Learning Multi-Sensory Integration with Self-Organization and Statistics

Recently, we have presented a self-organized artificial neural network algorithm capable of learning a latent variable model of its high-dimensional input and to optimally integrate that input to compute and population-code a probability density function over the values of the latent variables of that model. We did take our motivation from natural neural networks and reported on a simple experiment with simulated multi-sensory data. However, we focused on presenting the algorithm and evaluating its performance, leaving a comparison with natural cognition for future work. In this paper, we show that our algorithm behaves similar, in important behavioral and neural aspects, to a prime example of natural multi-sensory integration: audio-visual object localization.