DISPLAYING:

Events happening on (October 15, 2026)

• 15
  Oct

  TRAINABILITY AND INTERPRETABILITY IN QUANTUM NEURAL NETWORKS: A SYMPLECTIC FRAMEWORK AND AN APPLICATION TO CREDIT SCORING

  Date: October 15, 2026
  Time: 11:00 AM
  Location: Conference Room (9-117), Department of Physics, SBASSE, LUMS

  Supervisor: Dr. Aeysha Khalique
  Student(s): Abdul Samad (2024-12-0006) MS Physics Student at LUMS

  Quantum neural networks (QNNs) promise new computational capabilities for near-term devices, but two separate barriers stand between that promise and real-world deployment: many circuit ansätze are difficult or impossible to train efficiently, and even a successfully trained QNN's decisions are hard to trust, since the post-hoc explanation tools built for classical neural networks do not transfer to quantum circuits without modification. This thesis addresses both barriers. The  first contribution, DyLA.jl, is a Julia library for computing the Dynamical Lie Algebra (DLA) generated by a quantum many-body system's Hamiltonian   the algebraic quantity that governs controllability and, under well-defined conditions, classical simulability and trainability via the barren-plateau phenomenon. Standard matrix-based methods cost O(N3L) time and O(N2L) memory per commutator for L qudits of local dimension N, rendering exact DLA construction intractable beyond modest system sizes. DyLA.jl instead encodes operators as points in a discrete symplectic vector space, reducing every commutator evaluation to an O(L), allocation-free integer computation. The library reproduces 28 published DLA complexity classes exactly, and extends DLA analysis to qudit (N ≥ 3) systems without constructing the full operator basis, revealing, for one 3-local qutrit interaction, a dimension-8 subalgebra isomorphic to su(3) hidden inside a nominally 6560-dimensional operator space. The second contribution, IQNN-CS, is a hybrid quantum-classical architecture for interpretable, multiclass credit scoring, evaluated on two real-world  financial datasets. Its central contribution is Inter-Class Attribution Alignment (ICAA), a new metric quantifying whether a QNN's per-class explanations are actually distinguishable. Under an identical architecture, the same model achieved 100% test accuracy on one dataset and, on a harder dataset, 77.3% accuracy alongside markedly less class-specific attribution structure - a difference invisible to accuracy or F1 score alone. Finally, this thesis argues that these two contributions are complementary stages of a single trust pipeline for QNNs - a DLA check before training, and an ICAA check after -  and proposes a concrete, falsifiable hypothesis and experiment connecting the two tools to the specific interpretability failure observed in this work.

  • All Events
  • Physics

  Read More