Matthew Toews

Javad Manashti, Pouyan Pirnia, Alireza Manashty et al.

This project aimed to determine the grain size distribution of granular materials from images using convolutional neural networks. The application of ConvNet and pretrained ConvNet models, including AlexNet, SqueezeNet, GoogLeNet, InceptionV3, DenseNet201, MobileNetV2, ResNet18, ResNet50, ResNet101, Xception, InceptionResNetV2, ShuffleNet, and NASNetMobile was studied. Synthetic images of granular materials created with the discrete element code YADE were used. All the models were trained and verified with grayscale and color band datasets with image sizes ranging from 32 to 160 pixels. The proposed ConvNet model predicts the percentages of mass retained on the finest sieve, coarsest sieve, and all sieves with root-mean-square errors of 1.8 %, 3.3 %, and 2.8 %, respectively, and a coefficient of determination of 0.99. For pretrained networks, root-mean-square errors of 2.4 % and 2.8 % were obtained for the finest sieve with feature extraction and transfer learning models, respectively.

Mehraveh Javan, Matthew Toews, Marco Pedersoli

Downsampling layers, including pooling and strided convolutions, are crucial components of the convolutional neural network architecture that determine both the granularity/scale of image feature analysis as well as the receptive field size of a given layer. To fully understand this problem, we analyse the performance of models independently trained with each pooling configurations on CIFAR10, using a ResNet20 network, and show that the position of the downsampling layers can highly influence the performance of a network and predefined downsampling configurations are not optimal. Network Architecture Search (NAS) might be used to optimize downsampling configurations as an hyperparameter. However, we find that common one-shot NAS based on a single SuperNet does not work for this problem. We argue that this is because a SuperNet trained for finding the optimal pooling configuration fully shares its parameters among all pooling configurations. This makes its training hard, because learning some configurations can harm the performance of others. Therefore, we propose a balanced mixture of SuperNets that automatically associates pooling configurations to different weight models and helps to reduce the weight-sharing and inter-influence of pooling configurations on the SuperNet parameters. We evaluate our proposed approach on CIFAR10, CIFAR100, as well as Food101 and show that in all cases, our model outperforms other approaches and improves over the default pooling configurations.

Laurent Chauvin, William Wells, Matthew Toews

This paper proposes to extend local image features in 3D to include invariance to discrete symmetry including inversion of spatial axes and image contrast. A binary feature sign $s \in \{-1,+1\}$ is defined as the sign of the Laplacian operator $\nabla^2$, and used to obtain a descriptor that is invariant to image sign inversion $s \rightarrow -s$ and 3D parity transforms $(x,y,z)\rightarrow(-x,-y,-z)$, i.e. SP-invariant or SP-symmetric. SP-symmetry applies to arbitrary scalar image fields $I: R^3 \rightarrow R^1$ mapping 3D coordinates $(x,y,z) \in R^3$ to scalar intensity $I(x,y,z) \in R^1$, generalizing the well-known charge conjugation and parity symmetry (CP-symmetry) applying to elementary charged particles. Feature orientation is modeled as a set of discrete states corresponding to potential axis reflections, independently of image contrast inversion. Two primary axis vectors are derived from image observations and potentially subject to reflection, and a third axis is an axial vector defined by the right-hand rule. Augmenting local feature properties with sign in addition to standard (location, scale, orientation) geometry leads to descriptors that are invariant to coordinate reflections and intensity contrast inversion. Feature properties are factored in to probabilistic point-based registration as symmetric kernels, based on a model of binary feature correspondence. Experiments using the well-known coherent point drift (CPD) algorithm demonstrate that SIFT-CPD kernels achieve the most accurate and rapid registration of the human brain and CT chest, including multiple MRI modalities of differing intensity contrast, and abnormal local variations such as tumors or occlusions. SIFT-CPD image registration is invariant to global scaling, rotation and translation and image intensity inversions of the input data.

Louis-François Bouchard, Mohsen Ben Lazreg, Matthew Toews

We report on a novel model linking deep convolutional neural networks (CNN) to biological vision and fundamental particle physics. Information propagation in a CNN is modeled via an analogy to an optical system, where information is concentrated near a bottleneck where the 2D spatial resolution collapses about a focal point $1\times 1=1$. A 3D space $(x,y,t)$ is defined by $(x,y)$ coordinates in the image plane and CNN layer $t$, where a principal ray $(0,0,t)$ runs in the direction of information propagation through both the optical axis and the image center pixel located at $(x,y)=(0,0)$, about which the sharpest possible spatial focus is limited to a circle of confusion in the image plane. Our novel insight is to model the principal optical ray $(0,0,t)$ as geometrically equivalent to the medial vector in the positive orthant $I(x,y) \in R^{N+}$ of a $N$-channel activation space, e.g. along the greyscale (or luminance) vector $(t,t,t)$ in $RGB$ colour space. Information is thus concentrated into an energy potential $E(x,y,t)=\|I(x,y,t)\|^2$, which, particularly for bottleneck layers $t$ of generic CNNs, is highly concentrated and symmetric about the spatial origin $(0,0,t)$ and exhibits the well-known "Sombrero" potential of the boson particle. This symmetry is broken in classification, where bottleneck layers of generic pre-trained CNN models exhibit a consistent class-specific bias towards an angle $θ\in U(1)$ defined simultaneously in the image plane and in activation feature space. Initial observations validate our hypothesis from generic pre-trained CNN activation maps and a bare-bones memory-based classification scheme, with no training or tuning. Training from scratch using combined one-hot $+ U(1)$ loss improves classification for all tasks tested including ImageNet.

