MIT Self-Assembling Neural Networks Learn Without a Prewired Training Circuit

MIT self-assembling neural networks learned pattern recognition and image classification without starting from a carefully engineered learning circuit. Instead, randomly connected units formed useful learning machinery under shared local plasticity rules.

That result changes the question researchers can ask about biologically plausible learning. A model no longer needs only to explain how an established circuit adjusts its connections. It can also explore how the circuit responsible for those adjustments emerges.

The central conflict is not simply MIT self-assembling neural networks versus conventional artificial intelligence. It is self-organized learning machinery versus backpropagation, an algorithm whose required coordination is difficult to map directly onto biology.

The researchers report performance similar to the backpropagation-trained networks they tested. However, they do not claim that their system replaces backpropagation in large AI models. They also do not claim that biological brains use the proposed mechanism.

The more important result is narrower. A collection of components can begin with random internal connections, follow local rules, and assemble motifs that approximate gradient-based learning. That shifts some design work from the network architect into the interactions among the network’s units.

What MIT Self-Assembling Neural Networks Actually Changed

The MIT study moves biological learning models one level deeper, from adapting a circuit to assembling the machinery that makes adaptation possible.

The research appeared online in the Proceedings of the National Academy of Sciences on October 9, 2026. MIT published its accompanying account on October 13.

The authors are Qianli Liao, Liu Ziyin, Yulu Gan, Brian Cheung, Mark T. Harnett, and Tomaso Poggio. Their work spans the McGovern Institute for Brain Research, MIT’s ICoN Center, CSAIL, and related MIT research groups.

Their peer-reviewed paper introduces a framework called Self-Assembling Learning. This framework asks how learning systems can arise from lower-level, distributed interactions.

Its first example is the Self-Assembling Motif, or SAM. A motif is a small recurring circuit structure that can serve as a functional building block.

SAM does not begin as a finished learning module. It emerges from initially random connectivity through heterosynaptic plasticity rules.

Heterosynaptic plasticity means that activity at one connection can influence changes at other connections. It gives a neuron more coordination than a rule that updates each synapse in isolation.

The team then composes these motifs hierarchically across multiple layers. According to the paper, the resulting network dynamics approximate a generalized form of stochastic gradient descent.

Gradient descent is an optimization method that repeatedly adjusts parameters to reduce errors. Stochastic gradient descent estimates those adjustments from subsets of training examples.

Modern neural networks usually calculate useful adjustments through backpropagation. Backpropagation sends error information backward through a network so each parameter receives a coordinated update.

That process works extremely well in software. Yet neuroscientists have long questioned whether biological tissue can reproduce its precise information flow, timing, and weight relationships.

The new result does not remove learning rules or supervision. The researchers still specify the local plasticity rules and provide data and task feedback.

What changes is the level at which the designer intervenes. The designer does not directly wire the final learning circuit represented by the motif.

The MIT research account describes multiple layers of randomly interconnected neuron-like units. Components within a layer follow the same rules for changing their connections.

Those units interact until organized behavior appears. Emerging motifs then become the machinery that supports learning across the network.

The researchers tested pattern recognition and image classification. The paper’s references identify benchmarks including MNIST, Fashion-MNIST, MedMNIST, and natural-image datasets.

These are meaningful tests for learning dynamics, but they are not equivalents of training a frontier language model. The study addresses the organization of learning circuits, not production-scale generative AI.

That distinction matters because “build their own learning machinery” can sound more autonomous than the experiment warrants. The network does not invent its task, data, or governing physics.

Instead, MIT self-assembling neural networks receive a lower-level specification. Their useful higher-level circuit organization then emerges from interactions within that specification.

The result creates the article’s central tension. Researchers can prescribe a learning circuit directly, or prescribe local conditions from which a learning circuit organizes itself.

The first approach optimizes control and performance. The second asks whether useful computation can emerge through a process closer to biological development and plasticity.

Why Local Rules Put Backpropagation Under Pressure

The study pressures models that call themselves biologically plausible while assuming that their most important learning circuit is already in place.

Backpropagation remains the practical standard for training modern artificial neural networks. Its effectiveness is not the disputed part of this story.

The biological question concerns implementation. A software framework can store activations, calculate an exact loss, and distribute derivatives across many layers.

A brain has no obvious central program performing that sequence. Neurons change through electrochemical signals, local activity, cellular processes, and feedback arriving through physical tissue.

A major backpropagation review explains several longstanding mismatches. These include weight transport, nonlocal credit assignment, and the separation of forward and backward computations.

