Neuromodulated
Local Learning

The credit assignment problem — determining how each parameter in a deep network should change to improve a global objective — is the central challenge of training neural networks.

Backpropagation[1] solved this elegantly by applying the chain rule recursively, propagating error gradients from the output back through every layer. This method has powered decades of progress in deep learning.

However, backpropagation's chain rule has a structural vulnerability: the gradient signal must travel through every layer in sequence. In deep, plain networks (those without skip connections or normalization layers), this signal degrades exponentially. Gradients shrink toward zero in early layers (the vanishing gradient problem[2]) or, less commonly, explode toward infinity. The practical consequence is well known: a 50-layer plain MLP trained with standard backpropagation often fails to learn at all.

The deep learning community has largely treated this as an architecture and optimization problem. Skip connections (ResNets[3]), gating mechanisms (LSTMs[4], Highway Networks[5]), normalization layers (BatchNorm[6], LayerNorm[7]), and careful initialization schemes[8] all create conditions where backpropagation's chain can function in deep networks. These solutions are effective, but they address the symptom (gradient degradation) by modifying the architecture to make the chain work — they do not question whether the chain itself is necessary.

A different line of work asks: what if layers could learn without the chain? Local learning methods assign each layer its own training objective, eliminating the need for gradients to flow between layers. Hinton's Forward-Forward algorithm[9] and the Mono-Forward approach[10] demonstrate that competitive accuracy is achievable without backpropagation, with significant savings in memory and energy. However, purely local learning sacrifices global coordination. Each layer optimizes its own objective in isolation, unaware of how its learned representations serve or hinder other layers. This gap is analogous to a team of workers who each do their individual tasks well but have no manager coordinating the overall effort.

We propose Neuromodulated Local Learning, a hybrid approach that restores global coordination to locally-learning layers without reintroducing backpropagation's gradient chain. The key idea: a small, external neuromodulator network observes summary statistics from all layers and broadcasts a learned modulation signal that scales each layer's local loss. This is directly inspired by neuromodulation in biological brains, where dopaminergic and other neuromodulatory systems broadcast global signals (reward prediction error, arousal, novelty) that modulate the intensity of local synaptic plasticity without specifying the direction of weight changes.[11]

This paper makes the following contributions:

• Novel framework: We introduce Neuromodulated Local Learning, the first (to our knowledge) application of a learned neuromodulatory signal to coordinate local learning in standard (non-spiking) deep networks.
• Depth invariance: We provide proof-of-concept experiments demonstrating that the neuromodulated hybrid approach is nearly depth-invariant, losing only 1.1% accuracy across a 10× increase in network depth, compared to 7.3% for local-only learning and 42.5% for backpropagation.
• Scaling advantage: We show that the neuromodulator's advantage scales with depth, suggesting it addresses a real coordination deficit in local learning rather than acting as a simple regularizer.

2.1  Mitigating Vanishing Gradients

The vanishing gradient problem was identified by Hochreiter[2] and Bengio et al.[12] in the 1990s. Architectural solutions include LSTM gating[4], Highway Networks[5], and residual connections.[3] Normalization techniques (BatchNorm[6], LayerNorm[7]) stabilize intermediate activations. Gradient clipping[13] caps gradient magnitudes. Layer-wise adaptive learning rates (LARS[14], LAMB[15]) scale updates per layer but use fixed heuristics rather than learned signals. These methods keep backpropagation's chain intact while engineering conditions for it to succeed.

2.2  Local Learning Methods

Local learning eliminates the backward chain entirely. Deep Supervision[16] attaches auxiliary classification heads at intermediate layers but still uses backpropagation for each head. Hinton's Forward-Forward algorithm[9] replaces the backward pass with a local "goodness" objective at each layer. The Mono-Forward approach[10] assigns local classification heads trained independently, reporting competitive accuracy with 31% of the memory and 41% less energy than backpropagation. Jaderberg et al.'s Synthetic Gradients[17] predict what gradients would be, enabling asynchronous updates but still approximating backpropagation's signal.

2.3  Learned Optimizers and Meta-Learning

Andrychowicz et al.[18] train an LSTM to replace the optimizer entirely, learning update rules from data. Hypernetworks[19] use one network to generate weights for another. GradNorm[20] learns to balance gradient magnitudes across tasks in multi-task learning. The Dopamine optimizer[21] uses perturbation-based reward prediction error as a global signal, inspired by biological dopamine but operating in a different regime than our approach.

2.4  Three-Factor Learning Rules

In computational neuroscience, three-factor learning rules combine pre-synaptic activity, post-synaptic activity, and a global neuromodulatory signal to govern synaptic plasticity.[11] This framework captures how dopamine, serotonin, and acetylcholine modulate learning intensity in biological neural circuits. However, three-factor rules have been studied almost exclusively in spiking neural networks and biologically-plausible models. Our work adapts this principle to standard deep learning: we implement the neuromodulatory factor as a learned network that broadcasts per-layer scaling factors to modulate standard gradient-based local learning. To our knowledge, this is the first application of a learned neuromodulatory signal in non-spiking deep networks.

