Training a single model on multiple input domains and/or output tasks allows for compressing information from multiple sources into a unified backbone hence improves model efficiency. It also enables potential positive knowledge transfer across tasks/domains, leading to improved accuracy and data-efficient training. However, optimizing such networks is a challenge, in particular due to discrepancies between the different tasks or domains: Despite several hypotheses and solutions proposed over the years, recent work has shown that uniform scalarization training, i.e., simply minimizing the average of the task losses, yields on-par performance with more costly SotA optimization methods. This raises the issue of how well we understand the training dynamics of multi-task and multi-domain networks. In this work, we first devise a large-scale unified analysis of multi-domain and multi-task learning to better understand the dynamics of scalarization across varied task/domain combinations and model sizes. Following these insights, we then propose to leverage population-based training to efficiently search for the optimal scalarization weights when dealing with a large number of tasks or domains.

In this paper, we propose a conditional early exiting framework for efficient video recognition. While existing works focus on selecting a subset of salient frames to re-duce the computation costs, we propose to use a simple sampling strategy combined with conditional early exiting to enable efficient recognition. Our model automatically learns to process fewer frames for simpler videos and more frames for complex ones. To achieve this, we employ a cascade of gating modules to automatically determine the earliest point in processing where an inference is sufficiently reliable. We generate on-the-fly supervision signals to the gates to provide a dynamic trade-off between accuracy and computational cost. Our proposed model outperforms competing methods on three large-scale video benchmarks. In particular, on ActivityNet1.3 and mini-kinetics, we outperform the state-of-the-art efficient video recognition methods with 1.3× and 2.1 less GFLOPs, respectively. Addition-ally, our method sets× a new state of the art for efficient video understanding on the HVU benchmark.

We propose Skip-Convolutions to leverage the large amount of redundancies in video streams and save computations. Each video is represented as a series of changes across frames and network activations, denoted as residuals. We reformulate standard convolution to be efficiently computed on residual frames: each layer is coupled with a binary gate deciding whether a residual is important to the model prediction, e.g. foreground regions, or it can be safely skipped, e.g. background regions. These gates can either be implemented as an efficient network trained jointly with convolution kernels, or can simply skip the residuals based on their magnitude. Gating functions can also incorporate block-wise sparsity structures, as required for efficient implementation on hardware platforms. By replacing all convolutions with Skip-Convolutions in two state-of-the-art architectures, namely EfficientDet and HRNet, we reduce their computational cost consistently by a factor of 3 ∼ 4× for two different tasks, without any accuracy drop. Extensive comparisons with existing model compression, as well as image and video efficiency methods demonstrate that Skip-Convolutions set a new state-of-the-art by effectively exploiting the temporal redundancies in videos.

Convolutional Neural Networks experience catastrophic forgetting when optimized on a sequence of learning problems: as they meet the objective of the current training examples, their performance on previous tasks drops drastically. In this work, we introduce a novel framework to tackle this problem with conditional computation. We equip each convolutional layer with task-specific gating modules, selecting which filters to apply on the given input. This way, we achieve two appealing properties. Firstly, the execution patterns of the gates allow to identify and protect important filters, ensuring no loss in the performance of the model for previously learned tasks. Secondly, by using a sparsity objective, we can promote the selection of a limited set of kernels, allowing to retain sufficient model capacity to digest new tasks. Existing solutions require, at test time, awareness of the task to which each example belongs to. This knowledge, however, may not be available in many practical scenarios. Therefore, we additionally introduce a task classifier that predicts the task label of each example, to deal with settings in which a task oracle is not available. We validate our proposal on four continual learning datasets. Results show that our model consistently outperforms existing methods both in the presence and the absence of a task oracle. Notably, on Split SVHN and Imagenet-50 datasets, our model yields up to 23.98% and 17.42% improvement in accuracy w.r.t. competing methods.

