Author: Maciej Pietrowski
Keywords: artificial neural networks, Python, convolutional neural networks, prednet
Nature has been inspiring us with its solutions for centuries. We observe it and try to imitate it. Artificial neural networks are one of the many examples of how much can be achieved with inspiration from biology. Nowadays, it is possible to model the functions of the human brain thanks to artificial neural networks. The human cerebral cortex predicts the movement of objects in the environment to adjust our behaviour. This allows us to react in a split second to changes taking place in our environment. The leading theory explaining this (rapid response) phenomenon is the predictive coding theory. It assumes that the internal models in the brain acquired through learning continuously generate predictions about the world around us. However, errors (differences between the predicted and actual images) are applied to update the model. The forecasts shorten the reaction time to an absolute minimum. PredNet is a neural network created to predict consecutive frames of a movie using the unsupervised learning method. According to Watanabe’s research, the neural network is capable of predicting movement on previously unseen stimuli and is susceptible to static motion illusions–still images that induce a sense of movement. The stimuli that mislead human perception seem to act similarly on the artificial neural network. The results obtained in this work confirm the network’s ability to predict movement on previously unseen material. PredNet neural network is prone to illusions, although, the majority of models generated by the network have produced erroneous predictions about the motion direction that would correspond to human perceived motion. No motion was detected by the network on the control stimuli presenting a static image. Similar results were obtained with nine other illusions. PredNet is undoubtedly susceptible to static motion illusions, moving elements are pointed correctly and arranged in a pattern corresponding to real displacement. In some cases, the movement direction is indicated incorrectly. The susceptibility of the neural network to illusions varies.