Neural networks that keep their algorithm clear for texture recognition
Specific Algorithmic Interpretability of Neural Networks: A Case Study on Textures
Machine LearningArtificial Intelligence
Summary
Neural networks often work like black boxes where it’s hard to understand how they make decisions. This paper shows a way to build neural networks that start by exactly following a clear algorithm for recognizing textures. They fine-tune the network afterward but keep it close to the original understandable algorithm, so the network remains interpretable. Their method helps explain what the network is doing internally by linking learned parameters back to a known process.
What this means in practice
- •For computer vision engineers: Build texture classification models that maintain clear links to their underlying algorithms for easier debugging and trust.
- •For machine learning practitioners: Develop neural network models initialized with interpretable algorithms, helping control training changes and improve generalization reliability.
Tested on one dataset.
Authors
Yanglin Zhang, Anneke von Seeger, Gilad Lerman, Ron Levie
Abstract
We develop a principled framework for constructing neural networks whose specific parameter realizations admit an explicit algorithmic interpretation. Existing algorithm-inspired architectures can explain the computational structure of a network, yet after standard training the learned parameters need not retain a clear relation to the motivating algorithm. We address this gap as follows. First, we model each data point as a sample of a class-dependent stochastic process and assume that statistics of this process can be estimated from a single sample and these statistics are sufficient to distinguish the classes. We then construct a neural network whose initial parameters exactly implement an algorithm for estimating these statistics, making the network fully interpretable. To account for mismatch between the idealized model and real data, we fine-tune this network while controlling its deviation from the algorithmic initialization. The trained network hence roughly retains the interpretation of the initial network. A PAC-Bayesian analysis yields a uniform generalization bound whose complexity term scales with the fine-tuning radius, providing a statistical motivation for our approach. We instantiate the framework for texture classification using the scattering transform to estimate the discriminative statistics.