PyTorch 1.3 Released; News From Arm TechCon 2019; Are Deepfakes More Dangerous Than Imagined?

Subscribe to Synced Global AI Weekly

PyTorch 1.3 Adds Mobile, Privacy, Quantization, and Named Tensors
Facebook is now advancing the platform further with the release of PyTorch 1.3, which includes experimental support for features such as seamless model deployment to mobile devices, model quantization for better performance at inference time, and front-end improvements.
(Facebook AI) / (Synced)

All You Need to Know about Arm TechCon 2019
– Arm TechCon 2019 | Arm enables custom instructions for embedded CPUs
– Arm TechCon 2019 | Arm shows roadmap for collaborating on embedded operating systems
– Arm TechCon 2019 | Arm unveils consortium for autonomous vehicle computing

There Are Now 15,000 Deepfake Videos on Social Media. Yes, You Should Worry.
Analyzing videos from around the Web, a company called Deeptrace has determined there are almost 15,000 deepfakes in existence (14,678 to be exact), which is a significant increase from just 7,964 they found a year ago. That is a whopping 84% increase. And the problem is getting worse.
(Forbes)

Technology

Neural Turtle Graphics for Modeling City Road Layouts
Researchers propose Neural Turtle Graphics (NTG), a novel generative model for spatial graphs, and demonstrate its applications in modeling city road layouts. Specifically, They represent the road layout using a graph where nodes in the graph represent control points and edges in the graph represent road segments.
(University of Toronto & Vector Institute & MIT &NVIDIA)

On Learning Symmetric Locomotion
Human and animal gaits are often symmetric in nature. By encouraging symmetric motion, the learning may be faster, converge to more efficient solutions, and be more aesthetically pleasing. Researchers describe, compare, and evaluate four practical methods for encouraging motion symmetry.
(University of British Columbia & University of California, Berkeley)

DISCOMAN: Dataset of Indoor SCenes for Odometry, Mapping and Navigation
Researchers present a novel dataset for training and benchmarking semantic SLAM methods. The dataset consists of 200 long sequences, each one containing 3000-5000 data frames. They generate the sequences using realistic home layouts.
(Samsung AI Center)