Abstract: A computer-implemented method for explaining an image classifier, the method comprising: receiving an initial image, the initial image having been wrongly classified by the image classifier; receiving an initial gradient of a function executed by the image classifier generated while classifying the initial image, the function being indicative of a probability for the initial image to belong to an initial class; converting the initial image into a latent vector, the latent vector being a representation of the initial image in a latent space; generating a plurality of perturbation vectors using the initial gradient of the function executed by the image classifier; combining the latent vector with each one of the plurality of perturbation vectors, thereby obtaining a plurality of modified vectors; for each one of the plurality of modified vectors, reconstructing a respective image, thereby obtaining a plurality of reconstructed images; transmitting the reconstructed images to the image classifier; for each one o

Abstract: A method for training a machine learning localization model to localize objects belonging to a given class within an image, the method comprising: receiving images each comprising objects of the given class; and for each image: receiving a heat map generated using the machine learning localization model; identifying proposals each corresponding to a potential object, each proposal having associated thereto an initial probability that the proposal corresponds to the potential object; for each proposal, correcting the initial probability using the heat map; selecting given ones of the proposals having a greatest corrected probability, thereby identifying object candidates; and calculating a loss for the machine learning localization model based on a location of the object candidates within the training image and the heat map; and providing the calculated loss to the machine learning localization model.

Abstract: A compatibility score generator implementing a neural network is trained for assessing compatibility of items. Elements of a feature vector representing each item and of a compatibility data structure indicating items considered compatible are retrieved. The neural network is trained using training data corresponding to the items and indicating compatibility between pairs of items. The compatibility data structure is modified by removing indications that items of a pair of items are compatible. An encoding function generating encoded representations for the items based on the compatibility data structure is evaluated. Encoded representations are provided to a decoder that learns a likelihood that the indication had been removed when modified. The neural network and the decoder are optimized based on a loss function that reflects the decoder's ability to correctly determine whether the indication had been removed. The encoded representations generate a compatibility score for at least two items of interest.

Abstract: There is provided a method and system for training an embedding model to perform relation predictions in a knowledge hypergraph to output a trained embedding model. A training dataset comprising tuples representing relations between entities in the knowledge hypergraph are received. The embedding model is trained to perform relation predictions for each given tuple from a subset of tuples in the training dataset by generating a respective entity vector for each entity and a respective relation matrix representing relations between the entities. The entity vectors and relation matrix are split into a plurality of windows, and interaction values between elements in each window are calculated. A relation score indicative of the relation in the given tuple being true is calculated. Parameters of the embedding model are updated based on the relation scores for the subset of tuples. The trained embedding model is then output.

Abstract: A computer-implemented method for explaining an image classifier, the method comprising: receiving an initial image, the initial image having been wrongly classified by the image classifier; receiving an initial gradient of a function executed by the image classifier generated while classifying the initial image, the function being indicative of a probability for the initial image to belong to an initial class; converting the initial image into a latent vector, the latent vector being a representation of the initial image in a latent space; generating a plurality of perturbation vectors using the initial gradient of the function executed by the image classifier; combining the latent vector with each one of the plurality of perturbation vectors, thereby obtaining a plurality of modified vectors; for each one of the plurality of modified vectors, reconstructing a respective image, thereby obtaining a plurality of reconstructed images; transmitting the reconstructed images to the image classifier; for each one o

Abstract: A method and a system for joint data augmentation and classification learning, where an augmentation network learns to perform transformations and a classification network is trained. A set of labelled images is received. During an inner loop iteration, an augmentation network applies a transformation on a given labelled image of the set to obtain a transformed image. The classification network classifies the transformed image to obtain a predicted class, and a training loss is determined based on the predicted class and the respective label. The parameters of the classification network are updated based on the classification loss. During an outer loop iteration, the classification network classifies another labelled image of the set to obtain another predicted class, and a validation loss is determined based on the other predicted class and the respective label. The parameters of the augmentation network are updated based on the validation loss.

Abstract: A method and a system for generating training images for training an instance segmentation machine learning algorithm (MLA). A set of image-level labelled images are received, where a given image is labelled with a label indicative of a presence of an object having an object class in the image. A classification MLA detects the object having the object class in each image. A class activation map (CAM) indicative of discriminative regions used by the classification MLA for detecting the object in each image is generated. A region proposal MLA is used to generate region proposals for each image. A pseudo mask of the respective object is generated based on the region proposals and the CAM, where a pseudo mask is indicative of pixels corresponding to the respective object class. The pseudo masks are used as a label with the image-level labelled images for training the instance segmentation MLA.

Abstract: A method and a system for joint data augmentation and classification learning, where an augmentation network learns to perform transformations and a classification network is trained. A set of labelled images is received. During an inner loop iteration, an augmentation network applies a transformation on a given labelled image of the set to obtain a transformed image. The classification network classifies the transformed image to obtain a predicted class, and a training loss is determined based on the predicted class and the respective label. The parameters of the classification network are updated based on the classification loss. During an outer loop iteration, the classification network classifies another labelled image of the set to obtain another predicted class, and a validation loss is determined based on the other predicted class and the respective label. The parameters of the augmentation network are updated based on the validation loss.

Abstract: A method and a system for generating training images for training an instance segmentation machine learning algorithm (MLA). A set of image-level labelled images are received, where a given image is labelled with a label indicative of a presence of an object having an object class in the image. A classification MLA detects the object having the object class in each image. A class activation map (CAM) indicative of discriminative regions used by the classification MLA for detecting the object in each image is generated. A region proposal MLA is used to generate region proposals for each image. A pseudo mask of the respective object is generated based on the region proposals and the CAM, where a pseudo mask is indicative of pixels corresponding to the respective object class. The pseudo masks are used as a label with the image-level labelled images for training the instance segmentation MLA.

Abstract: A compatibility score generator implementing a neural network is trained for assessing compatibility of items. Elements of a feature vector representing each item and of a compatibility data structure indicating items considered compatible are retrieved. The neural network is trained using training data corresponding to the items and indicating compatibility between pairs of items. The compatibility data structure is modified by removing indications that items of a pair of items are compatible. An encoding function generating encoded representations for the items based on the compatibility data structure is evaluated. Encoded representations are provided to a decoder that learns a likelihood that the indication had been removed when modified. The neural network and the decoder are optimized based on a loss function that reflects the decoder's ability to correctly determine whether the indication had been removed. The encoded representations generate a compatibility score for at least two items of interest.