Contrastive Hebbian Learning

• Contrastive Hebbian Learning (CHL) is an energy-based algorithm that employs two phases—free and clamped—to update local synaptic weights based on activity differences.
• CHL approximates backpropagation through local, biologically plausible updates and has been extended to architectures including spiking, convolutional, and meta-learning networks.
• Empirical studies show that CHL and its variants achieve competitive performance with conventional gradient methods while ensuring hardware efficiency and strict locality in updates.

Contrastive Hebbian Learning (CHL) is an energy-based, biologically inspired learning algorithm for supervised and unsupervised neural networks, defined by phase-based local updates that contrast neural coactivation statistics between two network equilibrium states—one with output units free and one with targets imposed. CHL embodies a credit-assignment computation that closely approximates backpropagation in layered networks, yet enforces only local synaptic plasticity. Recent theoretical and experimental advances have extended CHL to diverse architectures, introduced single-phase and hardware-oriented variants, and demonstrated empirical performance comparable to conventional gradient methods across deep, convolutional, spiking, and meta-learning systems.

1. Core Principle: Energy-Based Two-Phase Learning

CHL operates by defining a global network energy function—typically of Hopfield or generalized energy-based model (EBM) form, such as

$$E(x; W) = -\tfrac{1}{2} \sum_{i, j} w_{ij} x_i x_j - \sum_i b_i x_i,$$

where $W$ are symmetric weights and $x$ gives the state of all neurons (Kubo et al., 2022, Millidge et al., 2022). The learning protocol consists of two main phases per training pattern:

1. Free (negative) phase: Inputs are clamped; outputs and hidden units evolve under network dynamics (gradient flow on $E$) to a free equilibrium $x^-$.
2. Clamped (positive) phase: Outputs are hard-clamped to target values, and the network re-equilibrates to $x^+$ (with hidden and output units' activities now coupled to the teaching signal).

The update rule is local and Hebbian-contrastive:

$$\Delta w_{ij} = \alpha \left(x_i^+ x_j^+ - x_i^- x_j^-\right)$$

with learning rate $\alpha$, presynaptic $x_j$ and postsynaptic $x_i$ activities sampled at the end of each phase (Kubo et al., 2022, Mesnard et al., 2016). This rule generalizes naturally to multilayer, recurrent, and convolutional networks.

2. Theoretical Relationship to Gradient Descent and Backpropagation

CHL can be formally interpreted as stochastic gradient descent (SGD) on a contrastive energy objective

$F(W) = E(x^+; W) - E(x^-; W),$

where the superscripts denote the clamped (positive) and free (negative) equilibria (Millidge et al., 2022). By the envelope theorem, the synaptic update implements

$$\Delta w_{ij} \propto -\frac{\partial [E(x^+)-E(x^-)]}{\partial w_{ij}},$$

which, under suitable energy and feedback scaling (in the infinitesimal inference limit), recovers the exact gradient computed by backpropagation (Millidge et al., 2022, Mesnard et al., 2016). For small nudging strengths, the difference between the two phases is proportional to the backpropagated error, ensuring convergence toward backprop minimum. CHL thus provides a biologically plausible, fully local surrogate for error backpropagation, without explicit error routing or nonlocal weight symmetry at runtime.

3. Algorithmic Variants: Single-Phase, Random Feedback, Adaptation, and Neuromorphic Realizations

Multiple directions generalize or optimize CHL's basic two-phase structure:

Single-Phase (Dual Propagation and Adjoint-State Methods)

Dual Propagation (DP) introduces a dyadic neuron architecture with two internal states per neuron, encoding activity and error through mean and half-difference (Høier et al., 2023, Høier et al., 2024). A single closed-form or fixed-point inference phase replaces the two alternated free/clamped phases:

• For $L$ layers, DP solves for activity vectors $z_k^+, z_k^-$ at each layer using coupled update equations that, at equilibrium, encode both the nominal and nudged states.
• The weight update is local:

$$\Delta W_{k-1} \propto (z_k^- - z_k^+)(\bar{z}_{k-1})^T$$

with $\bar{z}_{k-1} = ½(z_{k-1}^+ + z_{k-1}^-)$ (Høier et al., 2024). This matches the backpropagation gradient as the nudging parameter tends to zero, but avoids explicit two-phase dynamics and supports hardware-efficient implementation.

Random Feedback and Feedback Alignment

Random Contrastive Hebbian Learning (rCHL) eliminates the requirement for symmetric (transposed) feedback by introducing fixed random matrices $G_{k+1}$ for top-down signals during the clamped phase:

$$dx_k/dt = - x_k + f_k( W_k x_{k-1} + γ G_{k+1} x_{k+1} + b_k )$$

This approach achieves high classification accuracy, with convergence and learning dynamics dependent on the pseudospectral properties of $G_{k+1}$; learning remains local and asynchronous (Detorakis et al., 2018).

Neuronal Adaptation and Dynamics Regularization

Adjusted adaptation augments CHL by introducing an adaptation trace $\hat x^{\rm adp}$ per neuron:

$$\hat x_{t+1}^{\rm adp} = (1-c)\, \hat x_{t}^{\rm adp} + c\, x^-$$

Post-clamped-phase activities are nudged toward prior free-phase values before the weight update, effectively shrinking gaps in neural responses and yielding smoother gradients better aligned with backpropagation (Kubo et al., 2022).

Spiking Neurons and Biophysical Substrates

CHL has been extended to leaky integrate-and-fire (LIF) spiking neuron models. Neural state variables (e.g., filtered spike trains or firing rates) replace rate-based activities. Synaptic updates remain Hebbian-contrastive, computed from low-pass filtered spike trains at the end of each phase (Mesnard et al., 2016).

