Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On the Interpretability of Quantum Neural Networks

Lirandë Pira, Chris Ferrie

TL;DR

This work explores the interpretability of quantum neural networks using local model-agnostic interpretability measures commonly utilized for classical neural networks, and generalizes a classical technique called LIME, introducing Q-LIME, which produces explanations of quantum neural networks.

Abstract

Interpretability of artificial intelligence (AI) methods, particularly deep neural networks, is of great interest. This heightened focus stems from the widespread use of AI-backed systems. These systems, often relying on intricate neural architectures, can exhibit behavior that is challenging to explain and comprehend. The interpretability of such models is a crucial component of building trusted systems. Many methods exist to approach this problem, but they do not apply straightforwardly to the quantum setting. Here, we explore the interpretability of quantum neural networks using local model-agnostic interpretability measures commonly utilized for classical neural networks. Following this analysis, we generalize a classical technique called LIME, introducing Q-LIME, which produces explanations of quantum neural networks. A feature of our explanations is the delineation of the region in which data samples have been given a random label, likely subjects of inherently random quantum measurements. We view this as a step toward understanding how to build responsible and accountable quantum AI models.

On the Interpretability of Quantum Neural Networks

TL;DR

This work explores the interpretability of quantum neural networks using local model-agnostic interpretability measures commonly utilized for classical neural networks, and generalizes a classical technique called LIME, introducing Q-LIME, which produces explanations of quantum neural networks.

Abstract

Interpretability of artificial intelligence (AI) methods, particularly deep neural networks, is of great interest. This heightened focus stems from the widespread use of AI-backed systems. These systems, often relying on intricate neural architectures, can exhibit behavior that is challenging to explain and comprehend. The interpretability of such models is a crucial component of building trusted systems. Many methods exist to approach this problem, but they do not apply straightforwardly to the quantum setting. Here, we explore the interpretability of quantum neural networks using local model-agnostic interpretability measures commonly utilized for classical neural networks. Following this analysis, we generalize a classical technique called LIME, introducing Q-LIME, which produces explanations of quantum neural networks. A feature of our explanations is the delineation of the region in which data samples have been given a random label, likely subjects of inherently random quantum measurements. We view this as a step toward understanding how to build responsible and accountable quantum AI models.
Paper Structure (10 sections, 4 equations, 4 figures, 3 algorithms)

This paper contains 10 sections, 4 equations, 4 figures, 3 algorithms.

Table of Contents

  1. Introduction
  2. Interpretability in AI
  3. Taxonomy
  4. Interpretability of neural networks
  5. Local interpretable methods
  6. The case for quantum AI interpretability
  7. Quantum LIME: Probabilistic local interpretability
  8. Local region of indecision
  9. Numerical experiments for interpreting QNNs
  10. Discussion & Future Work

Figures (4)

  • Figure 1: Categorization of interpretability techniques as they apply to classical and quantum resources. Here, the well-known QML diagram represents data and an algorithm or device, which can be classical (C) or quantum (Q) in four different scenarios. We consider a reformulation of interpretable techniques to be required in the CQ scenario. In the QC and QQ quadrants, the design of explicitly quantum interpretable methods may be required. The scope of this paper covers CQ approaches.
  • Figure 2: Depiction of the concept of the local region of indecision. The space within the dashed lines represents the region in the decision space where data samples exhibit ambiguous classification due to randomness. The figure showcases a two-class classification task in a two-dimensional space with two features represented along the horizontal and vertical axes. Here, $\epsilon$ is the pre-defined threshold. Likely, data samples, either inside or close to the band, are interpreted as randomly assigned.
  • Figure 3: The classification uncertainty for a chosen data sample (yellow). The two shadings represent the decision boundary of the QNN, which is clearly randomly defined. The synthetic data and the produced surrogate linear decision boundary are shown for a single instance of QNN labels. (top) A clear example of an undefined classification result for the selected data point. (bottom) Decision boundary unambiguously separates the who classes with respect to the selected data point.
  • Figure 4: The approximated local region of indecision. (top) An example of a marked data point that lies on the local region of indecision. (bottom) A data point that is outside of this region can be assessed for interpretability as per the interpretable techniques.