Visual Intelligence publications are financially supported by the Research Council of Norway, through its Centre for Research-based Innovation funding scheme (grant no. 309439), and Consortium Partners.
Visual Intelligence publications are financially supported by the Research Council of Norway, through its Centre for Research-based Innovation funding scheme (grant no. 309439), and Consortium Partners.
At Visual Intelligence we work across our innovation areas to extract knowledge from large volumes of visual data more efficiently through automatic and intelligent data analysis. The work to address the core research challenges in deep learning: working with limited training data, utilizing context and dependencies, providing explainability, confidence and uncertainty, are important in all the innovation areas.
In this essay, we review some recent advances in developing machine learning (ML) methods for marine science applications in Norway. We focus mostly on deep learning (DL) methods and review the challenges we have faced in the process, including data preparation, (lack of) labelled training data, and interpretability.
The objective function of a class of Gaussian mixture models can be rephrased as the loss function of a one-hidden layer autoencoder: the clustering module. Integrating the latter into a deep autoencoder yields a model able to jointly learn a clustering and an embedding and with state-of-the-art clustering performance.
Capturing global contextual representations in remote sensing images by exploiting long-range pixel–pixel dependencies has been shown to improve segmentation performance. We propose the Self-Constructing Graph (SCG) module that learns a long-range dependency graph directly from the image data.
To appear in CVPR 2021.
We identify several drawbacks with current state of the art methods for deep multi-view clustering, and present a simple baseline model that avoids these drawbacks. The baseline is expanded with a contrastive learning component, resulting in a model which outperforms the current SOTA on several benchmark datasets.
Published to AAAI-21.
An interpretable and differentiable dependence (or independence) measure that can be used to 1) train deep network under covariate shift and non-Gaussian noise; 2) implement a deep deterministic information bottleneck; and 3) understand the dynamics of learning of CNN. Code available.