| Summary: | Image representations have crucial importance in computer vision systems as they encode the pixels inner and relational information in a computationally tractable form, allowing algorithms to reason about the visual content and take decisions about it. Image representation learning aims to provide an automatized process for composing the most appropriate representations for a given computer vision task. The state-of-the-art of this research area - Deep Learning-based techniques - has achieved, in recent years, major advances in solving problems studied for decades by the Artificial Intelligence community and beat records in several pattern-recognition tasks. However, they usually present high computational complexity and demand a huge amount of resources such as storage memory, working memory, computational power, and energy consumption. Furthermore, they typically require large sets of labeled data to produce effective models. Motivated by these disadvantages, we combine three factors in order to pro- duce resource-efficient representations: incremental learning, that optimizes representations without constructing them from scratch avoiding complexity and high resource consumption; evolutionary algorithms, which provides scalable optimization, efficient search-space cover, and natural suitability for combinatorial problems; and quantization optimization, which often provides compaction without reducing the number of parameters. We address two important branches of image representation learning: shallow and deep representations. Regarding the former, we propose the optimization of shallow representations and introduce a Genetic-Algorithm based approach that optimizes the color-quantization of feature-engineered representations for improved ef- fectiveness and compactness. We evaluated this methodology in content-based image retrieval tasks and obtained representations with significantly improved precision and reduced size besides surpassing deep-learning-based baselines. Regarding the latter, we study the optimization of deep representations through model compression and pro- pose a post-training mixed-precision quantization method to optimize the weights and activations of convolutional neural models using a multi-objective Genetic-Algorithm search. We evaluated this methodology in image classification using Imagenet dataset and obtained compression in post-training quantization with small accuracy drops. Results confirm Genetic Algorithm optimization as a promising approach for highly effective and resource-efficient learning in future methodologies.
|
|---|