There has been a lot of progress in the field of invariant object recognition/categorization in the last decade with several methods trying to mimic functioning of the human visual system (e.g. Neocognitron, HMAX, VisNet). Examining those brain regions is a very difficult task with myriads of details to be considered. To simplify modeling approaches, Jeff Hawkins [1] proposed a framework of three basic principles that might underlie computations in regions of the neocortex. These also form the basis for a capable object recognition system named "Hierarchical Temporal Memory" (HTM).1. Learning of temporal sequences for creating invariance to transformations contained in the training data.2. Learning in a hierarchical structure, in which lower level knowledge can be reused in higher level context and thereby makes memory usage efficient.3. Prediction of future signals for disambiguation of noisy input by feedback.In my thesis I have developed and efficiently implemented two related artificial neural systems relying on these principles, the Temporal Correlation Graph (TCG) and the Temporal Correlation Net (TCN). Both are hierarchical neural networks made up of alternating levels of spatial and temporal neurons located at subsampled image positions called nodes. Spatial neurons represent spatial patterns, which on the lowest level are visual features from training images and on higher levels composed patterns of activities. Temporal neurons represent but groups of spatial patterns that tend to follow each other in time. Neural activities are stored in nodes which define the architecture of the systems. In each node any neuron of the same level can become active. Connections from temporal to the next higher spatial level are convergent collecting input from 3 × 3 spatial neurons in one node. Convergence is chosen thus that at the top of the network only one node of temporal neurons remains. These neurons represent the different object categories the system has learned. During training each of both systems observes sequences of images showing objects of different categories undergoing transformations in viewing conditions (scaling, rotation in depth, illumination changes etc.) to which the top level responses shall become invariant. First spatial patterns on the lowest level are learned, thenCorrespondence to:
Nearest neighbor search in metric spaces is an important task in pattern recognition because it allows a query pattern to be associated with a known pattern from a learned dataset. In low-dimensional spaces a lot of good solutions exist that minimize the number of comparisons between patterns by partitioning the search space using tree structures. In high-dimensional spaces tree methods become useless because they fail to prevent scanning almost the complete dataset. Locality sensitive hashing methods solve the task approximately by grouping patterns that are nearby in search space into buckets. Therefore an appropriate hash function has to be known that is highly likely to assign a query pattern to the same bucket as its nearest neighbor. This works fine as long as all the patterns are of the same dimensionality and exist in the same vector space with a complete metric. Here, we propose a locality-sensitive hashing-scheme that is able to process patterns which are built up of several possibly missing subpatterns causing the patterns to be in vector spaces of different dimensionality. These patterns can only be compared using a pseudosemimetric.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
customersupport@researchsolutions.com
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Copyright © 2025 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.