Diffusion of Neuromodulators for Temporal Credit Assignment in Recurrent Spiking Neural Networks
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Backpropagation through time (BPTT) is the foundation of most effective algorithms for training recurrent neural networks, as it enables precise credit assignment across temporal sequences of activity. Given its success, one might expect biological recurrent circuits to employ a similar mechanism for assigning credit through time. However, its complete implementation is biophysically implausible: neurons cannot store all past states or transmit exact error signals backward through time. Moreover, BPTT requires each synapse to access global information about all other synapses, which is incompatible with the locality of biological computation. The success of BPTT has inspired the search for biologically plausible approximations, such as eligibility propa- gation (e-prop) and random feedback local online (RFLO). These approaches combine synaptic eligibility traces with top-down feedback signals to guide plasticity. Yet, their performance critically depends on dense or precisely designed feedback pathways, a requirement that is difficult to reconcile with the sparsity of long-range feedback connections observed in the brain. We propose that spatially diffusing neuromodulatory signals provide a solution to this problem. Rather than assuming error signals reach target neurons with surgical precision, we model them as neuromodulators diffusing through neural tissue to influence neighboring cells over multiple timescales. In this way, neurons receive error signals from the local concentration of released and diffused modulators, regardless of the signals’ origin points. To assess the impact of error signal diffusion on learning, we trained spatially embedded Recurrent Spiking Neural Networks (RSNNs) on three benchmark tasks. We compared the performance of standard e-prop against our diffusion-enhanced variant. Under the constraint of sparse feedback connections, we found that the diffusion mechanism significantly enhances the learning capability of the RSNNs, achieving superior performance in all three tasks. Our results suggest that diffusion of neuromodulatory signals may play an important role in the learning capacity of neural circuits.
Fields of Study & Subjects
Neural dynamics and brain function
Advanced Memory and Neural Computing
Neural Networks and Applications
Publication details
- Journal / source
- MPG.PuRe (Max Planck Society)
- Publication date
- 1 March 2026
- Language
- EN