Credit assignment is the problem of determining which internal connection contributed to an error. Deep networks make that problem harder because outcomes depend on many interacting layers.

Engineered systems solve credit assignment by calculating gradients. The calculation tells each parameter which direction should reduce the network’s error.

Many biologically inspired proposals try to approximate that calculation with feedback alignment, predictive coding, local losses, or alternative circuit dynamics. However, they often begin with an architecture already arranged to support the chosen method.

The MIT team targets that starting assumption. A biologically plausible account should explain not only how synapses update, but how an update-supporting circuit becomes organized.

Liao summarized the issue directly in MIT’s coverage: “The brain doesn’t have an external designer or architect.” His point is about the source of organization, not the absence of constraints.

Brains develop under genetic programs, physical limits, sensory experience, and biochemical rules. Those influences do not specify every mature circuit connection individually.

A useful computational model should therefore separate two questions. Which rules govern local changes, and which large-scale learning structures emerge when those rules interact?

SAM offers one answer. Random microscopic connectivity gradually supports an organized motif whose dynamics become useful for optimization.

The paper describes self-assembly at two scales. First, a motif emerges from random connectivity under local plasticity.

Second, networks of those motifs organize into hierarchical dynamics. The authors mathematically connect those dynamics to a generalized form of stochastic gradient descent.

That two-scale claim distinguishes the work from simply randomizing a feedback matrix. The network is supposed to develop both a circuit building block and a system-level learning process.

If that account holds under further testing, it raises the standard for biological learning models. Researchers would need to justify why their learning circuits exist, not merely show what those circuits calculate.

The pressure extends beyond neuroscience. Engineers also care about systems that learn under local information and physical constraints.

Neuromorphic hardware places computation near memory and often limits global data movement. A learning method based on local interactions might fit those constraints better than exact digital backpropagation.

That possibility remains a research direction, not a product result. The study does not provide evidence that SAM currently reduces training costs or energy consumption in deployed hardware.

It also does not make conventional backpropagation obsolete. Backpropagation has decades of tooling, empirical knowledge, and compatibility with large accelerator clusters.

The sharper conclusion is that backpropagation no longer owns the conceptual link between useful gradients and carefully designed global machinery. The paper offers a route through which gradient-like optimization emerges from local plasticity.

That is why the comparison matters even when benchmark performance is only similar. Performance parity can serve as a control showing that self-organization did not destroy the ability to learn.

The study therefore pressures a research assumption, not today’s AI training market. Its target is the belief that effective gradient machinery must be wired into the system beforehand.

The Learning Machinery Assembles Itself From Random Connections

The core mechanism is a feedback process in which local plasticity shapes a motif, and the emerging motif then supports coordinated learning.

To understand how MIT neural networks learn, start with the difference between homosynaptic and heterosynaptic change. Homosynaptic change depends mainly on activity at the synapse being modified.

Heterosynaptic change allows activity elsewhere to affect that synapse. This wider interaction can coordinate multiple connections attached to the same neuron.

A familiar local principle says that neurons activated together strengthen their connection. That principle is often summarized as “neurons that fire together wire together.”

That alone does not reproduce deep credit assignment. A learning network needs a way to connect an output error with earlier internal activity.

SAM combines local interactions into a motif capable of processing such relationships. The motif’s connections are not inserted as a completed computational diagram.

They begin with random values or random microscopic structure. Repeated plasticity changes organize them into a more useful configuration.

The paper says the resulting dynamics provably approximate generalized stochastic gradient descent. “Provably” applies to the formal model and its assumptions, not to biological brains.

This boundary is essential. A mathematical proof can establish how a defined system behaves. It cannot establish that cortical tissue implements the same system.

The hierarchy also matters. One motif supplies a compact learning structure, while repeated motifs connect learning across layers.

That layered composition addresses a central weakness of purely local rules. A single local adjustment can be plausible yet fail to coordinate useful learning across a deep network.

In the authors’ model, coordination emerges through interaction among motifs. The network can therefore approach the behavior expected from gradient-based optimization without directly installing a standard backpropagation circuit.

The process still contains an instructive signal. Learning is not unsupervised merely because the circuit organization self-assembles.

Instead, the system changes how instructional information reaches and reorganizes the network. Local units respond under shared rules rather than receiving individually calculated design instructions.

This is the most important reversal in the study. Randomness is not treated only as noise that engineers must overcome.

Random initial connectivity supplies raw material. Plasticity then selects, strengthens, weakens, and coordinates relationships that support a useful learning pattern.

Biology often develops structure through comparable broad principles. Cells follow local chemical and electrical rules while tissues acquire larger organization.

