On-the-fly calibration of an image classifier

A computer-implemented method of using a trained image classifier, the method comprising:

• accessing trained parameters of the image classifier, the image classifier being configured to determine a prediction vector for an input image, the prediction vector including class probabilities for multiple respective classes, the image classifier being trained by:
• accessing trainable parameters of the image classifier and a training dataset including multiple training images and corresponding class labels,
• performing one or more initial training steps of training the image classifier to obtain current parameter values for the trainable parameters,
• selecting a training image and a corresponding class label from the training dataset,
• determining a current label vector for training the image classifier on the training image by:
• applying the image classifier to the training image according to the current parameter values to obtain a prediction vector;
• inferring a predicted class of the multiple classes from the prediction vector;
• comparing the predicted class to the class label;
• and, when the predicted class is equal to the class label, determining the current label vector as a soft label based on the prediction vector, and
• performing a current training step of training the image classifier on the training image using the current label vector;
• obtaining an input image;
• (i) applying the image classifier to the input image to determine a prediction vector for the input image and outputting a prediction based on the prediction vector;
• and/or (ii) applying a part of the image classifier to the input image to obtain a feature space representation of the input image and outputting the feature space representation.

A system for training an image classifier, the image classifier being configured to determine a prediction vector for an input image, the prediction vector including class probabilities for multiple respective classes, the system comprising:

• a data interface for accessing a training dataset including multiple training images and corresponding class labels, and for accessing trainable parameters of the image classifier;
• and
• a processor subsystem configured to train the image classifier by:
• performing one or more initial training steps of training the image classifier to obtain current parameter values for the trainable parameters;
• selecting a training image and a corresponding class label from the training dataset;
• determining a current label vector for training the image classifier on the training image by:
• applying the image classifier to the training image according to the current parameter values to obtain a prediction vector;
• inferring a predicted class of the multiple classes from the prediction vector;
• comparing the predicted class to the class label;
• and, if the predicted class is equal to the class label, determining the current label vector as a soft label based on the prediction vector;
• and
• performing a current training step of training the image classifier on the training image using the current label vector.

A system for using an image classifier, the system comprising:

• a storage including comprising trained parameters of the image classifier, the image classifier being configured to determine a prediction vector for an input image, the prediction vector including class probabilities for multiple respective classes, the image classifier being trained by:
• accessing trainable parameters of the image classifier and a training dataset including multiple training images and corresponding class labels,
• performing one or more initial training steps of training the image classifier to obtain current parameter values for the trainable parameters,
• selecting a training image and a corresponding class label from the training dataset,
• determining a current label vector for training the image classifier on the training image by:
• applying the image classifier to the training image according to the current parameter values to obtain a prediction vector;
• inferring a predicted class of the multiple classes from the prediction vector;
• comparing the predicted class to the class label;
• and, when the predicted class is equal to the class label, determining the current label vector as a soft label based on the prediction vector, and
• performing a current training step of training the image classifier on the training image using the current label vector;
• a processor subsystem configured to:
• obtain an input image, and
• (i) apply the image classifier to the input image to determine a prediction vector for the input image and output a prediction based on the prediction vector;
• and/or (ii) apply a part of the image classifier to the input image to obtain a feature space representation of the input image and output the feature space representation.

A non-transitory computer-readable medium on which is stored a computer program training an image classifier, the image classifier being configured to determine a prediction vector for an input image, the prediction vector including class probabilities for multiple respective classes, the computer program, when executed by a processor system, causing the processor system to perform the following steps:

• accessing trainable parameters of the image classifier and a training dataset including multiple training images and corresponding class labels;
• performing one or more initial training steps of training the image classifier to obtain current parameter values for the trainable parameters;
• selecting a training image and a corresponding class label from the training dataset;
• determining a current label vector for training the image classifier on the training image by:
• applying the image classifier to the training image according to the current parameter values to obtain a prediction vector;
• inferring a predicted class of the multiple classes from the prediction vector;
• comparing the predicted class to the class label;
• and, when the predicted class is equal to the class label, determining the current label vector as a soft label based on the prediction vector;
• and
• performing a current training step of training the image classifier on the training image using the current label vector.