Hybrid Quantum-Classical Neural Networks For Efficient Image Classification On Small Datasets

The modern deep learning models such as CNNs and Transformer models are very good at prediction but too big, require a truckload of computing resources, lots of data, and lots of memory so difficult to run on a budget laptop or a cheap smartphone.Quantum computing is an entirely different animal that is capable of handling information in large Hilbert spaces due to the ability of superposition and entanglement.I discuss a Hybrid Quantum Classical Neural Network (HQ-NN) in this paper, which combines the use of parameterised quantum circuits in a small stage of feature extraction, and a conventional fully-connected network in the actual predictions and optimisation.The quantum bit executes Aashna as an angle encoding of a flat image which the quantum encodes before executing a series of parameterised rotations together with CNOTs on the entangled bits and retrieves expectation values back as a classical feature-vector.The end-to-end training of the entire thing is a differentiable junction between the PennyLane and PyTorch implementations, and therefore the gradients flow straight across the quantum implementation.The new design and purely the quantum circuit outperform each other with an approximate 10 -percent difference in accuracy and a percentage difference of 20 -percent on general image benchmarks, which is a fraction of the number of parameters in a pure classical model.In even the case of adding simulated quantum noise, the predictions remain consistent to acceptable levels, indicating that this concept can be maintained on even real Noisy Intermediate-Scale Quantum (NISQ) machines.