Pattern recognition systems

Surveys are made of both character recognition and image
processing. The need to apply image processing techniques
to character recognition is pointed out. The
fields are then combined and tested in sample programs.
Simulations are made of recognition systems with and
without image preprocessing. Processing techniques
applied utilize Walsh-Hadamard transforms and l ocal window
operators. Results indicate that image prepro c ess i ng
improves recognition rates when noise degrades input
images. A system architecture is proposed for a hardware
based video speed image processor operating on local
image windows. The possible implementation of this
processor is outlined.

The existing Group Method of Data Handling (GMDH) algorithm has characteristics that are ideal for neural network design. This thesis introduces a new algorithm that applies some of the best characteristics of GMDH to neural network design and develops a Pruning based Regenerated Network by discarding the neurons in a layer which don't contribute for the creation of neurons in next layer. Unlike other conventional algorithms, which generate a network which is a black box, the new algorithm provides visualization of the network displaying all the neurons in the network. The algorithm is general enough that it will accept any number of inputs and any sized training set. To show the flexibility of the Pruning based Regenerated Network, this algorithm is used to analyze different combinations of drugs and determine which pathways in these networks interact and determine the combination of drugs that take advantage of these interactions to maximize a desired effect on genes.

This research aims at proposing a model for visual pattern recognition inspired by the neural circuitry in the brain. Our attempt is to propose few modifications in the Alopex algorithm and try to use it for the calculations of the receptive fields of neurons in the trained network. We have developed a small-scale, four-layered neural network model for simple character recognition as well as complex image patterns, which can recognize the patterns transformed by affine conversion. Here Alopex algorithm is presented as an iterative and stochastic processing method, which was proposed for optimization of a given cost function over hundreds or thousands of iterations. In this case the receptive fields of the neurons in the output layers are obtained using the Alopex algorithm.

This thesis discusses the implementation of a feed forward NN using time series model to predict the sudden rise or sudden crash of a company's stock prices. The theory behind this prediction system is Pattern recognition. Pattern recognition techniques for time-series prediction are based on structural matching of the current state of the time-series with previously occurring states in historical data for making predictions. This study reports the result of attempts to predict the Motorola stock price index using artificial neural networks (ANN). Daily data from January 1999 to December 2001 were taken from the NYSE. These data are classified based on criteria of an n% fall or rise of price corresponding to the previous day close price. A novel method using Hurst exponent is used in selecting the data set. These data are fed into a Back Propagated Neural Network. The number of hidden layers and number of neurons are systematically selected to implement a better predicting machine. The implemented model is tested using both interpolated and extrapolated data. Fundamental limitations and inherent difficulties when using neural networks for processing of high noise, small sample size signals are also discussed. Results of the prediction are presented and an elaborate discussion is made comparing the results.

In this thesis, application of GMDH Algorithm to real life problems is studied. A particular type of GMDH Algorithm namely TMNN is chosen for this purpose. An effort is made to forecast S&P Index Closing Value with the help of the forecaster. The performance of the TMNN Algorithm is simulated by implementing a tool in C++ for developing forecast models. The validation of this simulation tool is carried out with Sine Wave Values and performance analysis is done in a noisy environment. The noisy environment tests the TMNN forecaster for its robustness. The primary goal of this research is to develop a simulation software based on TMNN Algorithm for forecasting stock market index values. The main inputs are previous day's closing values and the output is predicted closing index.

An un-supervised learning algorithm on application level intrusion detection, named Graph Sequence Learning Algorithm (GSLA), is proposed in this dissertation. Experiments prove its effectiveness. Similar to most intrusion detection algorithms, in GSLA, the normal profile needs to be learned first. The normal profile is built using a session learning method, which is combined with the one-way Analysis of Variance method (ANOVA) to determine the value of an anomaly threshold. In the proposed approach, a hash table is used to store a sparse data matrix in triple data format that is collected from a web transition log instead of an n-by-n dimension matrix. Furthermore, in GSLA, the sequence learning matrix can be dynamically changed according to a different volume of data sets. Therefore, this approach is more efficient, easy to manipulate, and saves memory space. To validate the effectiveness of the algorithm, extensive simulations have been conducted by applying the GSLA algorithm to the homework submission system at our computer science and engineering department. The performance of GSLA is evaluated and compared with traditional Markov Model (MM) and K-means algorithms. Specifically, three major experiments have been done: (1) A small data set is collected as a sample data, and is applied to GSLA, MM, and K-means algorithms to illustrate the operation of the proposed algorithm and demonstrate the detection of abnormal behaviors. (2) The Random Walk-Through sampling method is used to generate a larger sample data set, and the resultant anomaly score is classified into several clusters in order to visualize and demonstrate the normal and abnormal behaviors with K-means and GSLA algorithms. (3) Multiple professors' data sets are collected and used to build the normal profiles, and the ANOVA method is used to test the significant difference among professors' normal profiles. The GSLA algorithm can be made as a module and plugged into the IDS as an anomaly detection system.

