By the late 1990s, increasing computer power and sophisticated algorithms enabled machines to automatically learn patterns from data, making object recognition faster, more flexible, and more accurate.
A pivotal moment occurred in 2009. That’s when Fei-Fei Li and her team, supported by an NSF Faculty Early Career Development Award, launched ImageNet, a public database containing more than 3 million images across 5,000 categories. ImageNet provided researchers with the large, high-quality datasets needed to train so-called deep learning systems that can recognize complex real-world images using artificial multilayer neural networks.
ImageNet also inspired the ImageNet Challenge, an annual competition to see which algorithm can most accurately identify a new image. In 2012, a deep learning model called AlexNet used CNNs and advanced graphics processors to analyze ImageNet’s massive dataset. We achieved record-breaking accuracy on the ImageNet challenge, cutting the error rate in half and proving that deep learning can significantly outperform previous approaches.
Along with ImageNet, NSF-supported centers such as the Temporal Dynamics of Learning Center have continued to expand our understanding of how the brain perceives objects and provided important insights that influence the development of today’s deep learning algorithms. The NSF AI Institute for Foundations of Machine Learning builds on this legacy by advancing the core mathematical and computational tools behind AI. For example, researchers are using deep generative models to reduce noise and sharpen low-quality or blurry images, such as MRI scans, to produce clearer, more accurate results for medical diagnosis and treatment.
