Deep Machine Learning
The problem: capturing both broad spatial and multi-scale temporal dependencies in observed input data in order to perform recognition and prediction tasks with complex datasets.
The challenge in constructing efficient, invariant representation of complex patterns pertains to both the spatial and temporal attributes. In the context of spatial information, high-dimensional inputs (e.g. millions of pixels in an image) span a very large space of possible observations. Mainstream “shallow” approaches reduce input dimensionality by applying pre-processing schemes that attempt to extract key features pertaining to the data observed. However, this feature extraction process is highly application-dependent and often fails to attain the necessary information that would enable accurate pattern inference.
The solution: HD-RNN, a scalable, multi-layer biologically-inspired auto-associative memory architecture that is able to capture, in an unsupervised manner, spatiotemporal dependencies that exist in observed input data.
Deep learning approaches overcome the limitations of pre-processing feature extraction schemes by constructing a hierarchical feature tensor that maps to a decision space using a simple classifier. The feature extraction process in HD-RNN is data-driven (rather than being engineered for a particular application domain) and is driven by regularities in the input data. This is a lossy process, in which not all pieces of the original data are stored with absolute accuracy; rather, detail is lost as one considers broader spatial scope and/or longer-term time intervals. As a result, salient features are formed, which can then be used for various robust pattern recognition applications.
A key factor in performing accurate recognition and prediction tasks pertains to the framework ability to consider information originating from different sensory sources, in order to reach a well-supported decision. Often times, complete understanding of a system's operational condition can only be reached by consolidating data received from these various sources. One of the unique advantages of deep-learning schemes is that they inherently facilitate fusion of information originating from different modalities.