Active learning of regular languages as an approach to neural language models verification.
This work tackles the general problem of verifying the behavior of sequence processing neural networks, specifically neural acceptors and neural language models. The contribution is a framework for extracting formal abstractions of the networks under analysis and verifying whether they satisfy given...
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | doctoralThesis |
| Sprache: | Englisch |
| Veröffentlicht: |
2024
|
| Online-Zugang: | https://hdl.handle.net/20.500.12008/45940 |
| Tags: |
Keine Tags, Fügen Sie das erste Tag hinzu!
|
| Zusammenfassung: | This work tackles the general problem of verifying the behavior of sequence processing neural networks, specifically neural acceptors and neural language models. The contribution is a framework for extracting formal abstractions of the networks under analysis and verifying whether they satisfy given requirements. This process steps on two core ideas: 1) treating the neural network as a black box, and 2) using a probabilistic framework to analyze how much the extracted model approximates the original one. For this matter, a series of active learning algorithms and techniques are proposed, developed and analyzed. For the case of neural acceptors a procedure for checking properties of neural networks is presented. This approach is able to check properties without explicitly building representations of the network. We show that this approach offers better guarantees and is more efficient than post-learning verification where the property is checked on a learned model of the network. Besides, it does not require resorting to an external decision procedure for verification nor fixing a specific requirement specification formalism. For the case of neural language models, a learning algorithm based on a congruence over strings which is parameterized by an equivalence relation over probability distributions is presented. The learning algorithm is implemented using a tree data structure and shown to be empirically more efficient than reference techniques. |
|---|