See how an improvement to the L-BFGS algorithm takes advantage of progressive batching, line search, and quasi-Newton updating.
This paper introduces a new neural network architecture and shows how improvements to a Cycle-GAN significantly reduces training time for translating images across domains.
Learn more about an award-wining method that uses noise unnoticeable to humans to confuse neural networks into making enormous mistakes.
A new test shows that some of the best state-of-the-art natural language models fail by changing even just a few words.
This research paper introduces DSOD, a framework that outperforms modern variations of regional proposal and classification networks.
This research shows how training low-bit deep neural networks has a much smaller memory footprint than full-precision models but with a small cost to accuracy.
Read about a mechanism to improve the memory capacity and time-scale of recurrent neural networks that allows for higher accuracy and faster training.
This robotic vision innovation uses Intel® RealSense™ technology and a Faster R-CNN to track and harvest sweet peppers as a test for helping fruit farmers.
A Fully Convolutional Model for Variable Bit Length and Lossy High-Density Compression of Mammograms
Using a convolutional autoencoder, a compression algorithm rivals JPEG and JPEG 2000 to reduce the size of mammogram images.
This paper presents a novel technique for detecting malware through classifying hardware-generated control flow traces using a deep neural network.