Titolo della tesi: Exploring Image Classification Problems with Sample- and Class-Deficient Data Distributions
In this thesis, we deal with images, i.e., 3D arrays, of different types such as RGB, grayscale, multi-spectral, or synthetically generated ones.
Image classification is among the most established tasks for machine learning models.
The problem is formally modeled by an unknown joint probability distribution over samples and corresponding labels from which training and testing data are sampled.
We are going to treat two variants of this modeling stemming from the hypothesis of having deficiencies induced to either one of the two marginals.
Precisely, sample deficiency, which derives from accessing only a limited portion of the sample distribution, and class deficiency, which stems from restricted exposure to the overall category space.
Practically speaking, a sample-deficient distribution generates training datasets of small numerosity for current standards (i.e., one or two orders of magnitudes) and hinders large-scale training of networks.
On the other hand, a class-deficient distribution does not provide samples of one or more classes.
The study of these two research problems has significant practical implications since many recent advances in machine learning have only been achieved by pre-training on massive datasets.
Unfortunately, labeling data is a costly and time-consuming process that can not always be achieved.
In many applications, large portions of the input space might be hardly accessible (sample deficiency), or eventually totally unknown (class deficiency).
Therefore, the practical use of deep neural networks is inherently related to the amount of available labeled training data.
Our contributions for sample-deficient scenarios comprise extensive empirical results concerning popular regularization strategies for deep networks such as data augmentation, dropout, and ensembles.
Because of the current fragmentation of the state of the art, we propose the first systematic literature overview and common benchmark to allow for objective comparisons between published methods.
The broad re-evaluation of state-of-the-art methods on our benchmark led us to the surprising and sobering result that the standard cross-entropy is a highly competitive baseline.
Following such findings, we propose a strong baseline.
Ultimately, for class-deficient classification problems, we delve deep into the application domain of cyber-physical systems.
We treat detection and open-world recognition of anomalies. We experiment with multiple log-to-image transformations, neural architectures, and modeling functions performing a large comparative analysis exploring the overmentioned directions.
The methodology presented is intended to be used as a guideline to face the challenging problems presented in this thesis, and, more importantly, to foster the progress of research in these research areas.
We provide our open-source implementations for the multiple datasets and solutions presented.