William Yolland

Jarrod Haas, William Yolland, Bernhard Rabus

We propose a simple modification to standard ResNet architectures--L2 normalization over feature space--that substantially improves out-of-distribution (OoD) performance on the previously proposed Deep Deterministic Uncertainty (DDU) benchmark. We show that this change also induces early Neural Collapse (NC), an effect linked to better OoD performance. Our method achieves comparable or superior OoD detection scores and classification accuracy in a small fraction of the training time of the benchmark. Additionally, it substantially improves worst case OoD performance over multiple, randomly initialized models. Though we do not suggest that NC is the sole mechanism or a comprehensive explanation for OoD behaviour in deep neural networks (DNN), we believe NC's simple mathematical and geometric structure can provide a framework for analysis of this complex phenomenon in future work.

Jarrod Haas, William Yolland, Bernhard Rabus

We demonstrate that L2 normalization over feature space can produce capable performance for Out-of-Distribution (OoD) detection for some models and datasets. Although it does not demonstrate outright state-of-the-art performance, this method is notable for its extreme simplicity: it requires only two addition lines of code, and does not need specialized loss functions, image augmentations, outlier exposure or extra parameter tuning. We also observe that training may be more efficient for some datasets and architectures. Notably, only 60 epochs with ResNet18 on CIFAR10 (or 100 epochs with ResNet50) can produce performance within two percentage points (AUROC) of several state-of-the-art methods for some near and far OoD datasets. We provide theoretical and empirical support for this method, and demonstrate viability across five architectures and three In-Distribution (ID) datasets.

Mohammadreza Mohseni, Jordan Yap, William Yolland et al.

One commonly used clinical approach towards detecting melanomas recognises the existence of Ugly Duckling nevi, or skin lesions which look different from the other lesions on the same patient. An automatic method of detecting and analysing these lesions would help to standardize studies, compared with manual screening methods. However, it is difficult to obtain expertly-labelled images for ugly duckling lesions. We therefore propose to use self-supervised machine learning to automatically detect outlier lesions. We first automatically detect and extract all the lesions from a wide-field skin image, and calculate an embedding for each detected lesion in a patient image, based on automatically identified features. These embeddings are then used to calculate the L2 distances as a way to measure dissimilarity. Using this deep learning method, Ugly Ducklings are identified as outliers which should deserve more attention from the examining physician. We evaluate through comparison with dermatologists, and achieve a sensitivity rate of 72.1% and diagnostic accuracy of 94.2% on the held-out test set.

Mohammadreza Mohseni, Jordan Yap, William Yolland et al.

Medical image diagnosis can be achieved by deep neural networks, provided there is enough varied training data for each disease class. However, a hitherto unknown disease class not encountered during training will inevitably be misclassified, even if predicted with low probability. This problem is especially important for medical image diagnosis, when an image of a hitherto unknown disease is presented for diagnosis, especially when the images come from the same image domain, such as dermoscopic skin images. Current out-of-distribution detection algorithms act unfairly when the in-distribution classes are imbalanced, by favouring the most numerous disease in the training sets. This could lead to false diagnoses for rare cases which are often medically important. We developed a novel yet simple method to train neural networks, which enables them to classify in-distribution dermoscopic skin disease images and also detect novel diseases from dermoscopic images at test time. We show that our BinaryHeads model not only does not hurt classification balanced accuracy when the data is imbalanced, but also consistently improves the balanced accuracy. We also introduce an important method to investigate the effectiveness of out-of-distribution detection methods based on presence of varying amounts of out-of-distribution data, which may arise in real-world settings.