Although the human visual system can recognize many concepts under challenging conditions, it still has some biases. In this paper, we investigate whether we can extract these biases and transfer them into a machine recognition system. We introduce a novel method that, inspired by well-known tools in human psychophysics, estimates the biases that the human visual system might use for recognition, but in computer vision feature spaces. Our experiments are surprising, and suggest that classifiers from the human visual system can be transferred into a machine with some success. Since these classifiers seem to capture favorable biases in the human visual system, we further present an SVM formulation that constrains the orientation of the SVM hyperplane to agree with the bias from human visual system. Our results suggest that transferring this human bias into machines may help object recognition systems generalize across datasets and perform better when very little training data is available.

Summary: This paper draws inspiration from work on psychophysics on classification images. Large-scale human experiments were run, where people were asked to classify images generated from random noise (randomly generated by inverting HOG or CNN feature spaces to more closely approximate the distribution of natural images). The results were used to 1) visualize human perception of different classes, 2) see how well classifiers trained on datasets of random noise would work on real images, and 3) use the results as an additional source of information to regularize classifiers trained on a small number of images. Quality: This is a very unusual paper. It is overall a high quality and well written paper where interesting and novel experiments were carried out; however it is unclear if the results or methods of the paper are of practical value.

Although the human visual system can recognize many concepts under challenging conditions, it still has some biases. In this paper, we investigate whether we can extract these biases and transfer them into a machine recognition system. We introduce a novel method that, inspired by well-known tools in human psychophysics, estimates the biases that the human visual system might use for recognition, but in computer vision feature spaces. Our experiments are surprising, and suggest that classifiers from the human visual system can be transferred into a machine with some success. Since these classifiers seem to capture favorable biases in the human visual system, we further present an SVM formulation that constrains the orientation of the SVM hyperplane to agree with the bias from human visual system. Our results suggest that transferring this human bias into machines may help object recognition systems generalize across datasets and perform better when very little training data is available.

Although the human visual system can recognize many concepts under challenging conditions,it still has some biases. In this paper, we investigate whether we can extract these biases and transfer them into a machine recognition system. Weintroduce a novel method that, inspired by well-known tools in human psychophysics, estimates the biases that the human visual system might use for recognition, but in computer vision feature spaces. Our experiments are surprising, andsuggest that classifiers from the human visual system can be transferred into a machine with some success. Since these classifiers seem to capture favorable biasesin the human visual system, we further present an SVM formulation that constrains the orientation of the SVM hyperplane to agree with the bias from human visual system. Our results suggest that transferring this human bias into machines may help object recognition systems generalize across datasets and perform betterwhen very little training data is available.