Partial least squares: a deep space odyssey

• Autor(a) principal: Artur Jordão Lima Correia
• Data de Publicação: 2020
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/36877
• topic: Computer Vision; Deep Learning; Pattern Recognition; Computação – Teses; Visão por computador – Teses; Teoria da estimativa – Teses; Reconhecimento de Padrões – Teses

Modern visual pattern recognition models are predominantly based on convolutional networks since they have led to a series of breakthroughs in different tasks. The reason for these achievements is the development of larger architectures as well as the combination of features from multiple layers of the convolutional network. Such models, however, are computationally expensive, hindering applicability on low-power and resource-constrained systems. To handle these problems, we propose three strategies. The first removes unimportant structures (neurons or layers) of convolutional networks, reducing their computational cost. The second inserts structures to design convolutional networks automatically, enabling us to build high-performance architectures. The third combines multiple layers of convolutional networks, enhancing data representation at negligible additional cost. These strategies are based on Partial Least Squares, a discriminative dimensionality reduction technique. We show that Partial Least Squares is an efficient and effective tool for removing, inserting, and combining structures of convolutional networks. Despite the positive results, Partial Least Squares is infeasible on large datasets since it requires all the data to be in memory in advance, which is often impractical due to hardware limitations. To handle this limitation, we propose a fourth approach, a discriminative and low-complexity incremental Partial Least Squares that learns a compact representation of the data using a single sample at a time, thus enabling applicability on large datasets. We assess the effectiveness of our approaches on several convolutional architectures and supervised computer vision tasks, which include image classification, face verification and activity recognition. Our approaches reduce the resource overhead of both convolutional networks and Partial Least Squares, promoting energy- and hardware-friendly models for the academy and industry scenarios. Compared to state-of-the-art methods for the same purpose, we obtain one of the best trade-os between predictive ability and computational cost.