Feature overview

Training a deep learning model is an iterative process, and each iteration requires identification of areas of weakness in the model and retraining the model on relevant data that will improve or eliminate the area of weakness. The model analyze feature allows intuitive and efficient identification of areas of model weakness through visual presentation and interactive statistical panels to help identify samples of data matching the model weakness areas and searching for more samples to build a curated training set.

The platform presents an aggregated and visual view of model errors using model metrics like Intersection over Union (IOU), object detection confidence score, class predictions, the position of ground truth and prediction bounding box within the image.

The Statistics panel gives detailed statistics on the model performance like the precision-recall curve and confusion matrix heatmap. The analysis can be interactively done at different IOU and confidence threshold values. The cells/rows/columns/diagonal of the confusion matrix can be selected to drill down to specific areas of the confusion matrix and perform follow-on operations like similarity search and addition to resultsets.

The model analyze feature is available for classification, object detection and instance segmentation models.