Machine Learning In The Realm Of Quantum: The State-Of-The-Art, Challenges, Future Vision and Applications Of It

Can today's quantum computers learn to read handwritten digits? Not "will quantum one day change ML" — that's a keynote question. This one is small and testable: take MNIST, run part of the classifier on a quantum circuit, measure the gap against an ordinary network. My thesis partner and I built two deliberately different quantum learners to find out. The first slides a tiny 4-qubit circuit across each image the way a convolutional filter would — a quanvolution. The second encodes images into beams of light (squeezers, beamsplitters, Kerr gates) following Xanadu's continuous-variable recipe, the Kerr gate playing the role of the activation function, and trains the way a normal neural network does: gradients flow through the quantum circuit via the parameter-shift rule — evaluate the same circuit at θ±π/2, subtract, and you have an exact analytic derivative — the same mathematics running live in this site's hero and in lesson 6 of /learn.

PennyLane with the Keras plugin, 10-class MNIST, simulators only — no real qubits, so no decoherence: the easy setting. Model 1 (quanvolutional, after Henderson et al. 2019): 2×2 patches angle-encoded via RY rotations into 4 qubits — the pixel is the rotation angle — a random variational layer, and Pauli-Z expectations giving 4 feature channels, turning each 28×28 image into a 14×14×4 stack for a classical dense head; 6,000 training and 1,000 test images on default.qubit. One point I want stated plainly: Model 1's quantum circuit is a fixed, random feature extractor — drawn once, frozen, never trained; only the classical head learns. Model 2 (CV-QNN, after Killoran et al.) is the one that actually trains quantum weights: dense classical layers compress each image to 14 parameters driving a 2-qumode photonic circuit — squeezing, interferometers, displacement, Kerr nonlinearity — with 4 quantum layers (56 quantum parameters, 111,334 end-to-end), trained on a 700-sample subset with cutoff dimension 4, because simulating photonic circuits in Fock space is expensive enough that a bigger run wasn't practical.

The classical baselines won — and the margins are the finding. The quanvolutional model reached 92% test accuracy against a 96% classical CNN, and converged to its optimum loss faster than the classical model, even though the classical model hit peak accuracy first — a quantum-preprocessed feature stack was at least not obviously worse to optimize. The CV-QNN reached 72% against an 88% classical baseline, its validation curve sitting below the classical one for the whole run, noisier and slower to settle: training 56 parameters through a truncated photonic simulation was simply harder than training the equivalent classical weights. No quantum advantage is claimed, and none was observed — but the gate-based approach got within four points: the honest state of the NISQ era, measured directly, and closer than either the hype or the cynicism predicts.

The thesis taught me that the interesting question isn't "is quantum faster?" but "where does the encoding bottleneck bite?" A neural network eats a vector; a quantum circuit eats a quantum state, and getting classical pixels into that state — the angle rotations, the 14-number compression, the choice of squeezers and Kerr gates — dominated every design decision in both models. If I picked the thread back up I'd start there: trainable data encodings instead of Model 1's frozen random feature map. The full story — the two distinct failure modes, derivations, and run-notes — lives in the writeup. That lesson also shaped this site: the interactive curriculum at /learn exists because the encoding intuition took me months and a 68-page thesis to build, and a visitor can now get it in six scroll-stops.