Severing the network to make it honest, and what that costs wetware

What the work claims

The argument is that the global entanglement of ordinary neural networks, where every weight can in principle touch every output, is the shared root of three failures: they are opaque, they are wasteful, and they forget. The proposed cure is structural. Their Decomposable Spiking Neural Network (D-SNN) partitions the hidden layers into K disjoint experts, one for each output class, and prohibits cross-talk by zeroing every weight outside the block-diagonal blocks; the severing is architectural, a matter of weights set to zero in software, not a physical cut.¹ Each expert is a self-contained circuit that verifies a single class in a One-vs-All scheme, so any classification can be read straight off which pathway lit up.

Two claims sit on top of this structure. First, a training objective the authors call a Push-Pull loss, inspired by the lateral inhibition of the insect antennal lobe, concentrates activity in the correct expert and suppresses the rest, so that in a leaky integrate-and-fire spiking network the non-target pathways stay largely inactive. Second, because pathways are physically separate, you can freeze the weights of an already-learned class and add new ones without erasing the old, addressing catastrophic forgetting. This is a machine-learning architecture paper; the biological language is analogy, and the system is trained entirely as a continuous network and then ported to spikes.

How the division of labor works

The pipeline is deliberately two-stage. A shared convolutional front end extracts features; the hidden layers are then either dense or partitioned into block-diagonal experts. Training happens in the continuous artificial-network domain using the Push-Pull loss, which compares total activity across all deep-layer neurons against activity in the correct expert and pushes the difference toward localization. Only after convergence are the learned weights transferred, unchanged, into a spiking network of identical shape; no learning happens in the spiking domain at all.¹ The authors justify this because it sidesteps the cost of backpropagating through non-differentiable spikes over time, and they report it runs roughly 10 to 50 times faster than a backpropagation-through-time baseline of equal size while reaching comparable accuracy.

The numbers are strong where the task is easy and weaken as it hardens. The isolated-expert model reaches about 98.15 percent on MNIST and about 91.68 percent on Fashion-MNIST using a few hundred thousand parameters, an order of magnitude fewer than the 3.3 to 3.9 million of comparable spiking networks, while matching a conventional cross-entropy baseline to within a few hundredths of a percent.¹ The efficiency story is cleaner in the spiking operations count: on MNIST the isolated experts perform roughly 37 thousand synaptic operations against about 2.56 million for the dense baseline, near a seventy-fold reduction, because activity concentrates in a single pathway. On CIFAR-10 every variant lands in the low 80s and the modular advantage disappears: the isolated experts reach about 82.7 percent, essentially tied with the dense hybrid at 83.3 percent and dense cross-entropy at 82.8 percent, so isolation buys nothing on the harder task. More troubling than that expected decline, the paper's own abstract quotes a far lower figure of around 66 percent that its tables contradict; the discrepancy is unexplained and is a reason to treat the CIFAR-10 claim cautiously. The most interesting result is the continual-learning test: fed classes one at a time up to the full ten, the isolated-expert model ends at 54.36 percent accuracy while the conventional cross-entropy baseline collapses toward random guessing at about 10 percent. Crucially, the authors are candid that structural isolation alone is not sufficient; they find that freezing the weights of learned pathways, a stand-in for biological memory consolidation, is required to hold representations, and even then 54.36 percent is a partial rescue, not a solved problem.

Where a skeptic should push

The load-bearing word is "biologically-inspired," and it is doing more work than the biology supports. Zeroing off-block weights is not something nervous systems do; real circuits are not severed into clean One-vs-All modules, and the antennal lobe the authors cite achieves segregation through inhibition within a connected network, not through physical excision. The paper actually contains the evidence for this tension: its dense-hybrid variant keeps full connectivity and acquires modularity only through the learned Push-Pull inhibition, and it captures much of the benefit. That suggests the biologically faithful mechanism is the loss function, and the block-diagonal severing is an engineering convenience layered on top. A second push: the whole system is trained as a continuous network and ported to spikes with no on-device learning, so its claims are about a silicon-style architecture, not about spiking computation that arises in place.

