Biological and Computer Vision

Gabriel Kreiman

Cambridge University Press. 2021. ISBN 9781108649995 

Additional Materials

Chapter IX: Towards a world with intelligent machines that can interpret the visual world

Fueled by the availability of large training datasets, increased computational power, and biologically inspired architectures, there has been rapid progress in the development of computer vision systems that find applicability in a plethora of real-world pattern recognition problems. Computer vision has thrived, and often surpasses humans, in tasks like face recognition, clinical diagnosis based on imaging techniques, classifying the shapes of galaxies, flora and fauna, and scene interpretation for self-driving cars. We may soon have computer vision systems that can help blind people interpret the world around them. An intriguing corollary of modern computer vision systems is the possibility of creating quasi-realistic images by inverting the architectures, as in generative adversarial networks. Despite the multiple successes of computer vision systems during the last decade, current algorithms remain fragile, can be easily fooled via adversarial images, and often fail to generalize. The quod-erat-demonstrandum (qed) will be a model that can pass the Turing for vision, that is, a model that can answer any arbitrary question about any image, and those answers are indistinguishable from the ones provided by humans.