A Review of Neuroscience-Inspired Machine Learning - CPU, GPU and Neuromorphic Credit Assignment Algorithms Survey

Aug 4, 2026

Upload by Aria

Deep learning faces a core limitation: backpropagation assigns credit in ways that are biologically implausible, creating practical constraints such as incompatibility with certain hardware and difficulties with non-differentiable implementations, which in turn increase energy demands. The survey argues that biologically plausible credit assignment is broadly compatible with diverse learning settings, energy-efficient, and suitable for modeling with physical systems and non-differentiable behavior, enabling real-time adaptive neuromorphic processing. It reviews key algorithms spanning neuroscience, cognitive science, and machine learning, highlighting benefits across CPUs, GPUs, and neuromorphic hardware.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

AccessFree to download

The Problem of Credit Assignment

Neurobiological implausibility of backpropagation

Biologically plausible credit assignment and energy efficiency

Organization of the Review

A REVIEW OF NEUROSCIENCE-INSPIRED MACHINE LEARNING arXiv :2403 . 18929v1 [ cs .NE] 16 Feb 2024 Alexander Ororbia Rochester Institute of Technology Rochester, NY 14623, USA [ago@cs.rit.edu](ago@cs.rit.edu) Ankur Mali University of South Florida Tampa, FL 33620, USA [ankurarjunmali@usf.edu](ankurarjunmali@usf.edu) Adam Kohan University of Massachusetts Amherst Amherst, MA, USA [akohan@umass.edu](akohan@umass.edu) Beren Millidge Zyphra, Palo Alto, CA University of Oxford, Oxford, UK [beren@millidge.name](beren@millidge.name) Tommaso Salvatori VERSES AI Research Lab, Los Angeles, USA TU Wien, Vienna, Austria [tommaso.salvatori@verses.ai](tommaso.salvatori@verses.ai) ABSTRACT One major criticism of deep learning centers around the biological implausibility of the credit assignment schema used for learning – backpropagation of errors. This implausibility translates into practical limitations, spanning scientific fields, including incompatibility with hardware and non-differentiable implementations, thus leading to expensive energy requirements. In contrast, biologically plausible credit assignment is compatible with practically any learning condition and is energy-efficient. As a result, it accommodates hardware and scientific modeling, e.g. learning with physical systems and non-differentiable behavior. Furthermore, it can lead to the development of real-time, adaptive neuromorphic processing systems. In addressing this problem, an interdisciplinary branch of artificial intelligence research that lies at the intersection of neuroscience, cognitive science, and machine learning has emerged. In this paper, we survey several vital algorithms that model bio-plausible rules of credit assignment in artificial neural networks, discussing the solutions they provide for different scientific fields as well as their advantageson CPUs, GPUs, and novel implementations of neuromorphic hardware. We conclude by discussing the future challenges that will need to be addressed in order to make such algorithms more useful in practical applications. 1 The Problem of Credit Assignment One of the key tasks in artificial intelligence is to construct mathematical and algorithmic solutions to what is known as the grand problem of credit assignment. Effective credit assignment reduces to: (i) the identification of which neural processing elements (NPEs), e.g., individual computational units in a computation graph, have an influence on a particular (task-specific) objective functional L(Θ); and (ii) modifying the synapses that connect all of the NPEs based on their degree of influence so as to optimize this objective. The synaptic adjustments that characterize the second step are made to improve the overall performance of the network that the set of NPEs constitutes. From the perspective of error-driven learning and adaptation, credit assignment is typically carried out by computing and assigning error values to each NPE based on the cost L(Θ) and, once these values have been obtained, yielding ∆ (the set of all adjustments to be made to the synapses within Θ), the current values of the ANN’s parameters are consequently updated. Note that, historically, error-driven adjustment of the kind described above has been theorized [59, 13] and experimentally observed in biological neuronal networks [75] . The problem with how this error is computed and allocated in modern-day ANNs – via backpropagation of errors [32](backprop)– is essentially what is considered to be neurobiologically implausible. Given the significant successes of modern deep learning, addressing this implausibility may seem a niche task of interest to neuroscientists alone. However, this is far from true; despite recent breakthroughs, important advancements are still required. Two of these include the need to: 1) develop more ro- bust and more general human-like capabilities, in service of the grander goal of constructing intelligent machines, and 2) to construct f

FAQs

What is the main criticism of deep learning discussed in the review?

The review focuses on the biological implausibility of the credit assignment mechanism used in deep learning, namely backpropagation of errors. This implausibility leads to practical limitations across scientific fields and hardware constraints.

Why is biologically plausible credit assignment considered more practical than backpropagation?

It is compatible with many learning conditions and is energy-efficient. It also supports learning with physical systems and non-differentiable behavior, and it can enable real-time adaptive neuromorphic processing.

What benefits does locality bring when implementing biologically plausible approaches on neuromorphic hardware?

Locality enables parallelized operations with low latency and low power consumption. It also supports training networks with cyclic and entangled topologies without storing gradients, unlike backpropagation through time.

Aug 4, 2026

Upload by Aria