Researchers from a wide range of fields have discovered the benefits of applying neural networks to pattern recognition problems. Although applications for neural networks have increased, development of tools to design these networks has been slower. There are few comprehensive network development methods. Those that do exist are slow, inefficient, and application specific, require predetermination of the final network structure, and/or result in large, complicated networks. Finding optimal neural networks that balance low network complexity with accuracy is a complicated process that traditional network development procedures are incapable of achieving. Although not originally designed for neural networks, the Group Method of Data Handling (GMDH) has characteristics that are ideal for neural network design. GMDH minimizes the number of required neurons by choosing and keeping only the best neurons and filtering out unneeded inputs. In addition, GMDH develops the neurons and organizes the network simultaneously, saving time and processing power. However, some of the qualities of the network must still be predetermined. This dissertation introduces a new algorithm that applies some of the best characteristics of GMDH to neural network design. The new algorithm is faster, more flexible, and more accurate than traditional network development methods. It is also more dynamic than current GMDH based methods, capable of creating a network that is optimal for an application and training data. Additionally, the new algorithm virtually guarantees that the number of neurons progressively decreases in each succeeding layer. To show its flexibility, speed, and ability to design optimal networks, the algorithm was used to successfully design networks for a wide variety of real applications. The networks developed using the new algorithm were compared to other development methods and network architectures. The new algorithm's networks were more accurate and yet less complicated than the other networks. Additionally, the algorithm designs neurons that are flexible enough to meet the needs of the specific applications, yet similar enough to be implemented using a standardized hardware cell. When combined with the simplified network layout that naturally occurs with the algorithm, this results in networks that can be implemented using Field Programmable Gate Array (FPGA) type devices.

The literature regarding biological motion suggests that people may accurately identify and recognize the gender of others using movement cues in the absence of typical identifiers. This study compared identification and gender judgments of traditional point-light stimuli to skeleton stimuli. Controlling for previous experience and execution of actions, the frequency and familiarity of movements was also considered. Watching action clips, participants learned to identify 4 male and 4 female actors. Participants then identified the corresponding point-light or skeleton displays. Although results indicate higher than chance performance, no difference was observed between stimuli conditions. Analyses did show better gender recognition for common as well as previously viewed actions. This suggests that visual experience influences extraction and application of biological motion. Thus insufficient practice in relying on movement cues for identification could explain the significant yet poor performance in biological motion point-light research.

In order to facilitate the development, discussion, and advancement of the relatively new subfield of Artificial Intelligence focused on generating narrative content, the author has developed a pattern language for generating narratives, along with a new categorization framework for narrative generation systems. An emphasis and focus is placed on generating the Fabula of the story (the ordered sequence of events that make up the plot). Approaches to narrative generation are classified into one of three categories, and a pattern is presented for each approach. Enhancement patterns that can be used in conjunction with one of the core patterns are also identified. In total, nine patterns are identified - three core narratology patterns, four Fabula patterns, and two extension patterns. These patterns will be very useful to software architects designing a new generation of narrative generation systems.

This thesis presents a comparative analysis of various low-frequency sonar signature representations and their ability to discriminate between proud targets of varying physical parameters. The signature representations used include: synthetic aperture sonar (SAS) beamformed images, acoustic color plot images, and bispectral images. A relative Mean-Square Error (rMSE) performance metric and an effective Signal-to-Noise Ratio (SNReff) performance metric have been developed and implemented to quantify the target differentiation. The analysis is performed on a subset of the synthetic sonar stave data provided by the Naval Surface Warfare Center - Panama City Division (NSWC-PCD). The subset is limited to aluminum and stainless steel, thin-shell, spherical targets in contact with the seafloor (proud). It is determined that the SAS signature representation provides the best, least ambiguous, target differentiation with a minimum mismatch difference of 14.5802 dB. The acoustic color plot and bispectrum representations resulted in a minimum difference of 9.1139 dB and 1.8829 dB, respectively