03/09/2019

At IMPA, IBM Research researcher details deep learning.

How can deep learning be used in the area of image processing and computer vision? In this Wednesday’s (4) seminar at the Computer Graphics Laboratory (Visgraf) at IMPA, Bianca Zadrozny, from IBM Research Brazil, will detail the technique, which has been used in applications in various areas of knowledge as a result of the expansion of the computational power of machines and the flood of data that we produce daily.

In “ Deep learning: why all the hype?”, a lecture open to the public at 1:30 PM in auditorium 3, Bianca will explain the technique by which a computer is able to learn from unstructured data (images and texts).

At IBM Research Brazil, Bianca develops advanced techniques to analyze critical problems in the natural resources sector. Combining state-of-the-art database management, statistics, and machine learning, she creates predictive models that help companies make informed decisions.

Luiz Henrique de Figueiredo, a senior researcher at IMPA in the field of computer graphics, observes that the existence of a huge volume of easily accessible images already allows, in certain cases, complicated algorithms to be replaced by deep neural networks.

In 2018, studies on deep neural networks earned researchers Yoshua Bengio, Geoffrey Hinton, and Yann LeCun the ACM AM Turing Award, a distinction considered the Nobel Prize of Computing. These networks mimic the processing done by the human brain, understanding human speech and visually recognizing objects. As a result, they are being used to solve a variety of problems in various fields of knowledge.

“Examples of problems with good solutions of this type include the recognition and classification of objects in images, noise removal and repair of damaged images, color transfer, and colorization of black and white images,” lists Figueiredo.

In addition to its use in image processing and computer vision, deep learning is applied to traditional computer graphics tasks such as realistic image synthesis, reconstruction of 3D models from points, physics-based animation, and geometric modeling.

In her presentation this Wednesday, Bianca will also discuss generative adversarial networks (GANs), a deep learning architecture that allows for data synthesis when trained on a real dataset. She will present an application developed by the IBM Research team that uses GANs to generate realistic seismic data from geoscientists' sketches.