Pattern Recognition Speech Image Handwriting etc Wajahat Qadeer Rebecca Schultz Ernesto Staroswiecki VOICE RECOGNITION Automatic conversion of speech into textual representation Preprocessing Partitioning and compression of speech into a stream of feature vectors Recognition Identification of words through an optimal path of a graph most time consuming VOICE RECOGNITION Preprocessing loop oriented with fixed bounds and no loop carried dependencies High DLP with provision for TLP Computationally intensive requiring floating point and integer operations Small working set and memory foot print with regular data access patterns High degree of spatial and temporal locality VOICE RECOGNITION Recognition Large working set with highly irregular control and data access patterns Big memory foot print during initialization requiring high bandwidth Large caches and bigger block size reduce cache misses Little ILP but TLP offers substantial gains Algorithmic changes can exploit data locality VOICE RECOGNITION Other Algorithms Dynamic Time Warping hidden Markov modeling Neural Networks etc Benchmarks Common benchmarks are RASTA pre processing and Sphinx recognition Scaling Trends Complex search mechanisms requiring more computational resources Large sets of databases requiring tremendous memory IMAGE RECOGNITION Also a 3 step process Edge detection Filtering Image processing Characterization Matching IMAGE RECOGNITION Processing Characterization We need to find image descriptors Shape contexts Fourier descriptors etc Similar characteristics to voice recognition preprocessing except Not necessary to use floating point or excessive computation Yet more points to look at which grow with the size of the image And although the memory access pattern is very regular is important to remember that now we are looking at a 2D window IMAGE RECOGNITION Matching Once again similar to voice recognition but problems really exacerbated Several algorithms SVMs Shortest Augmenting Path etc Remember that dictionary must be MUCH larger Little ILP some DLP but mostly TLP Topics to explore CAMs prefetching but be careful HANDWRITING RECOGNITION Special case of image recognition Similar algorithms for selecting descriptors and matching Neural Nets Hidden Markov Models etc Matching library is small and fixed size Rarely done in hardware Low data rate Scaling Constant number of descriptor points irrespective of sample size Limited opportunities for extensions