Invited talk, Uber AI Labs, San-Francisco, CA, US
A talk about recent results in application of generative networks for understanding of information encoded by high-level neurons in deep networks and macaque monkeys.
Conference talk, IEEE International Conference on Video and Audio Signal Processing in the Context of Neurotechnologies, St. Petersburg
In this talk I show how convolutional neural networks (CNNs) and Generative Adversarial Networks (GANs) can be used to study latent representations, coded by neurons in high visual areas. To explain neural preferences, I train a deep network to simulate the response of the neurons in the Inferior temporal (IT) cortex of macaque monkey. The model has high performance and explain > 0.65% of the variance in the neural data. I visualize latent representations of artificial neurons using a generative adversarial network. The approach allows to find an input signal that maximizes activation of an individual unit without limitations introduced by a dataset.
Conference talk, IEEE International Conference on Video and Audio Signal Processing in the Context of Neurotechnologies, St.Petersburg, Russia
I show how using an approach from signal detection theory can be applied to study high-level representations in a deep neural network to identify sensitive detectors. The results suggest that category-selective filters can be observed from the first layers of deep neural networks. I also show that a tuning curve of a category-selective filter differs from what is usually seen in a neural data. Representations of objects in weights of convolutional layers do not necessarily reflect the perceptual similarity of images. Instead, they are strongly dependent on a choice of a training set and cost function. This properties of neural networks may lead to the behavior when a model detect objects in their absence and fails to recognize obvious cases from a human-observer point of view.
Workshop talk, MODVIS Workshop, St. Pete Beach, Florida, US
It is common to compare properties of visual information processing by artificial neural networks and the primate visual system. Some remarkable similarities were observed in the responses of neurons in IT cortex and units in higher layers of CNNs. Here I show that latent representations formed by weights in convolutional layers do not necessarily reflect visual domain. Instead they are strongly dependent on a choice of training set and cost function. The most striking example is when an individual unit, which is highly selective to some members of a category is, nevertheless, inhibited by visually similar objects of the same category. And this surprising selectivity-profile cannot be attributed to incidental differences in low level statistics.
Download extended abstract here
Conference invited talk, Международная научная конференция молодых ученых «Психология XXI века: психология как наука, искусство и призвание», St.Petersburg, Russia