.jpg)
The Dale principle, fundamental in neuroscience, establishes that each neuron releases a single type of neurotransmitter, with its synapses being exclusively excitatory or inhibitory. Translating this biological restriction to artificial neural networks involves coordinating separate populations of excitatory and inhibitory neurons, which radically modifies credit assignment during learning. While traditional backpropagation requires transporting weights between layers, algorithms such as Error Diffusion propose an alternative pathway: routing global error signals to all layers without relying on random feedback matrices or transposed weights. However, until now it was unknown whether this approach could scale under the Dale principle in complex tasks, beyond binary classification. Recent research demonstrates that it is indeed possible, achieving 96.7% on MNIST and establishing a baseline of 61.7% on CIFAR-10, evidencing that representational learning is viable even under strict biological constraints.
A particularly relevant finding is that technical innovations —such as layer-specific sigmoid widths, batch-centered error signals, and asymmetric initialization— exhibit a relative importance that reverses between MNIST and CIFAR-10. This reveals task-dependent bottlenecks in credit assignment, invisible when evaluated on a single benchmark. In the realm of reinforcement learning, integrating Error Diffusion with Proximal Policy Optimization (PPO) shows competitive results against other backpropagation-free methods, opening the door to neural architectures more aligned with the principles of the real brain.
For companies seeking to incorporate these advances into their systems, having a specialized technology partner makes all the difference. At Q2BSTUDIO we develop artificial intelligence for businesses that leverages both classical techniques and emerging approaches, such as those based on biological principles. Our team designs custom applications that integrate deep learning models with realistic constraints, optimizing performance without sacrificing interpretability. Additionally, we offer custom software for sectors requiring high precision, such as cybersecurity or AWS and Azure cloud services, where computational efficiency is critical. We also implement business intelligence services with tools like Power BI to visualize the behavior of these models, and we develop AI agents capable of operating in open exploration environments, similar to those described in the most advanced research.
The context dependence in credit assignment is a reminder that there is no one-size-fits-all solution. That is why, at Q2BSTUDIO, we work side by side with each client to identify the optimal approach, whether through AI for businesses based on dual networks or through hybrid systems that combine supervised and reinforcement learning. Our experience in custom applications allows us to adapt these concepts to real needs, from process automation to strategic decision-making, always with a horizon of scalability and robustness.