Automated detection of cancer metastases in lymph nodes has the potential to improve the assessment of prognosis for patients. To enable fair comparison between the algorithms for this purpose, we set up the CAMELYON17 challenge in conjunction with the IEEE International Symposium on Biomedical Imaging 2017 Conference in Melbourne. Over 300 participants registered on the challenge website, of which 23 teams submitted a total of 37 algorithms before the initial deadline. Participants were provided with 899 whole-slide images (WSIs) for developing their algorithms. The developed algorithms were evaluated based on the test set encompassing 100 patients and 500 WSIs. The evaluation metric used was a quadratic weighted Cohen's kappa. We discuss the algorithmic details of the 10 best pre-conference and two post-conference submissions. All these participants used convolutional neural networks in combination with pre- and postprocessing steps. Algorithms differed mostly in neural network architecture, training strategy, and pre- and postprocessing methodology. Overall, the kappa metric ranged from 0.89 to -0.13 across all submissions. The best results were obtained with pre-trained architectures such as ResNet. Confusion matrix analysis revealed that all participants struggled with reliably identifying isolated tumor cells, the smallest type of metastasis, with detection rates below 40%. Qualitative inspection of the results of the top participants showed categories of false positives, such as nerves or contamination, which could be targeted for further optimization. Last, we show that simple combinations of the top algorithms result in higher kappa metric values than any algorithm individually, with 0.93 for the best combination.

This paper presents and evaluates a fully automatic method for detection of ductal carcinoma in situ (DCIS) in digitized hematoxylin and eosin (H&E) stained histopathological slides of breast tissue. The proposed method applies multi-scale superpixel classification to detect epithelial regions in whole-slide images (WSIs). Subsequently, spatial clustering is utilized to delineate regions representing meaningful structures within the tissue such as ducts and lobules. A region-based classifier employing a large set of features including statistical and structural texture features and architectural features is then trained to discriminate between DCIS and benign/normal structures. The system is evaluated on two datasets containing a total of 205 WSIs of breast tissue. Evaluation was conducted both on the slide and the lesion level using FROC analysis. The results show that to detect at least one true positive in every DCIS containing slide, the system finds 2.6 false positives per WSI. The results of the per-lesion evaluation show that it is possible to detect 80% and 83% of the DCIS lesions in an abnormal slide, at an average of 2.0 and 3.0 false positives per WSI, respectively. Collectively, the result of the experiments demonstrate the efficacy and accuracy of the proposed method as well as its potential for application in routine pathological diagnostics. To the best of our knowledge, this is the first DCIS detection algorithm working fully automatically on WSIs.

Variations in the color and intensity of hematoxylin and eosin (H&E) stained histological slides can potentially hamper the effectiveness of quantitative image analysis. This paper presents a fully automated algorithm for standardization of whole-slide histopathological images to reduce the effect of these variations. The proposed algorithm, called whole-slide image color standardizer (WSICS), utilizes color and spatial information to classify the image pixels into different stain components. The chromatic and density distributions for each of the stain components in the hue-saturation-density color model are aligned to match the corresponding distributions from a template whole-slide image (WSI). The performance of the WSICS algorithm was evaluated on two datasets. The first originated from 125 H&E stained WSIs of lymph nodes, sampled from 3 patients, and stained in 5 different laboratories on different days of the week. The second comprised 30 H&E stained WSIs of rat liver sections. The result of qualitative and quantitative evaluations using the first dataset demonstrate that the WSICS algorithm outperforms competing methods in terms of achieving color constancy. The WSICS algorithm consistently yields the smallest standard deviation and coefficient of variation of the normalized median intensity measure. Using the second dataset, we evaluated the impact of our algorithm on the performance of an already published necrosis quantification system. The performance of this system was significantly improved by utilizing the WSICS algorithm. The results of the empirical evaluations collectively demonstrate the potential contribution of the proposed standardization algorithm to improved diagnostic accuracy and consistency in computer-aided diagnosis for histopathology data.