Uncertainty, interpretability and dataset limitations in Deep Learning

Abstract

[eng] Deep Learning (DL) has gained traction in the last years thanks to the exponential increase in compute power. New techniques and methods are published at a daily basis, and records are being set across multiple disciplines. Undeniably, DL has brought a revolution to the machine learning field and to our lives. However, not everything has been resolved and some considerations must be taken into account. For instance, obtaining uncertainty measures and bounds is still an open problem. Models should be able to capture and express the confidence they have in their decisions, and Artificial Neural Networks (ANN) are known to lack in this regard. Be it through out of distribution samples, adversarial attacks, or simply unrelated or nonsensical inputs, ANN models demonstrate an unfounded and incorrect tendency to still output high probabilities. Likewise, interpretability remains an unresolved question. Some fields not only need but rely on being able to provide human interpretations of the thought process of models. ANNs, and specially deep models trained with DL, are hard to reason about. Last but not least, there is a tendency that indicates that models are getting deeper and more complex. At the same time, to cope with the increasing number of parameters, datasets are required to be of higher quality and, usually, larger. Not all research, and even less real world applications, can keep with the increasing demands. Therefore, taking into account the previous issues, the main aim of this thesis is to provide methods and frameworks to tackle each of them. These approaches should be applicable to any suitable field and dataset, and are employed with real world datasets as proof of concept. First, we propose a method that provides interpretability with respect to the results through uncertainty measures. The model in question is capable of reasoning about the uncertainty inherent in data and leverages that information to progressively refine its outputs. In particular, the method is applied to land cover segmentation, a classification task that aims to assign a type of land to each pixel in satellite images. The dataset and application serve to prove that the final uncertainty bound enables the end-user to reason about the possible errors in the segmentation result. Second, Recurrent Neural Networks are used as a method to create robust models towards lacking datasets, both in terms of size and class balance. We apply them to two different fields, road extraction in satellite images and Wireless Capsule Endoscopy (WCE). The former demonstrates that contextual information in the temporal axis of data can be used to create models that achieve

comparable results to state-of-the-art while being less complex. The latter, in turn, proves that contextual information for polyp detection can be crucial to obtain models that generalize better and obtain higher performance. Last, we propose two methods to leverage unlabeled data in the model creation process. Often datasets are easier to obtain than to label, which results in many wasted opportunities with traditional classification approaches. Our approaches based on self-supervised learning result in a novel contrastive loss that is capable of extracting meaningful information out of pseudo-labeled data. Applying both methods to WCE data proves that the extracted inherent knowledge creates models that perform better in extremely unbalanced datasets and with lack of data. To summarize, this thesis demonstrates potential solutions to obtain uncertainty bounds, provide reasonable explanations of the outputs, and to combat lack of data or unbalanced datasets. Overall, the presented methods have a positive impact on the DL field and could have a real and tangible effect for the society.

[cat] És innegable que el Deep Learning ha causat una revolució en molts aspectes no solament de l’aprenentatge automàtic però també de les nostres vides diàries. Tot i així, encara queden aspectes a millorar. Les xarxes neuronals tenen problemes per estimar la seva confiança en les prediccions, i sovint reporten probabilitats altes en casos que no tenen relació amb el model o que directament no tenen sentit. De la mateixa forma, interpretar els resultats d’un model profund i complex resulta una tasca extremadament complicada. Aquests mateixos models, cada cop amb més paràmetres i més potents, requereixen també de dades més ben etiquetades i més completes.

Tenint en compte aquestes limitacions, l’objectiu principal és el de buscar mètodes i algoritmes per trobar-ne solució. Primerament, es proposa la creació d’un mètode capaç d’obtenir incertesa en imatges satèl·lit i d’utilitzar-la per crear models més robustos i resultats interpretables. En segon lloc, s’utilitzen Recurrent Neural Networks (RNN) per combatre la falta de dades mitjançant l’obtenció d’informació contextual de dades temporals. Aquestes s’apliquen per l’extracció de carreteres d’imatges satèl·lit i per la classificació de pòlips en imatges obtingudes amb Wireless Capsule Endoscopy (WCE). Finalment, es plantegen dos mètodes per tractar amb la falta de dades etiquetades i desbalancejos en les classes amb l’ús de Self-supervised Learning (SSL). Seqüències no etiquetades d’imatges d’intestins s’incorporen en el models en una fase prèvia a la classificació tradicional.

Aquesta tesi demostra que les solucions proposades per obtenir mesures d’incertesa són efectives per donar explicacions raonables i interpretables sobre els resultats. Igualment, es prova que el context en dades de caràcter temporal, obtingut amb RNNs, serveix per obtenir models més simples que poden arribar a solucionar els problemes derivats de la falta de dades. Per últim, es mostra que SSL serveix per combatre de forma efectiva els problemes de generalització degut a dades no balancejades en diversos dominis de WCE.

Concloem que aquesta tesi presenta mètodes amb un impacte real en diversos aspectes de DL a la vegada que demostra la capacitat de tenir un impacte positiu en la societat.