Rosemary D. Paradis

Using High Performance Computing to Explore Large Complex Bioacoustic Soundscapes: Case Study for Right Whale Acoustics.

Abstract This paper describes ongoing work to investigate the development of a complex system designed for extracting information from large acoustic datasets. The system, called DeLMA is based on integrating advanced machine learning with high performance computing (HPC). The goal of this work is to provide the capability to accurately detect and classify whale sounds in large acoustics datasets collected using underwater sensors. The case study for this work is focused on detecting the acoustic communication signals of the North Atlantic Right Whale, Eubalaena glacialis , and uses data collected in the Stellwagen Bank National Marine Sanctuary (SBNMS), USA. A summary of the work done for developing a complex detection-classification system and brief description of several algorithms that are used for classifying whale sounds will be covered. A brief discussion on how standard detection algorithms can be incorporated, with no special modifications, into the HPC system for analysis will be mentioned, and two new right whale detection methods are presented, based on continuous region analysis (CRA) and histogram of oriented gradients (HOG). This paper presents a first-hand look at applying the DeLMA system and these algorithms on a large dataset containing over 60,000 channel-hours of acoustic data from the SBNMS. Results from these new detection methods are compared against Baseline algorithms. With the development of the DeLMA system, sound archives can now be explored using a powerful distributed processing architecture. This advancement will allow for rapid execution and visualization of the data using seasonal graphs called diel plots, which show the distribution of detections on a time-of-day vs. time-of-year plane. Diel plots of Baseline, CRA and HOG algorithm results reveal various large-scale features of the seasonality of whale calling behavior. Results are summarized and the authors discuss future areas for study, especially those relate to handling other big passive acoustic data projects.

Cognitive Category Learning

Abstract Categories play a fundamental role in our daily lives and are the basis for decision making in most professions (e.g., intelligence analysis, healthcare, and engineering). Categories are a set of objects or events that have similar features and are grouped together because of their similarity. Many categories are acquired as a child (e.g., tools, fruit) but we continue to learn and apply new categories throughout life. Categories range from concrete (e.g., a set of physical objects like houses , dogs… ) to abstract (e.g., political ideologies , movie genres… ) and narrow (e.g., rifles ) to broad (e.g., weapons ). People in a variety of knowledge intensive fields (e.g., analysts, commanders, medical doctors) recognize categories in streams of data and make a decision about how to act (or pass the information to a decision maker). Such a categorization system is important now, especially because of the large amount of information that people need to sift through to do their jobs on a regular basis. When decision support systems (DSSs) are applied to support human decision making by automatically recognizing categories, these systems are often not practical in complex dynamic real world environments. Rule based DSSs are limited because it is difficult to develop and maintain large complex rule sets. Current machine learning based DSSs are sometimes limited because they require data that encompasses all of the possible variations as examples to learn from and necessitate significant effort to develop and maintain this training data. This paper describes a cognitive category learning system that uses machine learning and natural language processing (NLP) techniques to categorize unstructured documents or semi-structured objects, such as emails, which we used in this experiment. Our system uses several methods to do this categorization of emails and then arbitrates the best solution based on the individual classifier results. This result provides a more confident answer with less chance of false positive and false negative outcomes. Our system also generates a metadata topic summary for each document or email.

Finding Semantic Equivalence of Text Using Random Index Vectors

Abstract The challenges of machine semantic understanding have not yet been satisfactorily solved by automated methods. In our approach, the semantics and syntax of words, phrases and documents are represented by deep semantic vectors that capture both the structure and semantic meaning of the language. Our experiment reproduces the experiment done by Patwardhan and Pedersen 2006, but uses random index vectors for the words, glosses and tweets. Our model first determines random index vectors from glosses and definitions for words from WordNet. From these foundational semantic vectors, random index vectors that represent phrases, sentences or tweets are determined. Our set of algorithms relies on high-dimensional distributed representations, and their effectiveness and versatility derive from the unintuitive properties of such representations: from the mathematical properties of high-dimensional spaces. High-dimensional vector representations have been used successfully in modeling human cognition, such as memory and learning. Our semantic vectors are high-dimensional and capture the meaning of a language expression, such as a word, phrase, query, news article, story or a message. A key benefit of our method is that the dimensionality of the vectors remains constant as we add data; this also allows good generalization to rarely seen words, which “borrow strength” from their more frequent neighbors.

Detection of Groups in Non-Structured Data

Abstract This paper describes a method to automatically discover features which distinguish the language use of cultural subgroups operating within the same broader language/culture. Sociolinguists have long known that special features such as vocabulary use, phonetic features (like accents), and syntactic characteristics develop within the in-group language of frequently interacting subgroups. These features set apart the groups language from the discourse of others speaking the same broader language. Our interest is to learn these features automatically and use them to distinguish the writing of one subgroup from another. The s pecial vocabulary and jargon of various subgroups has often been catalogued. This research focuses instead on syntactic differences which can be learned from digital text and the specialized use of vocabulary which is not topic or domain specific (e.g. we deliberately omit domain related jargon.) Our main data source is blogs and related discussions from a number of North American subculture groups, such as radical feminists and militia groups. In this paper we present our findings on looking for blogs whose participants have a particular subcultural affiliation, designated as “blogs of interest.” Our hypothesis is that we can ignore the particular topic of a blog discussion, through means described in the paper, and isolate other linguistic indicators that help us determine whether or not a blog is “of interest” We start with an overview of the process of training our system and describe its use in identifying blogs of the desired cultural subgroup. We then describe in detail the training process in which a series of grams are scored and aggregated to find key, highly indicative blog passages. The last section reports on an experiment we conducted that proved the concept against several North American English language blogging communities

Preface to Part II Computational Intelligence and Machine Learning

In computation intelligence and machine learning researchers often look to other fields of study, such as anthropology, biology, computer science, psychology, philosophy and neuroscience to shape the algorithms being developed. Intelligence, which is generally attributed to humans, animals and plants, can be characterized by abstract thought, understanding, communication, reasoning, learning, planning, emotional intelligence and problem solving. At the center of this capability are intelligent systems that have both the ability to adapt and the necessary complex analytics for autonomous behaviour that can provide intelligent control. Such abilities can aid those in industry and academia currently searching for novel and adaptable solutions to complex problems. Within industry, the main priorities for business and government research and development include decentralization, uncertainly and complexity. In the coming years there will be a need for programmatic efficiencies because of the insufficient manpower to support complex missions. There will also be harsh environments that do not reasonably permit humans to enter and sustain activity as well as new requirements for adaptive autonomous control of vehicle systems in the face of unpredictable environments and challenging missions. Problems of uncertainty and brittleness in this ever changing environment will need to be resolved with machine reasoning and intelligence, human/autonomous system interaction, and collaboration and scalable teaming of autonomous systems. Neural networks, fuzzy logic and evolutionary programming are just a few of the techniques that have been developed to solve some of these problems. The algorithms, techniques and applications described in this section move the fields of computational intelligence and machine learning in new directions and aim to contribute to the Complex Adaptive Systems Conference (2012) in the following categories: