Skip to main content
Talk

Towards Making Neural Networks Interpretable: Feature Visualisation from Deep Neural Networks

On December 13, 2017, Dr. Omar Costilla Reyes spoke on feature visualisation from deep neural networks at the Data Visualisation Summit in Manchester. The talk, “Towards making neural networks interpretable,” drew on models of tomography floor-sensor data.

Title slide of the talk presented at the Data Visualisation Summit 2017 in Manchester, December 13, 2017. Credit: University of Manchester; image courtesy of Dr. Omar Costilla Reyes.
The talk on feature visualization in progress at the Data Visualisation Summit 2017, Manchester, December 13, 2017. Credit: Image courtesy of Dr. Omar Costilla Reyes.

A talk between two commitments

On December 13, 2017, Dr. Omar Costilla Reyes, then a researcher at the University of Manchester, presented at the Data Visualisation Summit in Manchester. He was taking a break from an Alan Turing Institute Data Study Group in London, an intensive week in which participants work on data-science challenges, and he returned to London the same day. He had announced the talk the week before under the working title “Visualization of Deep Neural Networks for Learning Spatio-Temporal Features from Tomography Sensors,” and delivered it as “Towards making neural networks interpretable: feature visualisation from deep neural networks.” Afterward he reported plenty of questions from the audience.

Looking inside a neural network

A deep neural network learns its own internal features from data, and because those features are numbers spread across many layers, such networks are often described as black boxes. Feature visualization is a family of techniques for turning those internal values into pictures, for example by displaying what each learned filter responds to in an input. The talk drew on networks trained on spatio-temporal signals from tomography-based floor sensors, which record how footsteps press on a floor over space and time. For a data-visualization audience, the point is that a model’s output is easier to evaluate when people can inspect what shaped it: the pictures help compare learned patterns, form hypotheses about model behavior, and judge whether a model relies on meaningful structure in the signal.

Connection to published research

The visualizations came from the work published as “Deep neural networks for learning spatio-temporal features from tomography sensors” (IEEE Transactions on Industrial Electronics, 2017), in which the authors also visualize the learned filter maps. IEEE recognized the paper as the journal’s top-downloaded paper in January 2018. The interest in making models inspectable runs through his later research on interpretable machine learning and neurosymbolic programming.

Topics