Open problems in scaling, evaluating, and experimentally validating biologically plausible credit assignment algorithms

Establish effective methodologies to improve, scale, and rigorously test the capabilities of biologically plausible credit assignment algorithms, including predictive coding, contrastive Hebbian learning, and forward-only learning, and develop experimental protocols to empirically verify the specific claims these algorithms imply about cortical processing.

Background

The paper surveys several biologically plausible credit assignment frameworks—predictive coding, contrastive Hebbian learning (including equilibrium propagation), forward-only learning, and related approaches such as direct feedback alignment, target propagation, and local representation alignment. While these methods address key biological and hardware-related critiques of backpropagation, they have not consistently matched the practical performance and scalability of backpropagation on modern deep learning tasks.

The authors emphasize that artificial neural networks provide a test-bed for assessing the capability of neuromimetic algorithms and for rapidly probing hypotheses about cortical computation. However, they note that many challenges remain open: improving and scaling these methods, establishing rigorous evaluation practices, and experimentally verifying the cortical processing implications that these algorithms suggest. Addressing these open problems is necessary both for practical application and for connecting machine learning models to neuroscientific evidence.

References

In the last decade, a number of significant algorithms, which are reviewed in this work, have been proposed as solutions to the problem of cortical credit assignment. While these can serve as starting points for development, there remain many open problems both in improving, scaling, and testing the capabilities of these algorithms, as well as experimentally checking the claims that they imply about cortical processing.

— A Review of Neuroscience-Inspired Machine Learning  (2403.18929 - Ororbia et al., 2024) in Section 5, Discussion and Conclusion (Impact in the Neurosciences paragraph)

Several open directions and challenges remain for this line of research. First, substantial empirical effort is required to evaluate neurotransmitter-inspired mechanisms across a diverse range of learning problems, and across multiple application domains such as computer vision, audio understanding, and natural language processing. In particular, it is essential to establish whether neuromodulatory mechanisms provide consistent benefits when learning from large-scale, real-world datasets.

— Position Paper: Neurotransmitters as a Missing Dimension in Artificial Neural Networks  (2609.20083 - Li et al., 17 Sep 2026) in Section “Open Challenges and direction”

Beyond standard accuracy-based measures, additional metrics are needed to capture properties such as plasticity, learning efficiency, such as time to convergence, parameter efficiency, and memory utilisation. Moreover, higher-level qualitative attributes, including stability and flexibility, may also provide valuable insights into neuromodulatory behaviour. Developing unified evaluation protocols that incorporate these factors remains an open challenge.

— Position Paper: Neurotransmitters as a Missing Dimension in Artificial Neural Networks  (2609.20083 - Li et al., 17 Sep 2026) in Section “Open Challenges and direction”

However, this comes at the cost of slow learning, and it remains to be demonstrated whether their temporal Hebbian learning rule can scale to train large spiking networks.

— Dendritic structure enables powerful plasticity  (2608.23251 - Hünerbein et al., 24 Aug 2026) in Section 3.2, subsection “Global error broadcasting”

The transpose appearing in the neural update of equation~\ref{eq:neurons_synapses} does not pose a problem for analogue hardware, where crossbar arrays can be read in both the forward and reverse directions, although its biological plausibility remains an open question shared with predictive coding more broadly.

— Bio-inspired Learning and Decision-Making with Probabilistic In-Memory Computing Hardware: Part 1  (2609.11281 - Dalgaty et al., 10 Sep 2026) in Section "Algorithm Formulation"