Speakers
Description
How quantum computing and neural network architectures can be meaningfully integrated is a central question in quantum artificial intelligence. Current approaches mainly follow two paradigms: variational quantum circuits, and quantum algorithms for the linear-algebraic primitives of classical neural networks. Here we propose a third route, grounded in the complexity-theoretic analysis of classical neural networks and in shallow quantum circuit complexity. We show that a neuron with $n$ inputs and $m$ bits of precision can be implemented in a model of computation where constant-depth quantum circuits are combined with measurements, classical processing, and feedforward. The resulting quantum depth is \emph{independent} of both $n$ and $m$. Our construction relies on constant-depth implementations of arithmetic and comparison primitives, including addition, multi-number addition, sorting, one-hot encoding, and monotone activation functions. For these primitives we give explicit resource counts. The same construction can be applied coherently in superposition, enabling superpositions over inputs or over multiple neurons evaluated on the same input. It also admits a distributed evaluation protocol, does not require QRAM, and does not introduce barren plateaus. We validate the construction through a hybrid classifier for molecular pose prediction of protein--ligand complexes. We compare classical simulations with executions on a trapped-ion quantum computer. This provides a new perspective on the interplay between classical neural computation and shallow quantum computation.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Theoretical foundations of quantum learning |