Year: 2017
Type: dissertation
Abstract: Image classification is one of the most active challenging problems in computer vision field. Taking this to Deep Neural Networks with systems that are able to deal with large data sets as it can be ImageNet. Large Convolutional Networks as VGG-16 used in this work have recently demonstrated impressive classification performances. This work is focused on novel techniques for Incremental Learning stages for face recognition, which is an important open problem in artificial intelligence. The main challenge of this work is the development of incrementally learning systems that learn about more and more concepts over time. Most of the actual methods that use incremental learning in "online" or "offline" stages. This thesis focuses on "offline" incremental stages where the data available is distributed in batches of classes. Since the necessity to deal with a continuous training stages, some well-established methods for transfer learning are applied by the author to run the experiments. Preserving knowledge is the most challenge task to deal using incremental learning techniques. The actual research is on apply incremental learning in natural systems where for example, it is not considered to store all the old training data to make a new model when new data comes available. Another interesting concept for incremental learning systems is "lifelong" learning, which are related to the methods analyzed in this work since the system proposed also learn from a sequence of different tasks. The similarity of multi-task learning and "lifelong" learning is that they both use shared information across tasks to help learning, but also, multi-task learning is not able to grow the number of tasks over time preserving the knowledge.
Author: Eric Presas Valga
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2801395104