Then there is the strength of the results. MNIST and Fashion-MNIST are near-saturated benchmarks; the collapse to the low 80s and the inconsistent reporting on CIFAR-10 suggest the modular story may not survive richer inputs where classes share features and clean per-class pathways become harder to justify. The forgetting result is genuinely better than the baseline but the honest reading is that isolation buys interference-freedom, not memory: without the added weight-freezing the retained accuracy is modest, and the freezing trick requires knowing which weights belong to which task, which is trivially available here precisely because the architecture was cut that way in advance. In a network that organized itself, that knowledge would not be free.

Why wetware cannot be block-diagonalized

Read as a design principle rather than a benchmark, this paper describes properties a living computing substrate structurally cannot offer, and that is the uncomfortable implication for organoid intelligence. A cortical organoid or neuronal culture is the purest form of the "globally entangled" network the authors set out to fix: it wires itself, densely and recurrently, with no block-diagonal structure and no way to zero a weight. The two prizes the D-SNN wins by severing connections, an auditable decision that can be traced to one isolated pathway and resistance to interference between classes, are exactly the two things a self-organized tissue struggles to hand over. A trained organoid classifier is a black box whose output cannot be attributed to a clean module, and a plastic living network trained on new inputs will tend to overwrite old ones. The paper's own remedy, freezing the synapses that hold a previous task, does not carry over to tissue in the same way, and the reason is precise: not that biology lacks consolidation mechanisms, which it has, but that there is no known map from a task to the specific synapses that encode it and no external means to clamp them. In a world that increasingly demands verifiability from any deployed classifier, non-decomposability is not an aesthetic complaint; it is a governance liability that silicon can escape and wetware, absent imposed structure, largely cannot.

The opportunity is the mirror image, and the paper points to the more realistic route. If structural severing is impossible in tissue, functional segregation through inhibition may not be. The authors' own dense-hybrid result, full connectivity plus a Push-Pull inhibitory objective that recovers much of the modular benefit, is the biologically plausible analogue: living circuits do enforce sparse, stimulus-specific routing through lateral inhibition, so an organoid pushed toward strong inhibitory gating might approximate auditable, low-crosstalk pathways without any wires being cut. That reframes a concrete engineering target for the field: stop treating the homogeneous blob as the goal and start imposing architecture, whether through compartmentalized assembloids joined by defined tracts, micro-patterned cultures with engineered inhibitory boundaries, or training regimes that reward segregation. The threat and the opportunity share one mechanism: inhibition is the only lever a living substrate has for the modularity that silicon simply builds in, and this paper's quiet lesson is that inhibition might be enough, if the field stops asking tissue to be a monolith.

The bottom line

Established: physically isolating spiking experts, trained in a continuous network by an inhibition-style loss and ported to spikes, matches dense accuracy on simple benchmarks with an order of magnitude fewer parameters, yields decisions traceable to a single pathway, and, when combined with weight-freezing, resists catastrophic forgetting far better than a conventional baseline. Hypothesis, or worse, undecided: that the modular advantage survives hard, feature-shared datasets, given the CIFAR-10 degradation and the paper's own inconsistent numbers there, and that "biologically-inspired" means more than analogy, given that the faithful mechanism is the inhibitory loss and the severing is engineering. For organoid intelligence the transferable content is not the architecture, which tissue cannot instantiate, but the diagnosis: the auditability and interference-resistance that modularity buys are precisely what self-wired living computers lack, and the most plausible biological route to them runs through inhibition, not through a scalpel. What would confirm the broader thesis is a demonstration that inhibition-enforced functional modularity delivers auditability in a real network, spiking or living; what would break it is evidence that severing, not the loss, is doing the work, in which case the lesson does not cross into wetware at all.