Neural network training has relied for decades on backpropagation, an effective but biologically questionable algorithm. Recently, a research team has presented an alternative method called Error Diffusion that allows deep networks to be trained without the need to transport arrays of transposed weights, a requirement that the human brain probably cannot meet. This approach not only eliminates the problem of weight carrying, but also respects Dale's principle, which states that a neuron releases the same type of neurotransmitter (excitatory or inhibitory) in all its synapses.
The proposed architecture divides each layer into two parallel flows: one excitatory and one inhibitory. All weight matrices remain non-negative, and the cross-connections between flows incorporate negative signals in a structural, unlearned way. This implies that each layer requires four subarrays, which quadruples the number of parameters compared to a conventional network. However, this additional complexity allows learning to be completely local: each weight update depends solely on presynaptic activity, that derived from postsynaptic activation, and a single global error sign.
To scale this method beyond binary classification, the researchers introduced module-based error routing. Each hidden neuron is assigned to a specific output channel using the function r(i) = i mod C, where C is the output dimension. This structured mapping replaces the random feedback matrices used in other methods such as Direct Feedback Alignment (DFA). In addition, three key innovations for multiclass sorting were implemented: specific sigmoid widths per layer to avoid premature saturation, error centering by class in batches to correct imbalances, and an asymmetric initialization that favors excitatory connections over inhibitory connections in a 3:1 ratio.
The results obtained are remarkable: 96.7% is reached in MNIST and 61.7% in CIFAR-10, while the version without these improvements collapses to 50.4% and 11.6% respectively. Although DFA achieves higher scores (97.6% and 69.1%), it does so in violation of Dale's principle: it uses approximately 2.84 million negative pesos. The proposed method is the first to train convolutional networks in full respect of this biological constraint. The price of complying with Dale is quantified in a loss of between 0.9 and 7.4 percentage points against DFA.
Beyond classification, the team extended Error Diffusion to reinforcement learning, integrating the algorithm with Proximal Policy Optimization (PPO). In environments such as Brax's HalfCheetah, ED-PPO significantly outperforms BP-PPO (5494 vs. 3520) and ties with DFA-PPO. In Craftax, however, DFA-PPO turns out to be the weakest method (19.8 vs. 27.0 for BP-PPO), suggesting that random feedback that works for classification may fail in open-ended learning tasks.
This breakthrough has profound implications for neuromorphic and photonic hardware, where synaptic magnitudes are physically encoded as non-negative values. In addition, training induces implicit pruning: 37.3% of weights fall to the ground (10⁻⁴) after training, and cross-inhibitory connections are pruned up to 68.8%. This opens the door to model compression at no additional cost.
From a business perspective, the ability to train neural networks without backpropagation and respecting biological principles has practical applications in sectors such as robotics, embedded systems and explainable artificial intelligence. Companies such as Q2BSTUDIO, which specialises in artificial intelligence for companies, can take advantage of these advances to develop tailor-made applications that require energy efficiency or deployment on specialised hardware. Combining AI for business with AI agents that learn locally could revolutionize industries such as logistics or manufacturing, where latency and consumption are critical.
Integrating these methods with AWS and Azure cloud services would allow lighter models to be trained and deployed in edge environments with limited resources. In addition, the local learning approach is inherently more robust against adversarial attacks, making it an attractive option for cybersecurity and anomaly detection systems. The ability to automatically prune weights also aligns with compression strategies for business intelligence services that require fast and accurate models, using tools such as Power BI to visualize real-time performance metrics.
Custom software development based on Error Diffusion could be implemented on industrial automation platforms, where robots need to learn tasks without relying on a constant connection to the cloud. Q2BSTUDIO offers experience in the creation of complete solutions ranging from the design of the model to its integration into cloud infrastructures, guaranteeing scalability and regulatory compliance.
In summary, Error Diffusion represents a significant step towards more biologically plausible and computationally efficient neural networks. Although it is still far from the performance of conventional backpropagation in complex tasks, its combination of local learning, respect for Dale's principle, and extension to reinforcement learning make it a promising field of research. For companies looking to innovate in artificial intelligence, exploring these alternatives is key to differentiating themselves and building more robust and sustainable systems.
