The science of designing machines to extract meaningful information from digital images, videos, and other visual inputs is known as Computer Vision (CV). Deep learning algorithms cope CV problems by automatically learning task-specific features. Especially, Deep Neural Networks (DNNs) have become an essential component in CV solutions due to their ability to encode large amounts of data and capacity to manipulate billions of model parameters. Unlike machines, humans learn by rapidly constructing abstract models. This is undoubtedly due to the fact that good teachers supply their students with much more than just the correct answer; they also provide intuitive comments, comparisons, and explanations. In deep learning, the availability of such auxiliary information at training time (but not at test time) is referred to as learning by Privileged Information (PI). Typically, predictions (e.g., soft labels) produced by a bigger and better network teacher are used as structured knowledge to supervise the training of a smaller network student, helping the student network to generalize better than that trained from scratch. This dissertation focuses on the category of deep learning systems known as Collaborative Learning, where one DNN model helps other models or several models help each other during training to achieve strong generalization and thus high performance. The question we address here is thus the following: how can we take advantage of PI for training a deep learning model, knowing that, at test time, such PI might be missing? In this context, we introduce new methods to tackle several challenging real-world computer vision problems. First, we propose a method for model compression that leverages PI in a teacher-student framework along with customizable block-wise optimization for learning a target-specific lightweight structure of the neural network. In particular, the proposed resource-aware optimization is employed on suitable parts of the student network while respecting the expected resource budget (e.g., floating-point operations per inference and model parameters). In addition, soft predictions produced by the teacher network are leveraged as a source of PI, forcing the student to preserve baseline performance during network structure optimization. Second, we propose a multiple-model learning method for action recognition, specifically devised for challenging video footages in which actions are not explicitly visualized, but rather, only implicitly referred. We use such videos as stimuli and involve a large sample of subjects to collect a high-definition EEG and video dataset. Next, we employ collaborative learning in a multi-modal setting i.e., the EEG (teacher) model helps the video (student) model by distilling the knowledge (implicit meaning of visual stimuli) to it, sharply boosting the recognition performance. The goal of Unsupervised Domain Adaptation (UDA) methods is to use the labeled source together with the unlabeled target domain data to train a model that generalizes well on the target domain. In contrast, we cast UDA as a pseudo-label refinery problem in the challenging source-free scenario i.e., in cases where the source domain data is inaccessible during training. We propose Negative Ensemble Learning (NEL) technique, a unified method for adaptive noise filtering and progressive pseudo-label refinement. In particular, the ensemble members collaboratively learn with a Disjoint Set of Residual Labels, an outcome of the output prediction consensus, to refine the challenging noise associated with the inferred pseudo-labels. A single model trained with the refined pseudo-labels leads to superior performance on the target domain, without using source data samples at all. We conclude this dissertation with a method extending our previous study by incorporating Continual Learning in the Source-Free UDA. Our new method comprises of two stages: a Source-Free UDA pipeline based on pseudo-label refinement, and a procedure for extracting class-conditioned source-style images by leveraging the pre-trained source model. While stage 1 holds the same collaborative peculiarities, in stage 2, the collaboration exists in an indirect manner i.e., it is the source model that provides the only possibility to generate source-style synthetic images which eventually helps the final model in preserving good performance on both source and target domains. In each study, we consider heterogeneous CV tasks. Nevertheless, with an extensive pool of experiments on various benchmarks carrying diverse complexities and challenges, we show that the collaborative learning framework outperforms the related state-of-the-art methods by a considerable margin.

