Tanya Schmah

Kevin Raina, Tanya Schmah

Out-of-Distribution (OOD) detection is critical to AI reliability and safety, yet in many practical settings, only a limited amount of training data is available. Bayesian Neural Networks (BNNs) are a promising class of model on which to base OOD detection, because they explicitly represent epistemic (i.e. model) uncertainty. In the small training data regime, BNNs are especially valuable because they can incorporate prior model information. We introduce a new family of Bayesian posthoc OOD scores based on expected logit vectors, and compare 5 Bayesian and 4 deterministic posthoc OOD scores. Experiments on MNIST and CIFAR-10 In-Distributions, with 5000 training samples or less, show that the Bayesian methods outperform corresponding deterministic methods.

Aditya Ohri, Tanya Schmah

We have implemented a machine translation system, the PolyMath Translator, for LaTeX documents containing mathematical text. The current implementation translates English LaTeX to French LaTeX, attaining a BLEU score of 53.5 on a held-out test corpus of mathematical sentences. It produces LaTeX documents that can be compiled to PDF without further editing. The system first converts the body of an input LaTeX document into English sentences containing math tokens, using the pandoc universal document converter to parse LaTeX input. We have trained a Transformer-based translator model, using OpenNMT, on a combined corpus containing a small proportion of domain-specific sentences. Our full system uses both this Transformer model and Google Translate, the latter being used as a backup to better handle linguistic features that do not appear in our training dataset. If the Transformer model does not have confidence in its translation, as determined by a high perplexity score, then we use Google Translate with a custom glossary. This backup was used 26% of the time on our test corpus of mathematical sentences. The PolyMath Translator is available as a web service at www.polymathtrans.ai.

Kevin Raina, Uladzimir Yahorau, Tanya Schmah

Brain lesions, including stroke and tumours, have a high degree of variability in terms of location, size, intensity and form, making automatic segmentation difficult. We propose an improvement to existing segmentation methods by exploiting the bilateral quasi-symmetry of healthy brains, which breaks down when lesions are present. Specifically, we use nonlinear registration of a neuroimage to a reflected version of itself ("reflective registration") to determine for each voxel its homologous (corresponding) voxel in the other hemisphere. A patch around the homologous voxel is added as a set of new features to the segmentation algorithm. To evaluate this method, we implemented two different CNN-based multimodal MRI stroke lesion segmentation algorithms, and then augmented them by adding extra symmetry features using the reflective registration method described above. For each architecture, we compared the performance with and without symmetry augmentation, on the SISS Training dataset of the Ischemic Stroke Lesion Segmentation Challenge (ISLES) 2015 challenge. Using affine reflective registration improves performance over baseline, but nonlinear reflective registration gives significantly better results: an improvement in Dice coefficient of 13 percentage points over baseline for one architecture and 9 points for the other. We argue for the broad applicability of adding symmetric features to existing segmentation algorithms, specifically using nonlinear, template-free methods.

Dongyang Kuang, Tanya Schmah

We present a new unsupervised learning algorithm, "FAIM", for 3D medical image registration. With a different architecture than the popular "U-net", the network takes a pair of full image volumes and predicts the displacement fields needed to register source to target. Compared with "U-net" based registration networks such as VoxelMorph, FAIM has fewer trainable parameters but can achieve higher registration accuracy as judged by Dice score on region labels in the Mindboggle-101 dataset. Moreover, with the proposed penalty loss on negative Jacobian determinants, FAIM produces deformations with many fewer "foldings", i.e. regions of non-invertibility where the surface folds over itself. In our experiment, we varied the strength of this penalty and investigated changes in registration accuracy and non-invertibility in terms of number of "folding" locations. We found that FAIM is able to maintain both the advantages of higher accuracy and fewer "folding" locations over VoxelMorph, over a range of hyper-parameters (with the same values used for both networks). Further, when trading off registration accuracy for better invertibility, FAIM required less sacrifice of registration accuracy. Codes for this paper will be released upon publication.