Teacher-Student Curriculum Learning.

We propose Teacher-Student Curriculum Learning (TSCL), a framework for automatic curriculum learning, where the Student tries to learn a complex task, and the Teacher automatically chooses subtasks from a given set for the Student to train on. We describe a family of Teacher algorithms that rely on the intuition that the Student should practice more those tasks on which it makes the fastest progress, i.e., where the slope of the learning curve is highest. In addition, the Teacher algorithms address the problem of forgetting by also choosing tasks where the Student's performance is getting worse. We demonstrate that TSCL matches or surpasses the results of carefully hand-crafted curricula in two tasks: addition of decimal numbers with long short-term memory (LSTM) and navigation in Minecraft. Our automatically ordered curriculum of submazes enabled to solve a Minecraft maze that could not be solved at all when training directly on that maze, and the learning was an order of magnitude faster than a uniform sampling of those submazes.

Pulmonary nodule detection in CT scans with equivariant CNNs.

Convolutional Neural Networks (CNNs) require a large amount of annotated data to learn from, which is often difficult to obtain for medical imaging problems. In this work we show that the sample complexity of CNNs can be significantly improved by using 3D roto-translation group convolutions instead of standard translational convolutions. 3D CNNs with group convolutions (3D G-CNNs) were applied to the problem of false positive reduction for pulmonary nodule detection in CT scans, and proved to be substantially more effective in terms of accuracy, sensitivity to malignant nodules, and speed of convergence compared to a strong and comparable baseline architecture with regular convolutions, extensive data augmentation and a similar number of parameters. For every dataset size tested, the G-CNN achieved a FROC score close to the CNN trained on ten times more data.

3D scattering transforms for disease classification in neuroimaging.

Classifying neurodegenerative brain diseases in MRI aims at correctly assigning discrete labels to MRI scans. Such labels usually refer to a diagnostic decision a learner infers based on what it has learned from a training sample of MRI scans. Classification from MRI voxels separately typically does not provide independent evidence towards or against a class; the information relevant for classification is only present in the form of complicated multivariate patterns (or "features"). Deep learning solves this problem by learning a sequence of non-linear transformations that result in feature representations that are better suited to classification. Such learned features have been shown to drastically outperform hand-engineered features in computer vision and audio analysis domains. However, applying the deep learning approach to the task of MRI classification is extremely challenging, because it requires a very large amount of data which is currently not available. We propose to instead use a three dimensional scattering transform, which resembles a deep convolutional neural network but has no learnable parameters. Furthermore, the scattering transform linearizes diffeomorphisms (due to e.g. residual anatomical variability in MRI scans), making the different disease states more easily separable using a linear classifier. In experiments on brain morphometry in Alzheimer's disease, and on white matter microstructural damage in HIV, scattering representations are shown to be highly effective for the task of disease classification. For instance, in semi-supervised learning of progressive versus stable MCI, we reach an accuracy of 82.7%. We also present a visualization method to highlight areas that provide evidence for or against a certain class, both on an individual and group level.