How Can a Neural Network Reveal Hidden Defects with Few Tests?

Imagine having to find a few defects in a huge collection of items, but you only have a limited number of group tests. Instead of checking each item individually, you mix several items together and the test tells you how many defects are inside. This is the Quantitative Group Testing (QGT) problem. Researchers have found that a simple neural network, a multi‑layer perceptron, can decode these noisy results with high accuracy. After training, the network converts the measurement vector into a binary list indicating which items are defective. Even more, by examining the network's Jacobian, scientists extracted the pooling structure linking items to tests, proving the network didn’t just memorize examples but truly understood the underlying combinatorial relationships. Despite bounded noise, the model stays robust, recovering defects with few errors. This shows that standard feed‑forward architectures can learn verifiable inverse mappings in complex combinatorial problems, opening doors to new uses in diagnostics, cybersecurity, and resource optimization.

Quantitative Group Testing (QGT) generalizes classic Group Testing by returning the exact count of defectives in each pool. Formally, we have N items, K of which are defective (K≪N). A pooling matrix of size M×N is defined, where entry (i,j) is 1 if item j participates in test i. The test outcome is y_i=∑_{j=1}^N A_{ij} x_j + η_i, with x_j∈{0,1} indicating defectiveness and η_i representing noise. The goal is to reconstruct x from y and A. Traditional solutions use ℓ1‑minimization or bipartite graph algorithms. In this study, the authors trained a multi‑layer perceptron (MLP) with ReLU activations and sufficiently wide hidden layers to approximate the inverse map f^{-1}: y↦x. The loss function is cross‑entropy between predictions and ground truth, with L2 regularization for stability. To test generalization, sparse bounded perturbations were introduced (‖η‖_0≤s, ‖η‖_∞≤ε). Results show the network maintains an error rate below 2% even for s=5 and ε=0.1. Moreover, by computing the Jacobian J=∂f/∂y, the pooling matrix A can be recovered up to a permutation, as non‑zero derivatives correspond to test‑item connections. This provides verifiability: the learned structure is interpretable and comparable to the true configuration. The implication is that simple neural networks can solve combinatorial recovery problems with interpretability guarantees, reducing the need for specialized algorithms and enabling scalable solutions for large N.