Training objective functions for confidence-calibrated predictions from neural networks

  • Derived loss functions for confidence-calibrated neural network predictions for image classification and object detection
  • Shifted from two-step post-processing calibration methods, benchmarking one-stage object detectors (FAIR’s RetinaNet)
  • Achieved improved calibration scores on image classification along with occasional improvement in model performance, with key observations and analysis being discussed in the informal report

Related