Deep learning models of perception and cognition by Marco Zorzi (University of Padova, Italy)
Deep learning in stochastic recurrent neural networks with many layers of neurons (“deep networks”) is a recent breakthrough in neural computation research. These networks build a hierarchy of progressively more complex representations of the sensory data through unsupervised learning. Using examples from research in my laboratory, I will show that deep learning models represent a major step forward for connectionist modeling in psychology and cognitive neuroscience. I will also focus on a new model of letter perception, which shows that learning written symbols can recycle the visual primitives of natural images, thereby requiring only limited domain-specific tuning.