Integrating Motion into Vision Models for Better Visual Prediction

We demonstrate an improved vision system that learns a model of its environment using a self-supervised, predictive learning method. The system includes a pan-tilt camera, a foveated visual input, a saccading reflex to servo the foveated region to areas high prediction error, input frame transformation synced to the camera motion, and a recursive, hierachical machine learning technique based on the Predictive Vision Model. In earlier work, which did not integrate camera motion into the vision model, prediction was impaired and camera movement suffered from undesired feedback effects. Here we detail the integration of camera motion into the predictive learning system and show improved visual prediction and saccadic behavior. From these experiences, we speculate on the integration of additional sensory and motor systems into self-supervised, predictive learning models.

Saccadic Predictive Vision Model with a Fovea

We propose a model that emulates saccades, the rapid movements of the eye, called the Error Saccade Model, based on the prediction error of the Predictive Vision Model (PVM). The Error Saccade Model carries out movements of the model's field of view to regions with the highest prediction error. Comparisons of the Error Saccade Model on Predictive Vision Models with and without a fovea show that a fovea-like structure in the input level of the PVM improves the Error Saccade Model's ability to pursue detailed objects in its view. We hypothesize that the improvement is due to poorer resolution in the periphery causing higher prediction error when an object passes, triggering a saccade to the next location.

Unsupervised Learning from Continuous Video in a Scalable Predictive Recurrent Network

Understanding visual reality involves acquiring common-sense knowledge about countless regularities in the visual world, e.g., how illumination alters the appearance of objects in a scene, and how motion changes their apparent spatial relationship. These regularities are hard to label for training supervised machine learning algorithms; consequently, algorithms need to learn these regularities from the real world in an unsupervised way. We present a novel network meta-architecture that can learn world dynamics from raw, continuous video. The components of this network can be implemented using any algorithm that possesses three key capabilities: prediction of a signal over time, reduction of signal dimensionality (compression), and the ability to use supplementary contextual information to inform the prediction. The presented architecture is highly-parallelized and scalable, and is implemented using localized connectivity, processing, and learning. We demonstrate an implementation of this architecture where the components are built from multi-layer perceptrons. We apply the implementation to create a system capable of stable and robust visual tracking of objects as seen by a moving camera. Results show performance on par with or exceeding state-of-the-art tracking algorithms. The tracker can be trained in either fully supervised or unsupervised-then-briefly-supervised regimes. Success of the briefly-supervised regime suggests that the unsupervised portion of the model extracts useful information about visual reality. The results suggest a new class of AI algorithms that uniquely combine prediction and scalability in a way that makes them suitable for learning from and --- and eventually acting within --- the real world.