3.1  Problem Setting

Consider a deep MLP with L layers. In standard backpropagation, the loss gradient for layer l depends on all layers above it:

The product of Jacobians across many layers is the source of vanishing (or exploding) gradients. In local learning, each layer l has its own classification head and loss L_l, and activations are detached between layers so no gradient flows across the boundary. This eliminates the product entirely but leaves each layer learning in isolation.

3.2  The Neuromodulator

We introduce a small external network M (the neuromodulator) that observes summary statistics from every layer and outputs a per-layer modulation factor m_l ∈ [0.1, 2.0]. The modulated local loss for layer l is:

The neuromodulator M is a small feedforward network (2 hidden layers, 32 and 16 units) that takes as input, for each layer, four summary statistics:

These statistics are detached from the computational graph so no gradient from the neuromodulator flows into the layer weights. The modulation factor is bounded via a scaled sigmoid, ensuring no layer is ever fully silenced or over-amplified:

3.3  Neuromodulator Training Objective

The neuromodulator has its own dual objective, distinct from the task loss:

This dual objective captures the neuromodulator's role: it must both keep the training process healthy (balanced learning across layers) and direct the network toward good global performance. Crucially, the neuromodulator does not specify what each layer should learn (that comes from the local objectives), only how intensely each layer should learn at any given moment.

3.4  Training Procedure

Each training step proceeds in two phases:

4.1  Setup

We test three training methods on CIFAR-10 using plain MLPs (no skip connections, no normalization layers) at depths of 5, 10, 20, and 50 layers. All networks use 128 hidden units per layer with ReLU activations and Xavier initialization. This architecture is deliberately bare: it is the setting where vanishing gradients are most severe, providing the clearest test of our hypothesis.

4.2  Results

The results reveal three distinct behaviors with increasing depth:

Backpropagation collapses. At 50 layers, the loss is stuck at 2.3027 (log(10), the cross-entropy of a uniform distribution over 10 classes). The network is guessing randomly. No learning occurs. The gradient chain is dead.

Local learning degrades gracefully. Without a gradient chain to break, local-only learning continues to learn at depth. However, it loses 7.3 percentage points from 5 to 50 layers, reflecting the coordination deficit: deeper layers receive increasingly degraded input representations from earlier layers that have no incentive to produce useful features for them.

The neuromodulated hybrid is nearly depth-invariant. From 51.8% at 5 layers to 50.6% at 50 layers, the neuromodulated model loses only 1.1 percentage points across a 10× increase in depth. At 50 layers, it outperforms local-only learning by 6.6% and backpropagation by 40.6%.

Critically, the neuromodulator's advantage scales with depth. At 5 layers, where even backpropagation works fine, the neuromodulator adds only +0.4% over local-only. At 50 layers, where coordination matters most, the advantage grows to +6.6%. This scaling pattern supports the hypothesis that the neuromodulator addresses a real coordination deficit rather than acting as a simple regularizer.

4.3  Neuromodulator Behavior

Analysis of the learned modulation factors reveals that the neuromodulator develops distinct strategies at different depths.

At 5 layers, the neuromodulator learns a nuanced, evolving profile. Early in training, it boosts the first layer (factor ~1.73) while dampening later layers. Over 30 epochs, it progressively shifts emphasis to middle layers, eventually settling on a profile that emphasizes the central layers while dampening the first and last. This suggests a curriculum-like strategy: establish strong input representations first, then refine the middle of the network.

At 20 and 50 layers, the neuromodulator converges to a near-binary strategy: factors of 2.0 (maximum) on layers it finds productive and 0.10 (minimum) on layers it considers dead weight. At 50 layers, the modulation starts broadly distributed and progressively concentrates on fewer, more effective layers. This emergent behavior resembles a learned form of effective depth selection — the neuromodulator discovers which layers contribute meaningfully and allocates learning intensity accordingly, without any explicit architectural pruning.

5.1  Why It Works

The neuromodulator succeeds by solving a specific problem: in local learning, layers near the input receive raw, unprocessed features, while deeper layers receive the output of earlier layers. If an early layer learns a poor representation, all downstream layers suffer, but none can signal the problem back (there is no backward gradient chain). The neuromodulator provides exactly this missing signal. By observing statistics from all layers simultaneously, it can detect imbalances (one layer's loss much higher than others, dead neurons, collapsed activations) and adjust modulation factors to redirect learning intensity where it is needed most.

This is analogous to how dopamine functions in biological neural circuits: it does not tell neurons what to encode, but it modulates how strongly they should update based on a global assessment of whether learning is going well (reward prediction error). Our neuromodulator operates on the same principle at a higher level of abstraction.

5.2  Biological Parallels

Three-factor learning rules in computational neuroscience posit that synaptic plasticity depends on (1) pre-synaptic activity, (2) post-synaptic activity, and (3) a global neuromodulatory signal.[11] Our framework maps directly onto this structure: the local loss gradient captures factors (1) and (2), while the neuromodulator's output provides factor (3). The key difference from prior computational neuroscience work is that we implement this in standard deep networks with continuous activations and gradient-based updates, rather than in spiking neural networks with spike-timing-dependent plasticity.