That analogy motivates the work, but it remains an analogy. The researchers modeled neuron-like units rather than observing SAM formation in living neural circuits.

Recent experimental work has strengthened the case that dendrites carry structured teaching signals. In 2026, researchers reported instructive dendritic signals in cortical neurons.

Such evidence can support the plausibility of local, compartmental learning mechanisms. It does not independently verify the specific self-assembling motif proposed here.

The MIT paper turns this gap into testable questions. Researchers can look for motif-like connectivity, structured feedback, and plasticity patterns predicted by the model.

They can also ask whether an emerging motif remains stable after learning conditions change. A circuit that forms once but collapses across initializations would offer limited explanatory value.

The system’s public code release is therefore important. It includes benchmark implementations and a later implementation used for theory and alignment diagnostics.

Independent researchers can vary random seeds, connectivity, learning rates, and task distributions. They can test whether the same qualitative motif reliably emerges.

They can also compare the learned updates with exact gradients. Alignment matters because similar final accuracy can hide very different optimization dynamics.

For developers, this is a useful distinction between architecture search and self-organized learning. Architecture search evaluates alternative designs using an outer optimization process.

SAM instead places organization inside the modeled network dynamics. The local rules cause a learning circuit to take shape during interaction.

The designer has not disappeared. Design has moved downward, from specifying the final circuit toward specifying the conditions that generate it.

That reframing connects the research with a broader engineering pattern. Complex systems can become easier to adapt when designers specify composable rules instead of every final relationship.

A similar principle appears in knowledge work. A technical knowledge base becomes more useful when relationships emerge across accumulated material rather than relying on one fixed folder structure.

The analogy should not be taken literally. Documents do not change like synapses. Both cases simply illustrate the difference between fixed organization and organization that develops from repeated interactions.

Similar Performance Does Not Mean a Backpropagation Replacement

The reported comparison establishes research credibility, but it does not establish readiness for large models, biological validity, or superior efficiency.

The authors say their self-organizing networks reached performance similar to the backpropagation-trained networks tested. That is an important result because local learning schemes often lose accuracy as tasks become more demanding.

However, “backpropagation-level performance” needs careful interpretation. It refers to the experiments and network configurations evaluated in the paper.

It does not mean the method matches transformer training across language, multimodal reasoning, or long-horizon reinforcement learning. No such result appears in the cited paper or MIT report.

It also does not mean MIT self-assembling neural networks learned without human choices. Researchers selected the model, plasticity rules, tasks, objective conditions, and evaluation method.

The self-assembly claim concerns internal learning machinery. It does not imply open-ended autonomy.

A second uncertainty is sensitivity to initial connectivity. The model begins with randomness, so reproducibility across seeds becomes central.

A convincing self-organizing mechanism should produce comparable functional motifs from many plausible starting points. Occasional successful assembly would be much weaker evidence.

Researchers should therefore examine distributions, not only average accuracy. Failure rates, variance, convergence time, and motif consistency will matter.

A third uncertainty concerns scale. Local rules can become computationally expensive when simulated on conventional hardware.

Biological locality does not automatically produce digital efficiency. A process that avoids exact backpropagation may require more steps, additional state, or denser interactions.

The study does not establish a lower training bill. It also does not show reduced energy use on neuromorphic hardware.

A fourth issue is task breadth. Image classification provides a controlled setting for measuring optimization and representation learning.

Modern AI systems also handle sequence modeling, retrieval, planning, generation, and delayed feedback. Those tasks create different credit-assignment problems.

A method can match backpropagation on compact classifiers while struggling with recurrent or attention-based systems. Extension beyond demonstrated tasks remains open.

A fifth uncertainty is biological evidence. The paper’s mechanism is inspired by synaptic plasticity and cortical organization, but it remains computational.

Liao explicitly cautions that a self-organizing network does not necessarily learn as the brain does. Self-organization is a starting criterion for plausibility, not biological confirmation.

That restraint strengthens the result. It distinguishes a testable model from a declaration that a major neuroscience problem has been solved.

Future experiments would need to identify biological counterparts for the motif’s components. They would also need to observe the predicted organization during learning.

Researchers could then perturb specific feedback paths or plasticity processes. If learning fails in the predicted way, the model would gain empirical support.

Without that evidence, SAM remains one mechanism capable of producing the desired behavior. Other local mechanisms might reach similar gradient-like dynamics.

The comparison with backpropagation also contains an important asymmetry. Backpropagation is judged mainly as an engineering method, while SAM is judged partly as an explanatory model.