Temporal Contrastive Learning and In-situ Memory

Temporal Contrastive Learning (TCL) implements contrastive plasticity using integral feedback at each synapse, enabling a continuous “phase-less” protocol. Internal synaptic traces act as implicit non-equilibrium memory, facilitating local extraction of free/clamped differences without explicit phase switching or global memory (Falk et al., 2023). This approach imposes Landauer-like trade-offs between learning accuracy and thermodynamic dissipation.

4. Empirical Performance and Practical Considerations

CHL and its variants have been systematically benchmarked across a range of tasks:

Architecture	Dataset	Baseline	CHL/Variant Accuracy	Notes
MLP (782–6–10)	MNIST	—	17.3→12.0% (w/ adap.)	Adaptation: −5.4% error
MLP (784–128–64–10)	MNIST	—	CHL 97.7%, rCHL 97.5%	rCHL ≈ CHL, BP/FDA ≈97%
VGG16	CIFAR10	BP: 92.3%	DP: 92.3%	Runtime: DP = BP
VGG16	CIFAR100	BP: 69.6%	DP: 69.6%	DP matches accuracy
LIF Spiking	Regression	—	Error <0.02 (rate NNs)	Functional convergence
Autoencoder (784–36–784)	MNIST	—	CHL/rCHL codes learned	rCHL slightly noisier

Adaptation reduces test error, variance, and accelerates convergence. rCHL performs at parity with standard CHL and backpropagation, especially on small-scale or shallow networks; performance is slightly reduced for deeper/larger tasks, traceable to feedback alignment effects (Detorakis et al., 2018, Kubo et al., 2022). Dual Propagation empirically matches both BP and conventional CHL in accuracy and runtime, including for large-scale (ImageNet32x32) architectures (Høier et al., 2023, Høier et al., 2024).

Empirically, adaptation enhances stability: gradient-angle analysis reveals that CHL with adaptation produces weight updates closer in direction to those of backpropagation than vanilla CHL. Lowering the learning rate without adaptation fails to yield comparable gains, ruling out a simple effective $\alpha$ reduction (Kubo et al., 2022).

5. Biological Plausibility, Locality, and Interpretational Implications

CHL’s phase-based, local update rule is strongly motivated by principles of biological plausibility:

• Synaptic modifications depend only on locally accessible pre- and postsynaptic activities sampled at two moments;
• No requirement for explicit transport of error gradients, or nonlocal computation;
• Random feedback and compartmental/dyadic neuron formulations further relax biological constraints by avoiding bidirectional synaptic symmetry (Detorakis et al., 2018, Høier et al., 2023);
• Adaptation is directly inspired by observed neuronal adaptation currents, which regulate firing stability and context filling (Kubo et al., 2022).

Dyadic/compartmental neuron models (DP) and temporal/integral memory approaches further align with known neurobiological architectures, such as dendritic error signaling or homeostatic adaptation (Høier et al., 2024, Falk et al., 2023). TCL’s continuous non-equilibrium learning mechanism can be mapped onto diverse physical and biological substrates exhibiting Hebbian plasticity plus feedback (Falk et al., 2023).

6. Extensions: Meta-Learning and Theoretical Unification

Contrastive rules extend beyond standard supervised learning. In meta-learning, a contrastive “meta-Hebbian” update is applied by contrasting local parameter gradients at base-task and nudged meta-objective equilibria:

$$\Delta \theta = -\frac{1}{\beta} \left[ \partial_\theta \mathcal{L}(\hat\phi_{\theta,\beta}, \theta; \beta) - \partial_\theta \mathcal{L}(\hat\phi_{\theta,0}, \theta; 0) \right]^T$$

where $\mathcal{L}$ is an augmented loss with controllable meta-loss nudging parameter $\beta$ (Zucchet et al., 2021). This approach avoids backpropagation through time or second derivatives and empirically achieves parity or improvement over BPTT-base methods for supervised, few-shot, and meta-RL tasks.

Theoretical frameworks unify CHL, equilibrium propagation, and predictive coding within the EBM paradigm, showing that all recover backpropagation gradients in the limit of vanishing feedback or nudging strength. These results explicate the trade-off between local implementability and global credit assignment accuracy and inform practical algorithm design (choice of nudging strength, adaptation rate, phase duration) (Millidge et al., 2022).

7. Limitations, Trade-Offs, and Outlook

Classical CHL requires two distinct phases and potentially slow equilibrium relaxation, which can be impractical for very deep networks or neuromorphic systems. Single-phase approaches (DP, DP⁺) and TCL offer speed and hardware advantages but may require careful parameterization (nudging strength, compartmental symmetry) to ensure stability and biological plausibility (Høier et al., 2024, Falk et al., 2023).

CHL’s approximation error relative to backpropagation is controlled by the gap between free and clamped responses, which can be reduced by adaptation, random feedback design, or by incorporating higher-order corrections. Nevertheless, trade-offs remain between learning speed, fidelity to exact gradient descent, and strict locality/physical implementability.

A plausible implication is that CHL and its variants, by bridging the gap between global gradient-based learning and strictly local, phase-based synaptic plasticity, provide a tractable and physically implementable route for scalable learning in both artificial and biological networks (Kubo et al., 2022, Falk et al., 2023, Høier et al., 2024). The theoretical, algorithmic, and hardware directions emerging from CHL research continue to influence understanding of neural credit assignment and inform the development of neuromorphic and biologically inspired machine learning systems.