|
|
Biography |
|
|
José Ali Moreno is a former Professor of the Faculty of Science at Universidad Central de Venezuela, the premier university of Venezuela. He is a member and co-founder of the Group of Digital Science at Universidad Simon Bolivar, where he continues his pioneering research on emergent computing and data mining. He is also the CIO of Corporation B4B, a Venezuelan start-up of digital predictive medicine. Dr. Moreno received his degree in physics from Universidad Central de Venezuela, in Caracas, and a scientist, he was trained at the Institut fur Physikalishe Chemie and Elektrochemie of the Universitaet Hannover in Germany and holds a Dr.rer.nat. from this university. He has about 80 scientific publications consisting of research papers, and letters, as well as book chapters edited by eminent scientists. Born in Caracas, Dr. Moreno has been speaker at universities of the USA, EU, and Latin America, as well as a consultant for the oil industry of Venezuela.
|
|
|
|
|
|
|
Abstract |
|
|
|
|
TopFoodMap: Self Organizing Kohonen Map for the Visualization of Foods |
|
|
|
|
|
Foods, as registered in food databases, are represented as multi-component nutrient content patterns that can be considered points in a high-dimensional "food space". Clearly the dimensionality of this space depends on the number of the considered nutrient components which typically is over 10. The problem with the handling of this kind of data is that most of us humans are not particularly adept at visualizing data of more than three dimensions. Hence many interesting questions that arise for foods in "food space" like, among others, How do the foods relate to each other? Where are the top quality foods located? How does a food change when one or more of its components are altered? Do they form clusters? How does these clusters relate? Is there any geometrical structure among them? What shape do they form? are very difficult to answer and understand. In this work we describe the application of the Kohonen Self Organizing Map (SOM) in the development of TopFoodMap, a visualization tool for foods represented by their nutrient content patterns. The Kohonen SOM is a class of dimensionality reduction technique that induces a mapping between "data space" and the topographic map that seeks to preserve some of the structure of the data in the geometric structure of the mapping. The term “geometric structure” refers to the relationships between distances in data space and the distances in the projection to the topographic map. The Kohonen Network is an unsupervised Artificial Neural Network that must be trained by a set of training patterns properly selected and representing the main topological properties of the data space. In this work the data space or food space is defined by foods in their nutrient content representation taken from the USDA National Nutrient Database in its abbreviated form. This database is produced by the United States Department of Agriculture and provides the nutritional content of many generic and proprietary-branded foods. Released in September 2015, the current release, Standard Reference 28 (SR28) contains data on 8,790 food items and up to 46 food components. The training set is a selection of 1450 instances of foods selected from the database with the criterion that they be raw generic or scarcely processed foods. The dimension of the nutrient content patterns of these foods was further reduced to 12 by considering different cases like mainly macro-nutrients, or rather the vitamin or mineral content. The map resulting from the training of the Kohonen network can be visualized using de U-matrix representation of the SOM where the Euclidean distance between the weight vectors of neighboring neurons is depicted in a color image. The topology preservation property implies that if the projections of two points are close in the map, it is because, they are similar, in nutrient content, in the original food space. The closeness criterion is usually the Euclidean distance between the data patterns. In this sense this map constitutes a 2-dimensional landscape where the nutrient content patterns, representing any kind of foods, can be plotted as high quality graphics allowing the exploitation of the powerful human visual processing capabilities in the answering of the aforementioned questions. The results are presented on-line by the visualization of different maps and corresponding examples produced by the TopFoodMap application implemented on the cloud. |
|
|
|
|