Jean-Baptiste Carluer, Laurent Chauvin, Jie Luo et al.

This work details a highly efficient implementation of the 3D scale-invariant feature transform (SIFT) algorithm, for the purpose of machine learning from large sets of volumetric medical image data. The primary operations of the 3D SIFT code are implemented on a graphics processing unit (GPU), including convolution, sub-sampling, and 4D peak detection from scale-space pyramids. The performance improvements are quantified in keypoint detection and image-to-image matching experiments, using 3D MRI human brain volumes of different people. Computationally efficient 3D keypoint descriptors are proposed based on the Binary Robust Independent Elementary Feature (BRIEF) code, including a novel descriptor we call Ranked Robust Independent Elementary Features (RRIEF), and compared to the original 3D SIFT-Rank method\citep{toews2013efficient}. The GPU implementation affords a speedup of approximately 7X beyond an optimised CPU implementation, where computation time is reduced from 1.4 seconds to 0.2 seconds for 3D volumes of size (145, 174, 145) voxels with approximately 3000 keypoints. Notable speedups include the convolution operation (20X), 4D peak detection (3X), sub-sampling (3X), and difference-of-Gaussian pyramid construction (2X). Efficient descriptors offer a speedup of 2X and a memory savings of 6X compared to standard SIFT-Rank descriptors, at a cost of reduced numbers of keypoint correspondences, revealing a trade-off between computational efficiency and algorithmic performance. The speedups gained by our implementation will allow for a more efficient analysis on larger data sets. Our optimized GPU implementation of the 3D SIFT-Rank extractor is available at https://github.com/CarluerJB/3D_SIFT_CUDA.

Liam Frija-Altrac, Matthew Toews

This paper proposes elementary information mechanics as a new model for understanding the mechanical properties of convolutional filtering with rectification, inspired by physical theories of special relativity and quantum mechanics. We consider kernels decomposed into orthogonal even and odd components. Even components cause image content to diffuse isotropically while preserving the center of mass, analogously to rest or potential energy with zero net momentum. Odd kernels cause directional displacement of the center of mass, analogously to kinetic energy with non-zero momentum. The speed of information displacement is linearly related to the ratio of odd vs total kernel energy. Even-Odd properties are analyzed in the spectral domain via the discrete cosine transform (DCT), where the structure of small convolutional filters (e.g. $3 \times 3$ pixels) is dominated by low-frequency bases, specifically the DC $Σ$ and gradient components $\nabla$, which define the fundamental modes of information propagation. To our knowledge, this is the first work demonstrating the link between information processing in generic CNNs and the energy-momentum relation, a cornerstone of modern relativistic physics.

Louis-François Bouchard, Mohsen Ben Lazreg, Matthew Toews

This paper proposes a novel hue-like angular parameter to model the structure of deep convolutional neural network (CNN) activation space, referred to as the {\em activation hue}, for the purpose of regularizing models for more effective learning. The activation hue generalizes the notion of color hue angle in standard 3-channel RGB intensity space to $N$-channel activation space. A series of observations based on nearest neighbor indexing of activation vectors with pre-trained networks indicate that class-informative activations are concentrated about an angle $θ$ in both the $(x,y)$ image plane and in multi-channel activation space. A regularization term in the form of hue-like angular $θ$ labels is proposed to complement standard one-hot loss. Training from scratch using combined one-hot + activation hue loss improves classification performance modestly for a wide variety of classification tasks, including ImageNet.

Mehraveh Javan Roshtkhari, Matthew Toews, Marco Pedersoli

Monte-Carlo Tree Search (MCTS) is a powerful tool for many non-differentiable search related problems such as adversarial games. However, the performance of such approach highly depends on the order of the nodes that are considered at each branching of the tree. If the first branches cannot distinguish between promising and deceiving configurations for the final task, the efficiency of the search is exponentially reduced. In Neural Architecture Search (NAS), as only the final architecture matters, the visiting order of the branching can be optimized to improve learning. In this paper, we study the application of MCTS to NAS for image classification. We analyze several sampling methods and branching alternatives for MCTS and propose to learn the branching by hierarchical clustering of architectures based on their similarity. The similarity is measured by the pairwise distance of output vectors of architectures. Extensive experiments on two challenging benchmarks on CIFAR10 and ImageNet show that MCTS, if provided with a good branching hierarchy, can yield promising solutions more efficiently than other approaches for NAS problems.

