An Inference Network Model for Goal-directed Attentional Selection

Cognitive Computational Neuroscience (2019)
doi: 10.32470/CCN.2019.1431-0
2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany


"Listen to the cello in this symphony!" How can we direct selective attention according to different goals, even in distracting environments which we haven't experienced before? It is an essential cognitive ability of the brain, but remains challenging for machines. We developed a computational model that can identify individual digits in images containing multiple overlapping digits, without ever having seen overlapping digits during training. The goal-driven attentional selection is modelled as inferring the posterior distribution of latent variables (the attended target) in a generative model, conditioned on both sensory input and different semantic goals. A neural network model has been build to efficiently carry out the the inference process by predicting the most likely results, instead of using classic per-sample based iterative optimization methods which may not naturally map onto neural structures. Our model also help to understand how top-down and bottom-up attention are combined during perception in the brain.