Collaborative Learning in Computer Vision

AHMED, WAQAR
2022

Abstract

The science of designing machines to extract meaningful information from digital images, videos, and other visual inputs is known as Computer Vision (CV). Deep learning algorithms cope CV problems by automatically learning task-specific features. Especially, Deep Neural Networks (DNNs) have become an essential component in CV solutions due to their ability to encode large amounts of data and capacity to manipulate billions of model parameters. Unlike machines, humans learn by rapidly constructing abstract models. This is undoubtedly due to the fact that good teachers supply their students with much more than just the correct answer; they also provide intuitive comments, comparisons, and explanations. In deep learning, the availability of such auxiliary information at training time (but not at test time) is referred to as learning by Privileged Information (PI). Typically, predictions (e.g., soft labels) produced by a bigger and better network teacher are used as structured knowledge to supervise the training of a smaller network student, helping the student network to generalize better than that trained from scratch. This dissertation focuses on the category of deep learning systems known as Collaborative Learning, where one DNN model helps other models or several models help each other during training to achieve strong generalization and thus high performance. The question we address here is thus the following: how can we take advantage of PI for training a deep learning model, knowing that, at test time, such PI might be missing? In this context, we introduce new methods to tackle several challenging real-world computer vision problems. First, we propose a method for model compression that leverages PI in a teacher-student framework along with customizable block-wise optimization for learning a target-specific lightweight structure of the neural network. In particular, the proposed resource-aware optimization is employed on suitable parts of the student network while respecting the expected resource budget (e.g., floating-point operations per inference and model parameters). In addition, soft predictions produced by the teacher network are leveraged as a source of PI, forcing the student to preserve baseline performance during network structure optimization. Second, we propose a multiple-model learning method for action recognition, specifically devised for challenging video footages in which actions are not explicitly visualized, but rather, only implicitly referred. We use such videos as stimuli and involve a large sample of subjects to collect a high-definition EEG and video dataset. Next, we employ collaborative learning in a multi-modal setting i.e., the EEG (teacher) model helps the video (student) model by distilling the knowledge (implicit meaning of visual stimuli) to it, sharply boosting the recognition performance. The goal of Unsupervised Domain Adaptation (UDA) methods is to use the labeled source together with the unlabeled target domain data to train a model that generalizes well on the target domain. In contrast, we cast UDA as a pseudo-label refinery problem in the challenging source-free scenario i.e., in cases where the source domain data is inaccessible during training. We propose Negative Ensemble Learning (NEL) technique, a unified method for adaptive noise filtering and progressive pseudo-label refinement. In particular, the ensemble members collaboratively learn with a Disjoint Set of Residual Labels, an outcome of the output prediction consensus, to refine the challenging noise associated with the inferred pseudo-labels. A single model trained with the refined pseudo-labels leads to superior performance on the target domain, without using source data samples at all. We conclude this dissertation with a method extending our previous study by incorporating Continual Learning in the Source-Free UDA. Our new method comprises of two stages: a Source-Free UDA pipeline based on pseudo-label refinement, and a procedure for extracting class-conditioned source-style images by leveraging the pre-trained source model. While stage 1 holds the same collaborative peculiarities, in stage 2, the collaboration exists in an indirect manner i.e., it is the source model that provides the only possibility to generate source-style synthetic images which eventually helps the final model in preserving good performance on both source and target domains. In each study, we consider heterogeneous CV tasks. Nevertheless, with an extensive pool of experiments on various benchmarks carrying diverse complexities and challenges, we show that the collaborative learning framework outperforms the related state-of-the-art methods by a considerable margin.
25-feb-2022
Inglese
MURINO, VITTORIO
DEL BUE, ALESSIO
VALLE, MAURIZIO
Università degli studi di Genova
File in questo prodotto:
File Dimensione Formato  
phdunige_4611071_1.pdf

accesso aperto

Dimensione 16.39 MB
Formato Adobe PDF
16.39 MB Adobe PDF Visualizza/Apri
phdunige_4611071_2.pdf

accesso aperto

Dimensione 12.17 MB
Formato Adobe PDF
12.17 MB Adobe PDF Visualizza/Apri

I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14242/67832
Il codice NBN di questa tesi è URN:NBN:IT:UNIGE-67832