Quantum Computing: An AI Approach
Project Description
A hands-on quantum machine learning project (Colab + PennyLane) that builds a trainable quantum classifier from scratch. It goes deep on one application rather than broad: starting from a single “quantum neuron” (a rotation angle θ standing in for a classical weight), students scale up to parameterized multi-qubit circuits with entanglement, then train them with gradient descent using the parameter-shift rule. The capstone is a variational quantum classifier (VQC) trained on a non-linearly-separable dataset, followed by an honest head-to-head against classical baselines. The project deliberately surfaces the field’s central tension: on a classical simulator a quantum model has zero quantum advantage, so the real question is where, on actual hardware, quantum models might genuinely win. Basic ML background (loss functions, gradient descent, train/test splits) is recommended.
Dataset
The “circles” dataset: 200 synthetically generated 2D points arranged as two concentric rings (one ring class −1, the other class +1), e.g. via make_circles. Features are scaled to the range [0, π] — the informative range for RY angle encoding — then split into training and test sets. The rings are intentionally not linearly separable, so any linear model fails and a non-linear classifier is required.
Relevant Skills You May Apply
Python and NumPy, core machine learning concepts (neurons/weights, sigmoid, loss functions, gradient descent, learning rate, train/test split, classification, decision boundaries), and comfort running Colab notebooks.
Skills You May Gain
Quantum machine learning with PennyLane: the quantum neuron and rotation gates (RY/RX), parameterized/variational circuits, angle encoding of classical data, entanglement (CNOT) for model expressiveness, the parameter-shift rule (why backpropagation is impossible on a quantum state), training a VQC with Adam, and critically evaluating quantum vs. classical models and the limits of simulation.
Total Time
Approximately 8–12 hours, self-paced. The classifier training loop (120–200 steps) takes a few minutes per run, which is normal.
Milestones
Milestones
- The Quantum Neuron — rotation angle θ as a trainable weight; P(0)=cos²(θ/2); classical sigmoid vs. periodic quantum neuron; RY→RX and 50% crossover TODO.
- Parameterized Circuits — multi-qubit rotations with and without CNOT; heatmap sweep showing why entanglement adds expressiveness (written question).
- Teaching the Circuit — MSE loss, gradient descent, and the parameter-shift rule via
opt.step(). - A Real Quantum Classifier — angle encoding, trainable Rot layers, CNOT entanglement, 13 parameters; train on the circles dataset and inspect the non-linear decision boundary (written question + range(120)→range(200) TODO).
- Quantum vs. Classical — VQC compared against logistic regression and a small neural network; why a simulator yields no quantum advantage (written questions).
Deliverables
A completed notebook with the RY→RX / 50%-angle TODO answered (M1), the entanglement heatmap question answered (M2), the classifier trained and its decision boundary visualized (M4), the 120→200 step experiment run and explained, and written answers on the biggest practical disadvantage (training time) plus the short QML explanation (M5).
Before editing anything, all learners must open the notebook and select “File” → “Save a copy in Drive” so they are working on their own personal copy in their Google Drive. Each member of a team should make and edit their own copy rather than working in the shared original. Once finished, completed materials are submitted to the NSDC HQ team to receive a virtual certificate of completion.