A method can lose some efficiency yet remain scientifically valuable if it explains how biological organization might arise. Conversely, matching accuracy alone cannot validate its biological interpretation.

This difference shapes who should care now. Neuroscientists gain a formal hypothesis connecting local plasticity with organized credit assignment.

Machine learning researchers gain another approach to local learning and emergent optimization. Hardware researchers gain a possible design principle for physically distributed learning systems.

Enterprise AI buyers should not expect an immediate change in commercial models. The work does not announce a new foundation model, training service, or deployment framework.

Developers should also resist translating the result into claims about self-improving AI. The model learns within externally chosen tasks and rules.

The pressure on backpropagation is therefore intellectual before it is commercial. The study shows that useful gradient machinery need not be explicitly installed in its final form.

Whether that principle becomes technically competitive depends on evidence the current paper does not yet provide. Scale, stability, efficiency, and task diversity remain decisive tests.

Three Signals Will Show Whether the Result Travels

The next phase should test reproducible assembly first, harder tasks second, and biological predictions third.

The first signal is replication across random initial conditions. Independent teams should run the released implementation with broader seed ranges and connectivity patterns.

If stable motifs repeatedly emerge, the paper’s self-organization claim becomes stronger. If success depends on narrow initial conditions, the claim becomes less general.

Replication should track more than final classification accuracy. Researchers should report convergence, update alignment, motif structure, and unsuccessful runs.

They should also test whether the motif survives reasonable changes in hyperparameters. A self-assembling mechanism should not require precise manual tuning that effectively recreates circuit design.

The second signal is performance beyond the demonstrated classification settings. Larger visual tasks, sequential problems, and delayed credit assignment would provide harder tests.

Success on those tasks would show that MIT neural networks versus backpropagation is not limited to compact benchmark conditions. Persistent gaps would confine the method to a narrower explanatory role.

Compute measurements should accompany those comparisons. Accuracy without training time, memory use, and energy data cannot establish engineering competitiveness.

Researchers should also compare the method with other biologically motivated approaches. Standard backpropagation is an essential baseline, but it is not the only relevant one.

Feedback alignment, predictive coding, equilibrium methods, and local-error systems address overlapping questions. SAM’s distinct advantage should come from assembling its learning machinery, not merely using a different update formula.

The third signal is experimental neuroscience evidence. Researchers should search for the predicted motifs and learning-dependent changes in real circuits.

The strongest evidence would connect local heterosynaptic rules with emerging structures that perform identifiable credit-assignment functions. That would require measurement before, during, and after learning.

Perturbation would make the case stronger. Disrupting the proposed local interaction should prevent motif organization or change learning in a predicted direction.

Evidence that cortical circuits carry structured instructive signals would support the broader research program. It would not automatically confirm SAM’s exact mechanism.

A failure to find the motif would also be informative. The mathematical result could remain valid while its biological interpretation weakens.

These signals create a clear sequence. Reproducible computation tests whether the mechanism is stable. Harder benchmarks test whether it generalizes.

Biological experiments test whether the mechanism explains nature rather than only supplying an alternative algorithm. Each stage can strengthen or narrow the paper’s importance.

Readers should also watch how the authors separate the 2024 and 2025 benchmark implementation from the 2026 theory implementation. Differences between those versions may reveal which components drive performance.

Transparent ablations would help. An ablation removes or alters one component to show whether the claimed mechanism actually depends on it.

For example, researchers can weaken heterosynaptic interactions or prevent structural organization. If performance remains unchanged, the motif may not explain the result.

They can also freeze an assembled motif and test new tasks. That experiment would show whether the machinery becomes reusable or must reorganize for each setting.

Another useful test would change the data distribution after assembly. Stable adaptation would make the mechanism more relevant to continual learning.

Catastrophic forgetting, where new learning damages older knowledge, is a major challenge for neural systems. The current reporting does not establish that SAM improves it.

The paper’s relevance will therefore depend on disciplined follow-up, not dramatic extrapolation. Its strongest claim already deserves attention without being inflated.

MIT self-assembling neural networks show that local rules can generate a learning structure whose dynamics approximate gradient optimization. They offer a concrete bridge between circuit formation and learning.

They do not yet show that the brain uses this bridge. They also do not show that today’s AI labs should replace backpropagation.

The immediate action for researchers is straightforward: reproduce the assembly process, stress its stability, and test its biological predictions. For everyone else, the key question is equally clear.

Can a learning system keep organizing useful machinery when its starting connections, tasks, and physical constraints change, or does the result depend on the laboratory conditions that first revealed it?