spiny stellate cells, and layer-6 corticothalamic neurons, so that only independently corroborated pre-post coincidences are expressed as weight changes (Cox et al., 2010).

A different circuit-level solution embeds the error in inhibition itself. In the dis-inhibitory control framework, each local unit contains an excitatory neuron and an inhibitory interneuron with dynamics

τEddtuiE(t)=−uiE(t)+Wiri−1E(t)−riI(t),\tau_E \frac{d}{dt}u_i^E(t)=-u_i^E(t)+W_i r_{i-1}^E(t)-r_i^I(t),

τIddtuiI(t)=−uiI(t)+riE(t)−Qic(t).\tau_I \frac{d}{dt}u_i^I(t)=-u_i^I(t)+r_i^E(t)-Q_i c(t).

At steady state, the local error decoded at the excitatory cell is

ei=(Qic)i=riE−uiI,e_i=(Q_i c)_i=r_i^E-u_i^I,

and the exact learning rule becomes

τWddtWi=[(riE−uiI)⊙ϕ′(uiE)](ri−1E)⊤.\tau_W \frac{d}{dt}W_i= \big[(r_i^E-u_i^I)\odot \phi'(u_i^E)\big](r_{i-1}^E)^\top.

Its single-synapse form,

Δwij=ϕ′(uiE)(riE−uiI)rjE,\Delta w_{ij}= \phi'(u_i^E)(r_i^E-u_i^I)r_j^E,

is again a three-factor update. In the absence of inhibition it reduces to a postsynaptic-threshold rule,

Δw∝rpre(rpost−θ)ϕ′(rpost),\Delta w \propto r_{\text{pre}}(r_{\text{post}}-\theta)\phi'(r_{\text{post}}),

which the paper presents as consistent with in vitro phenomenological plasticity models (Rossbroich et al., 2023).

These two lines of work treat the third factor differently. Proofreading emphasizes selective approval that suppresses false updates caused by crosstalk, whereas dis-inhibitory control interprets the modulatory term as a locally decoded error current. Both reject the view that biologically plausible plasticity must be only correlation-based.

5. Realizations in deep networks, spiking systems, and unsupervised representation learning

Error-modulated Hebbian updates are now used in supervised, reinforcement, and unsupervised settings, with different granularities of the third factor: scalar reward, neuron-wise vector, synapse-wise sign, or block-local structural error.

Approach Setting Reported result
FFA / Hebbian FFA MNIST, single-layer, symmetric probability Analog FFA 95.10%; Hebbian/spiking FFA 92.72%; online Hebbian K=1K=1 94.36%
GHL ImageNet, ResNet-50 Top-1 73.14; Top-5 91.04
Online Hebbian BCI SNN Zenodo Indy, MC Maze Pearson R≥0.63R \ge 0.63 and R≥0.81R \ge 0.81; 28–35% memory reduction
SPHeRe CIFAR-10 / CIFAR-100 / Tiny-ImageNet 81.11% / 56.79% / 40.33%
MOHQA Hardest POMDPs and Malmo At least 33% improvement vs. baselines

FFA supplies one explicit bridge from local contrastive learning to neo-Hebbian updates. With squared Euclidean goodness, analog and spiking realizations produced similar accuracy and latent distributions on MNIST, with sparse latent activity vectors having Hoyer Index >0.95>0.95 and T-SNE separability index ≥0.95\ge 0.95. The same paper argues that the forward-only, local nature of the rule makes analog FFA directly relevant for neuromorphic deployment (Terres-Escudero et al., 2024).

In large-scale supervised vision, Global-guided Hebbian Learning (GHL) separates magnitude and direction. The local magnitude is the competitive Oja term

ΔwikSWTA=η uk (xi−ykwik),uk=eyk/τ∑leyl/τ,\Delta w_{ik}^{\text{SWTA}}=\eta\,u_k\,(x_i-y_k w_{ik}), \qquad u_k=\frac{e^{y_k/\tau}}{\sum_l e^{y_l/\tau}},

while direction is provided by the per-synapse sign of the backpropagated gradient,

Δwik=η sign⁡ ⁣(∂L∂wik) ∣uk(xi−ykwik)∣.\Delta w_{ik}= \eta\,\operatorname{sign}\!\Big(\frac{\partial L}{\partial w_{ik}}\Big)\, \big|u_k(x_i-y_k w_{ik})\big|.

The paper reports competitive results on CIFAR-10/100 and ImageNet, including ResNet-50 Top-1 $73.14$ and Top-5 $91.04$, and notes that the method maintained performance even for ResNet-1202 on CIFAR-10 (Hua et al., 29 Jan 2026).

