Speaker
Description
TQML is an integrated quantum machine learning platform that spans the full workflow, from data to trained model, inside ordinary PyTorch and JAX. Quantum machine learning today covers a wide ansatz design space, from hardware-efficient and NISQ-style circuits through data re-uploading and tensor-network models to symmetry- and particle-number-preserving constructions such as fixed-Hamming-weight sectors. General-purpose simulators treat all of these as dense $2^{n}$ statevectors, which caps differentiable training near 25 to 30 qubits. TQML takes the opposite approach: it exposes the whole spectrum as composable machine-learning layers and backs them with a simulator that adapts its state representation to the structure of the chosen ansatz.
The platform shares a single Python namespace. Its layer library, tqnet, presents every quantum layer as both a PyTorch and a JAX module generated from one circuit recipe, declared by design axes rather than assembled gate by gate. Its simulator, tqsim, decouples what is computed from how it runs across Torch, JAX, and C++ backends and dense, tensor-network, and subspace representations; a deferred optimiser and auto-router accelerate structured circuits. Circuit analysis, quantum-aware optimisers, architecture suggestion, and benchmarking close the design loop.
The subspace engine simulates excitation-preserving circuits directly in the $\binom{n}{k}$-dimensional Hamming-weight sector. On this engine, TQML computes a full forward pass and gradient for circuits of up to 100 qubits in seconds on a laptop CPU, and trains a Hamming-weight-preserving quantum convolutional network to $86.6\%$ test accuracy on MNIST.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Hybrid quantum-classical approaches for learning and optimization |