Learning informative representations of data is one of the primary goals of deep learning, but there is still little understanding as to what representations a neural network actually learns. To better understand this, subspace match was recently proposed as a method for assessing the similarity of the representations learned by neural networks. It has been shown that two networks with the same architecture trained from different initializations learn representations that at hidden layers show low similarity when assessed with subspace match, even when the output layers show high similarity and the networks largely exhibit similar performance on classification tasks. In this note, we present a simple example motivated by standard results in commutative algebra to illustrate how this can happen, and show that although the subspace match at a hidden layer may be 0, the representations learned may be isomorphic as vector spaces. This leads us to conclude that a subspace match comparison of learned representations may well be uninformative, and it points to the need for better methods of understanding learned representations.
Self--supervised machine learning algorithms now achieve accuracies comparable to their supervised counterparts on benchmark classification tasks. The representations learned by these algorithms generalize to a range of downstream tasks beyond classification. We apply self--supervised learning to better understand auroral morphology. Specifically, we demonstrate using a recently released auroral image dataset that a modified version of the Simple framework for Contrastive Learning of Representations (SimCLR) algorithm (Chen et. al. 2020) learns sufficiently high-quality representations for a simple linear classifier to classify them with an average precision of over 94%, an 11 percentage-point improvement over the previous benchmark. The representations learned by our model cluster into more clusters than exist manually assigned categories, suggesting that the existing categorization is overly coarse and may obscure important connections between auroral types, near-earth solar wind conditions, and geomagnetic disturbances at the earth's surface. By combining our self-supervised representation learning strategy with semi-supervised automatic labelling, we aim to classify all of the auroral image data available from Time History of Events and Macroscale Interactions during Substorms (THEMIS) all-sky imagers, producing a labelled, machine-learning ready auroral image database many orders of magnitude greater than any currently available.
Meso-scale structures, that have spatial scales between 100 to 500 km and temporal scales between 2 to 15 minutes, occur frequently during heightened levels of geomagnetic activity. These structures are responsible for significant energy dissipation from the Magnetosphere to the Ionosphere and can result in damage to critical technology. Our capabilities to predict the occurrence of meso-scale structures are lacking as the operational models traditionally use empirical relations to define the magnetospheric drivers. Optimized to provide a timely global solution, the operational models also employ coarser grid resolutions that are insufficient to resolve meso-scales. Ever-increasing computational resources enable more realistic numerical simulations of the geospace environment, while various missions and experiments have been generating data that motivate new assimilation and machine learning techniques. This presentation will introduce the predictive models for ionospheric convection and aurora classification, as well as techniques to incorporate incoherent scatter radar measurements as drivers for general circulation models with the aim of providing further insight into the meso-scale coupling of the Magnetosphere-Ionosphere systems.
Forecasting ground magnetic field perturbations has been a long-standing goal of the space weather community. The availability of ground magnetic field data and its potential to be used in geomagnetically induced current studies, such as risk assessment, have resulted in several forecasting efforts over the past few decades. One particular community effort was the Geospace Environment Modeling (GEM) challenge of ground magnetic field perturbations that evaluated the predictive capacity of several empirical and first principles models at both mid- and high-latitudes in order to choose an operative model. In this work, we use three different deep learning models-a feed-forward neural network, a long short-term memory recurrent network and a convolutional neural network-to forecast the horizontal component of the ground magnetic field rate of change (dB<sub>H</sub>/dt) over 6 different ground magnetometer stations and to compare as directly as possible with the original GEM challenge. We find that, in general, the models are able to perform at similar levels to those obtained in the original challenge, although the performance depends heavily on the particular storm being evaluated. We then discuss the limitations of such a comparison on the basis that the original challenge was not designed with machine learning algorithms in mind.
Unsupervised learning algorithms are beginning to achieve accuracies comparable to their supervised counterparts on benchmark computer vision tasks, but their utility for practical applications has not yet been demonstrated. In this work, we present a novel application of unsupervised learning to the task of auroral image classification. Specifically, we modify and adapt the Simple framework for Contrastive Learning of Representations (SimCLR) algorithm to learn representations of auroral images in a recently released auroral image dataset constructed using image data from Time History of Events and Macroscale Interactions during Substorms (THEMIS) all-sky imagers. We demonstrate that (a) simple linear classifiers fit to the learned representations of the images achieve state-of-the-art classification performance, improving the classification accuracy by almost 10 percentage points over the current benchmark; and (b) the learned representations naturally cluster into more clusters than exist manually assigned categories, suggesting that existing categorizations are overly coarse and may obscure important connections between auroral types, near-earth solar wind conditions, and geomagnetic disturbances at the earth's surface. Moreover, our model is much lighter than the previous benchmark on this dataset, requiring in the area of fewer than 25\% of the number of parameters. Our approach exceeds an established threshold for operational purposes, demonstrating readiness for deployment and utilization.