Deep networks with asymmetric feedback provide another realization. There the feedforward and feedback weights, WW and RR, are separate, and both are updated locally: ΔWijl=η δilxjl−1,ΔRikl+1=η δkl+1xil.\Delta W^l_{ij}=\eta\,\delta^l_i x^{l-1}_j,\qquad \Delta R^{l+1}_{ik}=\eta\,\delta^{l+1}_k x^l_i. If RR is fixed random feedback, performance degrades quickly with depth; if RR is learned with the same local rule, performance remains close to ordinary backpropagation, and in the linear case the paper proves that updating the feedback weights accelerates convergence of the error to zero (Amit, 2018).

A complementary engineering route uses surrogate losses whose gradients are exactly Hebbian. With

L=−12y2,\mathcal L=-\frac12 y^2,

appropriate definitions of yy recover plain Hebbian, Grossberg’s instar, or Oja updates under autograd. A detached modulatory signal MM can be inserted as

L=−12∑k,u,vMk(u,v) yk(u,v)2,\mathcal L=-\frac12 \sum_{k,u,v} M_k(u,v)\,y_k(u,v)^2,

which yields a three-factor update Δw∝M y ∂y/∂w\Delta w \propto M\,y\,\partial y/\partial w. This makes error-modulated Hebbian learning compatible with modern convolutional frameworks (Miconi, 2021).

Unsupervised representation learning has also adopted local error modulation. SPHeRe defines

Lstruct=∥ZZ⊤−XX⊤∥F2,Lorth=∥Z⊤Z−I∥F2,L_{\text{struct}}=\|ZZ^\top-XX^\top\|_F^2, \qquad L_{\text{orth}}=\|Z^\top Z-I\|_F^2,

with the auxiliary structural-projection path producing a local modulatory error

rZ=4(KZ−KX)Z.r_Z=4(K_Z-K_X)Z.

The paper presents the resulting weight change as an error-modulated Hebbian rule driven by presynaptic activity, postsynaptic activity, and a block-local structural mismatch, and reports state-of-the-art results among unsupervised synaptic-plasticity approaches on CIFAR-10, CIFAR-100, and Tiny-ImageNet (Deng et al., 16 Oct 2025).

6. Limitations, misconceptions, and open problems

A recurring limitation is that each derivation is tied to a specific objective and signal geometry. The FFA equivalence to a neo-Hebbian rule depends on the squared Euclidean goodness G(h)=∥h∥2G(h)=\|h\|^2; other goodness functions may not admit the same clean factorization. HSIC-based Information Bottleneck learning is sensitive to effective batch size NN and kernel bandwidth σ\sigma; small NN severely degrades performance because the kernel estimates become poor. Quantum probability-flow rules assume a single-flip approximation and require calibration of β\beta, Γ\Gamma, and ε\varepsilon. Policy-gradient and noise-based rules inherit high-variance estimators, sensitivity to stochasticity, and dependence on eligibility timescales. SPHeRe replaces backpropagated task errors with local Gram-matrix mismatch, but this introduces O(B2)O(B^2) batch coupling and reported diminishing gains with depth (Terres-Escudero et al., 2024, Daruwalla et al., 2021, Ohzeki, 1 Jun 2026, Bartlett et al., 2019, Fernández et al., 31 Mar 2025, Deng et al., 16 Oct 2025).

A common misconception is that every error-modulated Hebbian method is fully backpropagation-free. GHL explicitly obtains sign⁡(∂L/∂wik)\operatorname{sign}(\partial L/\partial w_{ik}) through standard backpropagation; asymmetric-feedback learning still requires backward error activities, even though they travel through separate learned feedback weights; and autograd-based Hebbian CNNs rely on conventional differentiation to realize local rules. Conversely, some biologically motivated methods remain highly structured: proofreading requires synapse-specific validation to avoid catastrophic crosstalk, and dis-inhibitory control assumes a controller and feedback matrices aligned with the network Jacobian (Hua et al., 29 Jan 2026, Amit, 2018, Miconi, 2021, Cox et al., 2010, Rossbroich et al., 2023).

The present literature therefore supports a narrower and more technical reading of the term. Error-modulated Hebbian updates are not a single algorithm but a family of local plasticity decompositions in which pre-post correlation is gated by a third factor. Depending on the model, that factor may be a scalar reward, a vector of layer-local teaching signals, a per-synapse sign, a synapse-specific validation gate, a softmax over instability channels, or a structural mismatch computed in an auxiliary projection. This suggests that the main unifying principle is not any particular biological story or optimization method, but the algebraic separation between local correlation and modulatory control.

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Error-Modulated Hebbian Updates.