Evaluating Attribution Methods using White-Box LSTMs
This addresses the challenge of reliably assessing interpretability tools for researchers, though it is incremental as it builds on existing evaluation concerns.
The paper tackled the problem of evaluating interpretability methods for neural networks by proposing a framework using manually constructed white-box networks with known behavior, and found that five attribution methods failed to produce expected explanations despite perfect task performance.
Interpretability methods for neural networks are difficult to evaluate because we do not understand the black-box models typically used to test them. This paper proposes a framework in which interpretability methods are evaluated using manually constructed networks, which we call white-box networks, whose behavior is understood a priori. We evaluate five methods for producing attribution heatmaps by applying them to white-box LSTM classifiers for tasks based on formal languages. Although our white-box classifiers solve their tasks perfectly and transparently, we find that all five attribution methods fail to produce the expected model explanations.