What Can't AI Do?
They paint strange nudes, colorize drawings, crush Minesweeper, and even run for political office. Is there anything artificial intelligences can't do? Their list of restrictions is getting smaller, and now we can add the reconstruction of corrupted and incomplete photographs. The project is led by a group of researchers under Nvidia, who trained a neural network with three special datasets on a platform based on the impressive Tesla V100 cards.
The Power Behind It: Tesla V100
One of the names that has most stimulated the development of AI is Nvidia, and the underlying reason is its hardware. The official page of the Tesla V100 card indicates that when used in groups of eight, it can train a neural network in a third of the time that eight cards of the previous generation, Tesla P100, would take. If we remember, the raytracing demo published by Epic Games used four Tesla V100, so imagine what that firepower can do for an AI. In fact... no, don't imagine. Better watch this video:
Reconstructing, Repairing, Replacing, and Correcting
A team of researchers at Nvidia trained a neural network on the PyTorch framework using the ImageNet, Places2, and CelebA-HQ datasets, plus a single 16GB Tesla V100 card, with the goal of reconstructing, repairing, replacing, and correcting different aspects in an image. Basically, the user creates 'holes' in the photograph, and the AI fills them with a convincing alternative. The demo shows the system has some strength in removing lines and other artificial objects, but has some difficulties processing human faces. Even so, the first touch-ups on Ernest Borgnine's face are notable.
According to the researchers, this is the first time that the usefulness and effectiveness of deep learning in models dedicated to processing irregularly shaped 'holes' (regardless of size, location, or distance from edges) has been demonstrated, while previous projects have focused on rectangular regions near the center of the image, which require expensive post-processing.
