A new open-source library wants your neural network to stop using the same generic neuron for every job.
TNLearn is a Python package that automates the construction of what its creators call "task-based neurons" - neuron designs tailored to a specific problem rather than a single, one-size-fits-all unit repeated across a network. The package handles building these custom neurons and assembling them into trainable networks, with documentation, an API reference, and example code published alongside the release. It's built on PyTorch and is now listed as a project in that framework's ecosystem. The code is available on GitHub under an open-source license.
The pitch borrows a page from neuroscience: brains use different neuron types for different tasks, not one universal building block, and the researchers argue software should follow suit. That's a real departure from the dominant trend in deep learning, where progress has mostly come from stacking more identical neurons into bigger architectures rather than redesigning the neuron itself. Whether task-specific neurons meaningfully outperform standard ones on real workloads isn't demonstrated in this release - the package is a tool for building and testing the idea, not proof it works better.
In a field where most "innovation" is a bigger version of the same transformer block, a package built around rethinking the neuron itself is at least asking a different question - even if the answer is still pending.