The purpose of Artificio is to: [1] design the next generation of machine vision systems that can 'see' like a human by reverse-engineering visual computations of biological organisms at the representational level via insights from computational models, psychophysics, and neural recordings (short-term); and [2] in doing so develop new technologies that may enhance visual capabilities of humans & treat visual diseases such as visual agnosia & prosopagnosia (long-term).

To accomplish this, we propose 3 main Research directions as the pillars for the study at Artificio that span the fields of vision science, computer vision, machine learning and robotics. The technologies discovered from our research will lead to new Developments in improving vision for humans & machines, that will results in robust reverse-image search, new autonomous driving API technologies, new state-of-the-art radiology screening detection systems, and treatments and diagnosis of visual processing disorders in humans.

To go beyond the role of architectural constraints in object recognition, Artificio will focus on understanding what are the computations in the ventral stream that may give rise to adversarial robustness by training and testing networks on specific types of stimuli. We believe that the most relevant unsolved problem in computer vision (and perhaps modern vision science) is the human perceptually misaligned noise-like adversarial images in machines which are a product of PGD attacks. Note that humans are not adversarially-robust, as they do have their own set of perceptual singularities that shape the family of visual illusions that trick our own visual system such as bi-stable images. Thus, the goal of modern machine vision should not be towards developing a machine that scores high accuracy on a benchmark or even that is adversarially robust, but rather to develop a machine that can make the same type of correct predictions and mistakes as a human across a wide range of stimuli including in-distribution, out-of-distribution, and adversarial images.

Recent advances in Reinforcement Learning (RL) have shown how multiple agents learn to walk, act, or assemble, given a reward function under constrained environments. In general, these ecologically-driven approaches have been applied purely to the action space while the visual sensor and perceptual module has been fixed as a spatially uniform processing visual sensor like a camera -- independent of the task. At Artificio we would like to design a set of new environments using Unity ML labs, where agents can evolve the wiring constraints of their visual sensors ; for example develop a foveated visual sensor in addition to potentially developing more than one type of visual array as can be seen in the animal kingdom (chameleons, humans, flies & arachnids). Artificio's 3rd stream of projects involve training RL agents to explore environments with spatially-adaptive and active vision with several animal retinal ganglion cell (RGC) convergence maps. Linking back to our first objective, we could also examine the learned representations as we vary the complexity of their virtual environments.