Visualising state space representations of LSTM networks
E. M. Smith, James Edward Smith, Phil Legg, Susan Francis · UWE Research Repository (UWE Bristol) · 2018
Long Short-Term Memory (LSTM) networks have proven to be one of the most effective models for making predictions on sequence-based tasks. These models work by capturing, remembering, and forgetting information relevant to their future predictions. The non-linear complexity of the mechanisms involved in this process means we currently lack tools for achieving interpretability. Ideally, we want these models to provide an explanation of why they make a particular prediction, given a specific input. Researchers have explored the idea of interpreting LSTMs in specific contexts such as natural language processing or classification, but they put minimal focus on approaches which are generalisable across different applications. To alleviate this, in this work, we demonstrate a method which enables the interpretation and comparison of LSTM states during time series predictions. We show that by reducing the dimensionality of network states one can scalably visualise patterns and explain model behaviours.