Backpropagation is elegant, efficient, and deeply entrenched. But if you spend any time with the mathematics of how biological brains actually learn, you start to notice that backpropagation, for all its power, does something brains almost certainly do not do: send error signals backward through every synapse in perfect, synchronized lockstep. The Reddit user asking these questions is trained in physics and biophysics, and they have noticed exactly this gap. Their instinct to explore alternatives like predictive coding, Hebbian learning, and Oja's rule is not a philosophical indulgence. It points toward a genuine structural limitation in how we currently train machines, and toward problems where that limitation actually matters.
Let's be precise about where local learning could win. In molecular dynamics and experimental biophysics, data arrives as continuous, high-dimensional streams from simulations or sensors. Unsupervised learning on such data is not a nice-to-have; it is often the only option because labeled states may not exist or may change with every new experiment. Backpropagation demands a fixed loss function and a carefully batched offline process. Local learning rules, which update weights based on information available at each neuron or synapse, can run as data arrives, adapting without replaying old examples. That is not a minor convenience. For someone trying to discover hidden collective variables in a simulation trajectory, a network that can learn on the fly and preserve older representations without catastrophic forgetting is not just interesting, it may be the only practical architecture.
This connects directly to work we have explored before on Smarter resource use: how batching transforms LLM performance. Batching is a workaround for backpropagation's need for global gradient information across many examples. Local learning rules side-step that need entirely, which suggests they could unlock efficiency gains in environments where batching is difficult, such as edge devices or real-time sensors. And the question about parameter efficiency is also relevant: if local rules can discover latent structure without the overhead of a global optimization pass, they may offer a different trade-off than methods like those in Train a neural network with 57x fewer parameters using this new PyTorch library, which reduce parameters at the cost of more complex latent optimization. Neither approach is universally better, but for unsupervised discovery in streaming data, local learning has a natural fit that parameter-count tricks do not.
The maturity question is fair. Predictive coding has been studied for decades, and we know it can approximate backpropagation under certain conditions. But approximations that are "close enough" for image classification are not necessarily "close enough" for discovering rare conformational states in a protein simulation. The real open problem is not whether local learning can match backpropagation on benchmarks. It is whether the inductive biases of local rules, especially their reliance on local prediction errors and correlations, produce representations that are more interpretable or more aligned with the latent physics of a system. That is a hypothesis worth testing with empirical comparisons on real scientific data. The answer will not be a universal yes or no, but it will tell us where to invest.