Achieving human-like visual abilities is a holy grail for machine vision, yet precisely how insights from human vision can improve machines has remained unclear. Here, we demonstrate two key conceptual advances: First, we show that most machine vision models are systematically different from human object perception. To do so, we collected a large dataset of perceptual distances between isolated objects in humans and asked whether these perceptual data can be predicted by many common machine vision algorithms. We found that while the best algorithms explain ∼ 70 percent of the variance in the perceptual data, all the algorithms we tested make systematic errors on several types of objects. In particular, machine algorithms underestimated distances between symmetric objects compared to human perception. Second, we show that fixing these systematic biases can lead to substantial gains in classification performance. In particular, augmenting a state-of-the-art convolutional neural network with planar/reflection symmetry scores along multiple axes produced significant improvements in classification accuracy (1-10 percent) across categories. These results show that machine vision can be improved by discovering and fixing systematic differences from human vision.

Neuroscientists have recently shown that images that are difficult to find in visual search elicit similar patterns of firing across a population of recorded neurons. The$L^{1}$distance between firing rate vectors associated with two images was strongly correlated with the inverse of decision time in behavior. But why should decision times be correlated with$L^{1}$distance? What is the decision-theoretic basis? In our decision theoretic formulation, we model visual search as an active sequential hypothesis testing problem with switching costs. Our analysis suggests an appropriate neuronal dissimilarity index, which correlates equally strongly with the inverse of decision time as the$L^{1}$distance. We also consider a number of other possibilities, such as the relative entropy (Kullback–Leibler divergence) and the Chernoff entropy of the firing rate distributions. A more stringent test of equality of means, which would have provided a strong backing for our modeling, fails for our proposed as well as the other already discussed dissimilarity indices. However, test statistics from the equality of means test, when used to rank the indices in terms of their ability to explain the observed results, places our proposed dissimilarity index at the top followed by relative entropy, Chernoff entropy, and the$L^{1}$indices. Computations of the different indices require an estimate of the relative entropy between two Poisson point processes. An estimator is developed and is shown to have near unbiased performance for almost all operating regions.

Recent advances in neural networks have revolutionized computer vision, but these algorithms are still outperformed by humans. Could this performance gap be due to systematic differences between object representations in humans and machines? To answer this question we collected a large dataset of 26,675 perceived dissimilarity measurements from 2,801 visual objects across 269 human subjects, and used this dataset to train and test leading computational models. The best model (a combination of all models) accounted for 68% of the explainable variance. Importantly, all computational models showed systematic deviations from perception: (1) They underestimated perceptual distances between objects with symmetry or large area differences, (2) They overestimated perceptual distances between objects with shared features. Our results reveal critical elements missing in computer vision algorithms and point to explicit encoding of these properties in higher visual areas in the brain.

We consider a visual search problem studied by Sripati and Olson where the objective is to identify an oddball image embedded among multiple distractor images as quickly as possible. We model this visual search task as an active sequential hypothesis testing problem (ASHT problem). Chernoff in 1959 proposed a policy in which the expected delay to decision is asymptotically optimal. The asymptotics is under vanishing error probabilities. We first prove a stronger property on the moments of the delay until a decision, under the same asymptotics. Applying the result to the visual search problem, we then propose a “neuronal metric” on the measured neuronal responses that captures the discriminability between images. From empirical study we obtain a remarkable correlation (r = 0.90) between the proposed neuronal metric and speed of discrimination between the images. Although this correlation is lower than with the L1metric used by Sripati and Olson, this metric has the advantage of being firmly grounded in formal decision theory.