Laurent Chauvin, Matthew Toews

Subject ID labels are unique, anonymized codes that can be used to group all images of a subject while maintaining anonymity. ID errors may be inadvertently introduced manually error during enrollment and may lead to systematic error into machine learning evaluation (e.g. due to double-dipping) or potential patient misdiagnosis in clinical contexts. Here we describe a highly efficient system for curating subject ID labels in large generic medical image datasets, based on the 3D image keypoint representation, which recently led to the discovery of previously unknown labeling errors in widely-used public brain MRI datasets

Laurent Chauvin, Kuldeep Kumar, Christian Desrosiers et al.

We propose a novel pairwise distance measure between image keypoint sets, for the purpose of large-scale medical image indexing. Our measure generalizes the Jaccard index to account for soft set equivalence (SSE) between keypoint elements, via an adaptive kernel framework modeling uncertainty in keypoint appearance and geometry. A new kernel is proposed to quantify the variability of keypoint geometry in location and scale. Our distance measure may be estimated between $O(N^2)$ image pairs in $O(N~\log~N)$ operations via keypoint indexing. Experiments report the first results for the task of predicting family relationships from medical images, using 1010 T1-weighted MRI brain volumes of 434 families including monozygotic and dizygotic twins, siblings and half-siblings sharing 100%-25% of their polymorphic genes. Soft set equivalence and the keypoint geometry kernel improve upon standard hard set equivalence (HSE) and appearance kernels alone in predicting family relationships. Monozygotic twin identification is near 100%, and three subjects with uncertain genotyping are automatically paired with their self-reported families, the first reported practical application of image-based family identification. Our distance measure can also be used to predict group categories, sex is predicted with an AUC=0.97. Software is provided for efficient fine-grained curation of large, generic image datasets.

Christian Wachinger, Matthew Toews, Georg Langs et al.

We introduce an approach for image segmentation based on sparse correspondences between keypoints in testing and training images. Keypoints represent automatically identified distinctive image locations, where each keypoint correspondence suggests a transformation between images. We use these correspondences to transfer label maps of entire organs from the training images to the test image. The keypoint transfer algorithm includes three steps: (i) keypoint matching, (ii) voting-based keypoint labeling, and (iii) keypoint-based probabilistic transfer of organ segmentations. We report segmentation results for abdominal organs in whole-body CT and MRI, as well as in contrast-enhanced CT and MRI. Our method offers a speed-up of about three orders of magnitude in comparison to common multi-atlas segmentation, while achieving an accuracy that compares favorably. Moreover, keypoint transfer does not require the registration to an atlas or a training phase. Finally, the method allows for the segmentation of scans with highly variable field-of-view.

Jie Luo, Alireza Sedghi, Karteek Popuri et al.

Estimating the uncertainty in (probabilistic) image registration enables, e.g., surgeons to assess the operative risk based on the trustworthiness of the registered image data. If surgeons receive inaccurately calculated registration uncertainty and misplace unwarranted confidence in the alignment solutions, severe consequences may result. For probabilistic image registration (PIR), the predominant way to quantify the registration uncertainty is using summary statistics of the distribution of transformation parameters. The majority of existing research focuses on trying out different summary statistics as well as a means to exploit them. Distinctively, in this paper, we study two rarely examined topics: (1) whether those summary statistics of the transformation distribution most informatively represent the registration uncertainty; (2) Does utilizing the registration uncertainty always be beneficial. We show that there are two types of uncertainties: the transformation uncertainty, Ut, and label uncertainty Ul. The conventional way of using Ut to quantify Ul is inappropriate and can be misleading. By a real data experiment, we also share a potentially critical finding that making use of the registration uncertainty may not always be an improvement.

Ahmad Chaddad, Behnaz Naisiri, Marco Pedersoli et al.

This paper proposes a principled information theoretic analysis of classification for deep neural network structures, e.g. convolutional neural networks (CNN). The output of convolutional filters is modeled as a random variable Y conditioned on the object class C and network filter bank F. The conditional entropy (CENT) H(Y |C,F) is shown in theory and experiments to be a highly compact and class-informative code, that can be computed from the filter outputs throughout an existing CNN and used to obtain higher classification results than the original CNN itself. Experiments demonstrate the effectiveness of CENT feature analysis in two separate CNN classification contexts. 1) In the classification of neurodegeneration due to Alzheimer's disease (AD) and natural aging from 3D magnetic resonance image (MRI) volumes, 3 CENT features result in an AUC=94.6% for whole-brain AD classification, the highest reported accuracy on the public OASIS dataset used and 12% higher than the softmax output of the original CNN trained for the task. 2) In the context of visual object classification from 2D photographs, transfer learning based on a small set of CENT features identified throughout an existing CNN leads to AUC values comparable to the 1000-feature softmax output of the original network when classifying previously unseen object categories. The general information theoretical analysis explains various recent CNN design successes, e.g. densely connected CNN architectures, and provides insights for future research directions in deep learning.