We expect a plate and spoon to be on a dining table but not on a high-way. More importantly we often get to see the same scene with and without objects like when we see a car come in and park in a parking lot or when someone takes away the plate and spoon from a table. We not only seem to have a reliable expectation about whether an object should be present but also about its location and size. Human vision then has an opportunity to learn independent representations of the scene and objects therein. How systematic are these expectations across humans? Do state-of-art machine vision algorithms also learn similar representations as humans? In this study we found that they indeed are and also that even good deep convolutional networks do not learn these expectations. We then went on to show that augmenting decisions of deep convolutional networks with human like priors improves their performance and this gives us new guidelines and constraints to train these machine vision models.

It is well known that humans have the highest visual acuity (resolution) at the location they are currently looking at. We asked whether human foveal blur has a bigger role to play than just being an information bottleneck and is indeed optimal for object recognition. How would one test whether human foveal blur profiles are indeed optimal? For this we trained deep convolutional networks with inputs at varying levels of spatial blur using a very large (~1 million) natural image dataset. The key insights we obtained were that (a) blur profiles matched to human foveal blur gave an advantage over other kinds of blur and even seeing the full resolution intact scene. (b) Deep convolutional networks had faster learning rates and higher accuracy when the input was subjected to human foveal blur profiles. (c) This advantage came about because object identity is more discriminable in coarser spatial frequencies at the periphery as opposed to the center of the visual field.