$\vec{\delta}$ Multiplexed Gradient Descent: Perturbative Learning with Astrocytes
Year: 2026
Type: conference-paper
Abstract: We present$\vec{\delta}$-Multiplexed Gradient Descent ($\vec{\delta}$MGD), a fully neuromorphic algorithm that consistently performs alongside backpropagation with orders of magnitude fewer gradient component calculations across a number of datasets and network architectures. This performance is enabled by the finding that perturbative methods perform near-optimally when parameters are perturbed in the same direction as the true ternary gradient, and that even a noisy estimate of the gradient yields marked improvements in learning performance and convergence time. We show that local synaptic information can be used to make strong estimates of the ternary gradient using only presynaptic activation and postsynaptic bias gradients. Moreover, we find the latter can be cast into learnable classwise representations, which we call DeltaVecs. According to these findings we realize a reduction of$N_{\text {samples }} \cdot N_{\text {neurons }}^{2} \rightarrow N_{\text {classes }} \cdot N_{\text {neurons }}$in terms of gradient components that must be explicitly computed per epoch. We further find that astrocyte-inspired subnetworks are able to learn DeltaVec representations in order to drive gradient-informed perturbations. Finally, we stack these methods together and show that using only neuro-inspired hardware-friendly operations we can perform alongside backpropagation in networks on the scale of millions of parameters, in DNNs, CNNs, and even transformers. Altogether, we offer an algorithm implementing novel machine learning methods and neuro-inspired operations to realize compelling neuromorphic intelligence.
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7165533024