Introduction to Grad-CAM Neural Network Visualization
As part of monitoring and managing your experiments, you might wish to visualize your data. Visualizing data assists in analyzing the experiment when solving classification issues.
One method of visualization supported in MissingLink is Grad-CAM (gradient class activation map) - an algorithm for using heat maps to visualize which part of the image provided to a model is important to it.
By using Grad-CAM with MissingLink, the dashboard can assist you in tuning the model by showing gradient-based areas of an image. This functionality can be combined with the test monitoring to enable a drill-down from the confusion matrix to the actual data points with their respective visualizations.
Grad-CAM is only applicable to classification problems. To deal with non-classification problems, see Visual Back Prop.
If you have performed basic integration with the MissingLink SDK, learn now how to generate Grad-CAM for neural network visualization: