Prospective neurons solve timing delays for deeper neural network learning

A team led by Nicolas Zucchet, including Qianqian Feng, Axel Laborieux, Friedemann Zenke, Walter Senn, and João Sacramento, published a paper on arXiv (2511.14917) demonstrating how 'prospective neurons' can solve a fundamental timing problem in hierarchical neural networks. The issue arises because working memory relies on slowly integrating neurons that maintain persistent activity over time. In a multi-layer network, these delays accumulate, causing teaching signals (error feedback) to arrive out of sync with the neural activity that produced the behavior.

To fix this, the researchers enhanced neurons with an adaptive current that allows them to respond to stimuli prospectively—effectively predicting future inputs. This predictive synchronization ensures error signals align with the relevant neural states. They showed that prospective neurons enable teaching signal synchronization across multiple learning algorithms (e.g., backpropagation through time) and validated the approach on motor control tasks, where agents learned to form and retrieve memories over long timescales. Mathematical analysis confirmed the coding efficiency. The work bridges computational neuroscience and AI, offering a biologically plausible mechanism to train deep spiking networks without the classic credit assignment lag.