Generate Accurate Climate Change Images Using AI

2019-10-24

Generative AI models can synthesize various objects such as faces, apartments, and butterflies. Recently, a professor has developed a new subclass of AI generative models with the aim of raising awareness about climate change by illustrating the consequences of catastrophic floods. In order to quantify the accuracy of these synthesized climate change images, researchers from the University of Montreal and Stanford University have recently detailed several evaluation methods in a preprint paper. They state in the paper that although their work is still in its early stages, they have begun to bridge the gap between automation and quantitative generation. This research was jointly conducted by Turing Award winner and professor Yoshua Bengio from the University of Montreal. Professor Yoshua Bengio was the first person to combine neural networks with probabilistic models of sequences. Nearly 20 years ago, in a paper he published, he introduced the concept of word embeddings, which is a paradigm for language modeling and feature learning. This paradigm maps words and phrases from the vocabulary to real-valued vectors, making significant contributions to machine translation, image generation, audio synthesis, and other fields. "Historically, the difficulty in mobilizing collective action against climate change lies in the fact that people find it hard to simulate the impact of climate change psychologically, often perceiving it as a distant problem," the paper states. "Climate communication literature has argued that effective communication arises from emotionally rich information that is personally relevant, with visualized images being the most effective in capturing people's attention." The researchers point out that existing evaluation methods for generated climate change images have significant limitations as they have low correlation with human judgment. Therefore, they propose an alternative approach involving the participation of volunteers. The volunteers' task is to evaluate image style combinations generated by the model based on input images of different locations and building types (houses, farms, streets, cities). Each image style has dozens of AI-generated patterns. Evaluators choose between real and semi-generated images, and the researchers calculate the average error rate, reflecting the proportion of images judged as real by the evaluators. A higher value indicates a more realistic image. The team evaluated a total of eight different automated methods and reported that the best-performing embedding model came from the intermediate layer of an AI model using the Frechet Inception Distance (FID). FID is a measurement method that obtains photos from the target distribution and the evaluated model, and uses an object recognition system to identify similarities between important features. The paper concludes, "The ultimate goal of this work is to create an ML architecture that can generate the most realistic images of extreme weather phenomena caused by climate change. We will continue to enhance the realism of the images and expand support for generated climate change categories."

RECENT AI NEWS

RECENT AI TOOLS