The problem with artificial intelligence is that everybody thinks it should do everything perfectly right out of the box, according to Rama Chellappa, an expert in computer vision and machine learning. There are always tradeoffs," said Chellappa, a Bloomberg Distinguished Professor in electrical and computer engineering and biomedical engineering, during a virtual discussion on Thursday--the latest session in the Johns Hopkins Congressional Briefing series. AI, Chellappa noted, is built on data, and data is not always perfect or complete. Errors in how data is collected, compiled, or processed will affect the success of any AI tool that relies on it. That is how challenges such as bias creep into autonomous systems. The process for reaching for fully functional AI, he said, yields new discoveries along the way. "[AI] is doing good things in a lot of ways, but it is not a perfect technology," he said. "For example, take the autonomous car.

A prized attribute among law enforcement specialists, the expert ability to visually identify human faces can inform forensic investigations and help maintain safe border crossings, airports, and public spaces around the world. The field of forensic facial recognition depends on highly refined traits such as visual acuity, cognitive discrimination, memory recall, and elimination of bias. Humans, as well as computers running machine learning (ML) algorithms, possess these abilities. And it is the combination of the two--a human facial recognition expert teamed with a computer running ML analyses of facial image data--that provides the most accurate facial identification, according to a recent 2018 study in which Rama Chellappa, Distinguished University Professor and Minta Martin Professor of Engineering, and his team collaborated with researchers at the National Institute of Standards and Technology and the University of Texas at Dallas. Chellappa, who holds appointments in UMD's Departments of Electrical and Computer Engineering and Computer Science and Institute for Advanced Computer Studies, is not surprised by the study results.

Attributes are human describable features, which have been used successfully for face, object, and activity recognition. Facial attributes are intuitive descriptions of faces and have proven to be very useful in face recognition and verification. Despite their usefulness, to date there is only one large-scale facial attribute dataset, CelebA. Impressive results have been achieved on this dataset, but it exhibits a variety of very significant biases. As CelebA contains mostly frontal idealized images of celebrities, it is difficult to generalize a model trained on this data for use on another dataset (of non celebrities). A typical approach to dealing with imbalanced data involves sampling the data in order to balance the positive and negative labels, however, with a multi-label problem this becomes a non-trivial task. By sampling to balance one label, we affect the distribution of other labels in the data. To address this problem, we introduce a novel Selective Learning method for deep networks which adaptively balances the data in each batch according to the desired distribution for each label. The bias in CelebA can be corrected for in this way, allowing the network to learn a more robust attribute model. We argue that without this multi-label balancing, the network cannot learn to accurately predict attributes that are poorly represented in CelebA. We demonstrate the effectiveness of our method on the problem of facial attribute prediction on CelebA, LFWA, and the new University of Maryland Attribute Evaluation Dataset (UMD-AED), outperforming the state-of-the-art on each dataset.

In November 2005, three suicide bombers walked into three hotels in Jordan and blew themselves up, killing 63 and injuring more than 100. While the world is alert to such deadly threats, the challenge remains: how to detect approaching suicide bombers from a safe distance. X-ray machines can obviously see a concealed bomb, but they are dangerous to humans–and a bomber could detonate himself and kill people at the checkpoint. Video surveillance can help, but it requires personnel trained to scan crowds and pick out suspicious individuals. A new radar-imaging technology expected to reach market later this year could solve the problem by directing low-power radar beams at people–who can be 50 yards or more away–and analyzing reflected radar returns to reveal concealed objects.