Engaging undergraduate students in large classes is a constant challenge for many lecturers, as student participation and engagement can be limited. This is a concern since there is a positive correlation between increased engagement and student success. The lack of student feedback on content delivery prevents lecturers from identifying topics that would benefit students if reviewed. Implementing novel methods to engage the students in course content and create ways by which they can inform the lecturer of the difficult concepts is needed to increase student success. In the present study, we investigated the use of Twitter as a scalable approach to enhance engagement with course content and peer-to-peer interaction in a large course. In this pilot study, students were instructed to tweet the difficult concepts identified from content delivered by videos. A software program automatically collected and parsed the tweets to extract summary statistics on the most common difficult concepts, and the lecturer used the information to prepare face-to-face (F2F) lectorial sessions. The key findings of the study were 1) the uptake of Twitter (i.e., registration on the platform) was similar to the proportion of students who participated in F2F lectorials, 2) students reviewed content soon after delivery to tweet difficult concepts to lecturer, 3) Twitter increased engagement with lecturers, 4) the difficult concepts were similar to previous years, yet the automated gathering of Twitter data was more efficient and time saving for the lecturer, and 5) students found the lectorial review sessions very valuable.
Invasive ductal carcinoma (IDC) comprises nearly 80% of all breast cancers. The detection of IDC is a necessary preprocessing step in determining the aggressiveness of the cancer, determining treatment protocols, and predicting patient outcomes, and is usually performed manually by an expert pathologist. Here, we describe a novel algorithm for automatically detecting IDC using semi-supervised conditional generative adversarial networks (cGANs). The framework is simple and effective at improving scores on a range of metrics over a baseline CNN.
The COVID-19 pandemic created a host of issues for institutions of higher education over the past year, including the issue of how to effectively assess student learning when courses are taught remotely. In this work-in-progress paper we present our experience using remote oral assessments in five introductory courses in two subject areas: computing and mathematics. We discuss our motivation for adopting this new assessment format and how to successfully implement remote oral assessments to replace traditional written final exams. We conducted a post-assessment student survey to understand how students responded to the oral exam format. The purpose of the survey was to gather feedback on students' emotions during and after the assessment. Our preliminary quantitative results show that overall students experienced more positive than negative emotions in all courses, though students responded differently in computing and mathematics courses. Students generally favored the oral format, and those who did not have previous experience with this type of assessment had a similar positive response to students who were more familiar with the format. We expect to shed more light on students' experience with the oral assessment as we will continue our research and conduct a qualitative study of the open-ended responses in the survey.
Generative adversarial networks (GANs) are a recently introduced class of state-of-the-art generative models. GANs are characterized by a unique training process that, although unstable, enables them to accurately learn highly complex distributions. While much of the recent attention that GANs have received in the machine learning and computer vision communities is due to their ability to synthesize highly realistic images, this is but one of many potential uses for these models. In this chapter, we survey several recent applications of GANs in medical imaging, highlighting significant developments, and illustrating avenues for future work in this nascent area of research.
Jupyter notebooks are widely used in industry and in academic research, but have only begun to make inroads into the classroom. The design of the Jupyter notebook is in many ways well suited for teaching subjects in information technology and computer science, but it is a tool that departs significantly from a standard text editor or integrated development environment, and thus carries with it several unique advantages as well as several surprising potential pitfalls. As use of Jupyter notebooks has grown, so has criticism of the notebook, for varied reasons: notebooks can behave in unexpected ways, they can be difficult to reproduce, they open up potential security issues, and they may encourage poor coding practices. A set of best practices to guide instructors and help addressing these concerns when using Jupyter notebooks in the classroom is currently lacking. This paper addresses the strengths and weaknesses of the Jupyter notebook for education, drawing on existing literature as well as the author's experience teaching a range of courses with Jupyter notebooks for over five years, and recommends a set of best practices for teaching with the Jupyter notebook.
Jupyter notebooks are widely used in industry for a range of tasks. This is particularly so in areas that involve significant amounts of data analysis or machine learning; indeed, while 5% of Python developers surveyed in the 2018 JetBrains Python Developer Survey report using Jupyter notebooks for their primary development tool, when restricted to those working in data science roles, Jupyter notebooks tied with the PyCharm IDE as the most popular tool for Python development , and in the 2019 StackOverflow developer survey, 9.5% of developers surveyed listed Jupyter notebooks as their preferred development environment .
Automatic segmentation of microscopy images is an important task in medical image processing and analysis. Nucleus detection is an important example of this task. Mask-RCNN is a recently proposed state-of-the-art algorithm for object detection, object localization, and object instance segmentation of natural images. In this paper we demonstrate that Mask-RCNN can be used to perform highly effective and efficient automatic segmentations of a wide range of microscopy images of cell nuclei, for a variety of cells acquired under a variety of conditions.
The artistic style of a painting can be sensed by the average observer, but algorithmically detecting a painting's style is a difficult problem. We propose a novel method for detecting the artistic style of a painting that is motivated by the neural-style algorithm of Gatys et. al. and is competitive with other recent algorithmic approaches to artistic style detection.