The multi-chip ensemble VQC framework addresses key QML challenges on Noisy Intermediate-Scale Quantum (NISQ) devices
It partitions high-dimensional computations across multiple smaller quantum chips, each running a subcircuit, with classical aggregation of outputs
It partitions high-dimensional computations across multiple smaller quantum chips, each running a subcircuit, with classical aggregation of outputs
Each chip processes a subvector of the input data using its own quantum circuit and data encoding unitary
Classical aggregation combines the outputs from all chips using a function (e.g., weighted sum or shallow neural network) for final predictions
Training is hybrid quantum-classical, with parameters for all subcircuits optimized collectively via backpropagation and parameter-shift rules
Scalability is achieved horizontally by adding more chips, avoiding the need for larger single chips or deep circuits
The architecture mitigates the barren plateau problem by restricting entanglement within chips, increasing gradient variance and improving trainability
The controlled entanglement structure acts as implicit regularization, reducing overfitting and improving generalization
Noise resilience is enhanced by limiting circuit depth per chip and averaging uncorrelated noise across chips, reducing both error variance and bias
The framework is compatible with current and future modular quantum hardware, addressing limitations like sparse connectivity and limited qubit counts