The forward-backward disconnect: why living neural substrates need their own learning theory

What the work claims

The claim is structural. Neural models have reintroduced dynamical and biological structure into forward computation, recurrence, continuous-time evolution, fixed-point equilibria, and event-driven spikes, but the mechanisms that make these systems trainable remain concentrated around backpropagation, backpropagation through time, adjoint methods, implicit differentiation, and surrogate-gradient variants.¹ The authors name this asymmetry the forward-backward disconnect and propose a taxonomy with three coupled axes: state-dynamics structure, credit-assignment mechanism, and biological grounding, with grounding split separately into forward and learning dimensions.

This is a survey and synthesis, not a primary experimental paper. Its unit of analysis is the architecture-learning configuration rather than the architecture name alone. A spiking network trained with surrogate gradients and the same spiking architecture trained with spike-timing-dependent plasticity occupy different positions in the taxonomy because the credit-assignment mechanism changes.¹ The value is diagnostic: the taxonomy exposes which mechanisms a given configuration preserves, discards, or replaces with engineering substitutes.

How the taxonomy works

The horizontal axis partitions models by how their internal state evolves. Five classes are distinguished: static memoryless maps, discrete-time or sequence models with explicit hidden states, continuous-time models governed by differential equations, implicit-state models defined by fixed-point equations, and hybrid event-driven models that combine continuous subthreshold evolution with discrete spikes.¹

The vertical axis tracks credit assignment. Four categories cover the surveyed literature: global-gradient mechanisms such as backpropagation and BPTT; approximate or implicit-gradient mechanisms including feedback alignment, target propagation, surrogate gradients, equilibrium propagation, and adjoint methods; local plasticity rules such as Hebbian learning and STDP; and energy or unsupervised objectives such as contrastive divergence and the Forward-Forward algorithm.¹ The authors further annotate each mechanism by gradient relation, update scope, temporal requirement, objective source, and update timing.

The third axis is biological grounding, applied independently to the forward operator and the learning rule. Forward grounding ranges from none or historical inspiration through functional analogy and architectural constraint to mechanistic plausibility. Learning grounding ranges from none through weak and local or algorithmic to mechanistic plasticity. A surrogate-gradient-trained spiking network and an STDP-trained spiking network share the same LIF forward operator at mechanistic plausibility, but the former is weakly grounded on the learning axis while the latter is at mechanistic plasticity.¹

Where a skeptic should push

The most load-bearing assumption is that the taxonomy's categories are the right ones. The 32 configurations are a curated sample, not a systematic census of all published neural models, and the placement of borderline cases, modern Hopfield networks, predictive-coding variants, equilibrium propagation, Forward-Forward, requires adjudication that the authors document but that others might dispute. The counts therefore describe the surveyed literature, not an objective distribution of all possible configurations.

Second, the survey's conclusion that no alternative combines locality, temporal credit assignment, task breadth, scalability, and substrate alignment is a statement about the literature reviewed, not a proof of impossibility. The operational criterion is falsifiable, a strictly local rule reaching backpropagation-parity at ImageNet-1K or GPT-class scale would revise it, but until then the claim remains an inductive summary.

Third, the hardware-substrate discussion leans on well-known neuromorphic platforms and the hardware-lottery argument without new measurements. The statement that neuromorphic hardware removes the substrate penalty for event-driven inference but does not supply a scalable local learning rule is consistent with the surveyed literature, but it is not itself a hardware result.

Why organoid intelligence needs a credit-assignment theory

The non-obvious implication for organoid intelligence is that the hard problem is not building a biologically plausible forward model. Cultures already have recurrence, continuous-time membrane dynamics, spike-based signaling, and plasticity at scales no engineered substrate currently matches. The hard problem is credit assignment: how to train or steer those dynamics toward a useful computation without global gradients, stored activations, or a backward pass through the tissue. The forward-backward disconnect reframes organoid intelligence as a learning-theory problem first and a hardware problem second.

The specific mechanism behind this reframing is the temporal credit-assignment sub-problem. Backpropagation through time solves temporal credit assignment by unrolling time into a layered graph with tied weights, which requires storing the computational history and replaying it in reverse. Biological systems do not appear to do this.¹ Eligibility traces and three-factor rules, including e-prop, offer a biologically plausible alternative: a local transient memory holds pre-post coincidence until a modulatory signal arrives. But the survey notes that these approaches have not matched the scale or generality of global-gradient systems. For organoids, this means the question is not whether the tissue can compute, but whether it can be taught without an implausible reverse-time replay mechanism.

The opportunity is conceptual clarity. The taxonomy gives organoid researchers a way to state exactly what kind of learning they are claiming. A demonstration that a culture changes its response to a stimulus after repeated pairing is at the local-plasticity, mechanistic-plasticity corner. A demonstration that a readout decoder is trained by backpropagation on recorded firing rates is at the global-gradient, none-biologically-grounded corner. Most organoid-computing papers mix the two without labeling them. Using the taxonomy would force honest accounting: which part of the system is biological, which is engineered, and which learning rule connects them.

The threat is that the disconnect may be fundamental. If scalable credit assignment genuinely requires global error propagation, then living tissue, which lacks the wiring for such propagation, may never be trainable to perform deep computations on demand. That would relegate organoid intelligence to inference on prestructured dynamics or to small, local adaptations rather than to learning complex tasks from examples. The survey does not prove this pessimistic case, but it shows that no counterexample has yet satisfied the operational criterion of backpropagation-parity at scale.

A second threat is substrate mismatch. The paper maps the dense-gradient path, GPUs and TPUs optimized for matrix multiplication and global gradient transport, against the event-driven path, neuromorphic substrates optimized for sparse events and local state. The current empirical bridge is asymmetric: large spiking networks are trained off-chip with surrogate gradients and deployed on neuromorphic hardware for inference.¹ Organoid systems risk the same asymmetry: train a model in silicon, then ask the tissue to run inference. That is not biological computing; it is biological deployment. The survey's research agenda, configuration-level benchmarks that jointly report accuracy, locality, temporal behavior, memory, and energy, is exactly what organoid intelligence needs and currently lacks.

The bottom line

Established as a useful synthesis: forward dynamics in neural models have diversified across five state-dynamics classes, while scalable credit assignment remains concentrated in global-gradient methods and their close relatives. The taxonomy is a coherent diagnostic tool and the forward-backward disconnect is a productive framing. Not established: that the disconnect is insurmountable, or that any specific alternative is impossible; the survey explicitly frames its central claim as empirically falsifiable.

What would strengthen the survey's argument for organoid intelligence: a demonstration of a local or mechanistic-plasticity rule that reaches backpropagation-parity on a nontrivial benchmark, which would directly challenge the disconnect. What would break the optimistic reading: continued failure of such rules to scale, which would support the conclusion that living tissue is better suited to inference and local adaptation than to deep learning from examples. The immediate takeaway is methodological: organoid experiments should report credit-assignment mechanism, biological grounding, and substrate alignment as carefully as they report accuracy.