Neurobiological mechanism for credit assignment across time and depth

Determine the mechanisms by which the brain performs credit assignment and learning across both recurrent temporal dynamics and hierarchical cortical depth (cortical layers and regions).

Background

The paper argues that backpropagation through time (BPTT) is implausible for biological implementation and reviews forward-mode alternatives like RTRL and E-prop. Despite proposing a mathematical extension of E-prop to deep networks, the authors emphasize that the broader neurobiological question of how the brain achieves credit assignment simultaneously over time and hierarchical depth remains unresolved.

This open problem frames the motivation for biologically plausible algorithms that can handle both temporal recurrence and multi-layer cortical organization.

References

However, how exactly the brain solves the credit assignment and learning problem across both time (recurrent dynamics) and depth (cortical layers and regions) is still unknown.

— Generalising E-prop to Deep Networks  (2512.24506 - Millidge, 30 Dec 2025) in Introduction (first paragraph after Abstract), Page 1

However, this relegates the weight transport problem to a weight update transport problem, and it remains unclear how these symmetric updates can be achieved given the omnipresent temporal noise and morphological variability in neuronal substrates.

— Dendritic structure enables powerful plasticity  (2608.23251 - Hünerbein et al., 24 Aug 2026) in Box “The weight transport problem in biologically plausible credit assignment”

While it remains to be shown how any of these mechanisms can allow the kind of deep credit assignment and complex real-time modulation that the dendritic models are capable of, their presence in biology and their importance for plasticity is uncontroversial.

— Dendritic structure enables powerful plasticity  (2608.23251 - Hünerbein et al., 24 Aug 2026) in Section 5, paragraph “Information separation: beyond compartments”