5.3  Limitations

5.4  Future Directions

• CNN and Transformer experiments: Testing on ResNet-style (without skip connections) and Transformer architectures would establish whether the neuromodulator's benefits transfer to practical architectures.
• Architecture search: Recurrent neuromodulators, attention-based neuromodulators, and neuromodulators that observe richer statistics (gradient norms, weight distributions) are all unexplored.
• Multiple neuromodulatory signals: Biological brains use multiple systems (dopamine, serotonin, norepinephrine, acetylcholine), each with different roles. Multiple neuromodulators with distinct objectives could provide richer coordination.
• Transfer and continual learning: The neuromodulator's ability to identify productive versus unproductive layers may be useful for transfer learning (freeze layers with low modulation) or continual learning (protect well-trained layers).
• Extreme depth: Testing at 100, 200, or 500 layers would reveal whether the neuromodulator's depth-robustness holds at extreme scales.

We introduced Neuromodulated Local Learning, a hybrid training framework that combines the gradient-chain-free robustness of local learning with a learned global coordination signal inspired by biological neuromodulation. In experiments on plain MLPs at depths up to 50 layers, the neuromodulated approach achieved near-depth-invariant accuracy (1.1% drop over a 10× depth increase), substantially outperforming both standard backpropagation (which collapsed entirely) and local-only learning (which lost 7.3%).

The neuromodulator's advantage scales with depth, reaching +6.6% over local-only learning at 50 layers, directly supporting the hypothesis that a learned global signal addresses a real coordination deficit in locally-learning layers. Analysis of the learned modulation patterns reveals emergent behaviors — curriculum-like emphasis shifting at shallow depths and effective depth selection at large depths — that were not explicitly designed.

These results suggest that the gap between local learning and backpropagation can be substantially narrowed by restoring a lightweight form of global coordination, without reintroducing the backward gradient chain that causes backpropagation to fail at depth. We believe this direction — drawing from neuroscience's three-factor learning rules to coordinate locally-learning artificial neural networks — is a promising avenue for training very deep networks efficiently and robustly.

1. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536.
2. Hochreiter, S. (1991). Untersuchungen zu dynamischen neuronalen Netzen. Diploma thesis, Technische Universität München.
3. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. CVPR, 770–778.
4. Hochreiter, S. & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780.
5. Srivastava, R. K., Greff, K., & Schmidhuber, J. (2015). Highway networks. arXiv:1505.00387.
6. Ioffe, S. & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. ICML, 448–456.
7. Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer normalization. arXiv:1607.06450.
8. Glorot, X. & Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. AISTATS, 249–256.
9. Hinton, G. (2022). The forward-forward algorithm: Some preliminary investigations. arXiv:2212.13345.
10. Pau, D. P. & Aymone, F. M. (2025). Mono-Forward: Backpropagation-free algorithm — a layer-parallel approach. arXiv:2501.00762.
11. Gerstner, W., Lehmann, M., Liakoni, V., Corneil, D., & Brea, J. (2018). Eligibility traces and plasticity on behavioral time scales: experimental support of neoHebbian three-factor learning rules. Frontiers in Neural Circuits, 12, 53.
12. Bengio, Y., Simard, P., & Frasconi, P. (1994). Learning long-term dependencies with gradient descent is difficult. IEEE Transactions on Neural Networks, 5(2), 157–166.
13. Pascanu, R., Mikolov, T., & Bengio, Y. (2013). On the difficulty of training recurrent neural networks. ICML, 1310–1318.
14. You, Y., Gitman, I., & Ginsburg, B. (2017). Large batch training of convolutional networks. arXiv:1708.03888.
15. You, Y., Li, J., Reddi, S., et al. (2020). Large batch optimization for deep learning: Training BERT in 76 minutes. ICLR.
16. Lee, C.-Y., Xie, S., Gallagher, P., Zhang, Z., & Tu, Z. (2015). Deeply-supervised nets. AISTATS, 562–570.
17. Jaderberg, M., Czarnecki, W. M., Osindero, S., et al. (2017). Decoupled neural interfaces using synthetic gradients. ICML, 1627–1635.
18. Andrychowicz, M., Denil, M., Gomez, S., et al. (2016). Learning to learn by gradient descent by gradient descent. NeurIPS, 3981–3989.
19. Ha, D., Dai, A., & Le, Q. V. (2017). HyperNetworks. ICLR.
20. Chen, Z., Badrinarayanan, V., Lee, C.-Y., & Rabinovich, A. (2018). GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks. ICML, 794–803.
21. Zhong, C., Jiang, Z., & Chen, T. (2025). Dopamine: A dopamine-inspired reward-based optimizer. arXiv:2510.08652.

All experiments were conducted using PyTorch. The complete code is available as a Jupyter notebook runnable on Google Colab's free tier (T4 GPU). Random seeds were fixed at 42 for reproducibility. Total training time for all 12 runs (3 methods × 4 depths) was approximately 2 hours on a single T4 GPU. The neuromodulator's parameter count ranges from 1,285 (5 layers) to 7,810 (50 layers), representing 0.3–0.6% of the total model parameters.