Image Classification via Attribute Detection Kylie McCarty Dr. Gong Abdullah Jamal
New Challenges
What this means... Getting good accuracy just learning to always predict 0 for every attribute – despite a seemingly good score, not meaningful results if we want to go on to use these attribute predictions for classification Proposed solutions: Find more meaningful metric to judge performance Tweak the model to better learn to predict 1's AUC = area under
How? Weighted Sigmoid Cross Entropy Loss Layer Give more weight to the positive examples to encourage the net to learn meaningful positive predictions Additional performance metrics: AUC: Area Under the ROC Curve Plots false positive and true positive rate MAP: Mean Average Precision Score Area under the precision-recall curve