This is a Keras-Pennylane implementation of MNIST classification using classical and quantum layers, inspired by "Continuous-variable quantum neural networks" Physics Review 1 033063 (2019). https://arxiv.org/pdf/1806.06871v1.pdf The model in the paper is for binary classification, hence there is a slight modification to the proposed model.
The modification takes advantage of the measurement flexibility not present in the qubit model. A quantum state in photonic quantum computing is an infitie sum of density matrices of number basis as shown in this expression.
The notion of cutoff dimension allows for the flexibility of approximating the true quantum state with a desired number of basis. When using 2 qumodes, the probability measurement of the computational basis (density matrix) returns a vector of length (cutoff dimension)^2. By setting cutoff dimension to 4, we get vectors of length 4^2 = 16. Viewing each element of the vector as the probability of finding 0 - 9 with 6 irrelevant entries allows us to use it as label prediction. We one-hot encode the labels and pad 6 zeros for each label to match the output vector of our hybrid network.
After 100 epochs on 600 training samples,
Loss:
Accuracy:
