Computer Vision News - April 2022

45 Magda Paschali Improve Training Dynamics 3D Convolutional DNNs are widely used for volumetric medical imaging data such as MRI and CT Scans. However, compared to 2D DNNs, such models have more training parameters and are more prone to overfitting when trained with limited data. To that end, we proposed 3DQ, the first ternaryquantizationmethod for 3DDNNs. 3DQperformedweight quantization and utilized two trainable scaling factors and a normalization parameter to increase model capacity while maintaining compression. 3DQ managed to not only reduce the model size by 16 times but also enhanced the training dynamics and increased the Dice Score achieved by large volumetric models for hippocampus and whole-brain segmentation trained on limited scans. 3DQ constitutes a solid approach for space-critical applications, like patient- specific models or model weight transfer for Federated Learning. Model Evaluation with Adversarial Examples Model robustness evaluation is necessary for DNNs deployed on critical applications. Thus, we proposed a novel benchmarking strategy that utilized adversarial examples to evaluate state-of-the-art models for classification and segmentation. Our method highlighted that models that achieve similar or identical performance on clean test data had substantial differences regarding robustness to adversarial attacks. That could be attributed to notable differences in the models’ exploration of the underlying data manifold, resulting in varying robustness capabilities.