Confidence From Invariance

This is the official TensorFlow implementation of the error and novelty detection method described in the arXiv paper Confidence from Invariance to Image Transformations by Yuval Bahat and Gregory Shakhnarovich.

The core idea behind this method is to base an image classification confidence measure on the variations in the classifier's outputs corresponding to different natural transformations of the image at hand. This is done by training a simple multi-layer-perceptron on a given classifier's outputs corresponding to the original image and its transformed versions. Note that our method does not require any access to the classifier model parameter, and instead operates on any given image classifier by treating it as a "black box". Please see the paper for more details.

If you find our work useful in your research or publication, please cite it:

@article{bahat2018confidence,
  title={Confidence from Invariance to Image Transformations},
  author={Bahat, Yuval and Shakhnarovich, Gregory},
  journal={arXiv preprint arXiv:1804.00657},
  year={2018}
}

---

General usage:

The TensorFlow transformations operators, which constitute the core of this method, are implemented in file Transformations.py. These operators can be incorporated for both error detection and novelty detection on any given pre-trained image classifier, as described in the paper. Be sure to apply the transformations directly on the images, before applying any standartization or whitening processes (e.g. subtracting images' mean). If the classifier at hand requires such operations, apply them after applying the transformations operator.

Example of error detection on a pre-trained CIFAR-10 classifier:

We provide here an example usage of error detection on a pre-trained CIFAR-10 classifier. The classifier model code was modified from the CIFAR-10 classifier by TensorFlow.

Requirements:

• TensorFlow

Detector training:

To train a detector, set TRANSFORMATIONS_LIST (in train_detector.py) to include the desired transformations (from the list in Transformations.py). Setting it to an empty list means the input to the detector will consist only the logits of the original input image, without any transformations. To start training, run

python train_detector.py -layers_widths <LAYERS_WIDTHS> -descriptor <MODEL_DESCRIPTOR> -train,

where the number of hidden layers and the number of channels in each layer in the detector is set using LAYERS_WIDTHS, and MODEL_DESCRIPTOR is a name used for saving the model and comparing ROC curves. For example,

python train.detector.py -layers_widths 70 70 -descriptor 70_70 -train

will train a detector with 2 hidden layers, each with 70 channels, and will use the name 70_70.

To resume training of a pre-trained detector use -resume_train instead of -train, and use neither flags when evaluating a pre-trained detector.

Features normalization:

Using the optional -data_normalization flag will normalize the feature vectors prior to feeding them into the detector.

Detector training and evaluation sets:

The detector is trained on a portion of the original validation set, by assigning this portion to be used as detector training images, while the rest of the images are used for its evaluation. In order to compare different detector configurations and compare to other methods, the same assignment is reused by saving it to ValidationSetSplit_<TRAIN_PORTION>.npz. This repository already includes an assignment for the default TRAIN_PORTION=0.5 split.

Saved models and figures:

Detector models are regularly saved into a sub-folder to allow later evaluation or further training. Further more, each time a detector is evaluated (either during training or evaluation runs), a Receiver Operating Characteristic (ROC) curve corresponding to the current detector is saved to a figure, comparing it with ROC curves corresponding to all previously trained detectors and to the Maximal Softmax Response (MSR) method.