In this research, statistical and deep learning models are applied to determine factors that affect motorcycle crash-injury severities. Four methodological challenges are considered: 1) imbalanced data (because fatal injuries are an exceedingly small portion of all resulting injury outcomes); 2) unobserved heterogeneity (because many unobserved factors will influence resulting injury severities); 3) quantification of variable effects; and 4) the possibility of temporally shifting relationships among variables. Convolutional neural networks and deep neural networks are the deep learning models considered, and random parameters logit models with heterogeneity in means and variances is the statistical model considered. Extensive experimentation indicated that data imbalance and unobserved heterogeneity could be best handled in deep learning models with a Bayesian deep neural network with a random generator and weighted loss function. With statistical modeling indicating significant shifts in model parameters over time, the data were segmented by year and both statistical and deep learning models were estimated. While techniques are available for deep learning to potentially handle data imbalance and unobserved heterogeneity, the quantification of variable effects and temporal shifts remains a challenge. For example, a comparison of variable effects show that the deep learning estimates of variable effects are generally inconsistent with the plausible values generated by the statistical models in terms of magnitudes and occasionally in terms of direction, indicating a need for improvements in deep-learning variable-effect extraction methods. The findings also show the need for future work to isolate the effect of complex temporal relationships which are currently imbedded in deep learning approaches, because the segmentation of data that has been used in statistical models to isolate temporal effects, and even the use of all data and defining new time-dependent variables, may not be a viable deep learning option due to the potential loss in predictive performance.

Causality and Representation Learning are foundational to advancing AI systems capable of reasoning, generalizing, and understanding the complex structure of the world. Causality provides tools to uncover the underlying causal structure of a system, understand cause-effect relationships, and reason about interventions. Representation Learning, on the other hand, transforms raw data into structured abstractions essential for modeling the underlying system and decision-making. Causal Representation Learning bridges these paradigms by using representation learning to extract high-level abstractions and entities and integrating causal reasoning principles to uncover cause-effect relationships between these entities. This approach is crucial for real-world systems, where causal relationships are typically defined between high-level entities, such as objects or interactions, rather than low-level sensory inputs like pixels. This thesis explores two key paradigms presented as a collection of two papers: the challenges in the evaluation of Bayesian Causal Discovery, and the effectiveness of structured representations, with a focus on object-centric representations in visual reasoning.

In the first paper, we study the challenges in the evaluation of Bayesian Causal Discovery methods. By analyzing existing metrics on linear additive noise models, we find that current metrics often fail to correlate with the true posterior in high-entropy settings, such as with limited data or non-identifiable causal models. We highlight the importance of considering posterior entropy and recommend evaluating Bayesian Causal Discovery methods on downstream tasks, such as causal effect estimation, for more meaningful evaluation in such scenarios.

In the second paper, we investigate the effectiveness of object-centric representations in visual reasoning tasks, such as Visual Question Answering. We reveal that while large foundation models often match or surpass object-centric models in performance, they require larger downstream models and more compute due to their less explicit representations. In contrast, object-centric models provide more interpretable representations but face challenges on more complex datasets. Combining object-centric representations with foundation models emerges as a promising solution, reducing computational costs while maintaining high performance. Additionally, we provide several additional insights such as segmentation performance versus downstream performance, and the effect of factors such as dataset size and question types, to further improve our understanding of these models.

Kausalitet och representationsinlärning är grundläggande för att utveckla AI-system som kan resonera, generalisera och förstå världens komplexa strukturer. Kausalitet tillhandahåller verktyg för att avslöja den underliggande kausala strukturen i ett system, förstå orsak-verkan-relationer och resonera kring interventioner. Representationsinlärning, å andra sidan, omvandlar rådata till strukturerade abstraktioner som är avgörande för modellering av det underliggande systemet och beslutsfattande. Kausal representationsinlärning sammanför dessa paradigm genom att använda representationsinlärning för att extrahera högre nivåers abstraktioner och enheter samt integrera principer för kausalt resonemang för att avslöja orsak-verkan-relationer mellan dessa entiteter. Detta tillvägagångssätt är avgörande för verkliga system, där kausala relationer vanligtvis definieras mellan högre nivåers entiteter, såsom objekt eller interaktioner, snarare än lågupplösta sensoriska data som pixlar. Denna avhandling undersöker två centrala paradigm presenterade som en samling av två artiklar: utmaningarna i utvärderingen av Bayesiansk kausal upptäckning och effektiviteten av strukturerade representationer, med fokus på objektcentrerade representationer inom visuellt resonemang.

I den första artikeln studerar vi utmaningarna i utvärderingen av metoder för Bayesiansk kausal upptäckning. Genom att analysera befintliga mått på linjära additiva brusmodeller finner vi att nuvarande metoder ofta misslyckas med att korrelera med den sanna posteriorn i högentropiska inställningar, såsom vid begränsad data eller icke-identifierbara kausala modeller. Vi framhäver vikten av att beakta posteriorns entropi och rekommenderar att Bayesiansk kausal upptäckning-metoder utvärderas på nedströmsuppgifter, såsom orsakseffektsberäkning, för att uppnå en mer meningsfull utvärdering i sådana scenarier.

I den andra artikeln undersöker vi effektiviteten av objektcentrerade representationer i visuella resonemangsuppgifter, såsom Visual Question Answering. Vi avslöjar att även om stora grundmodeller ofta kan matcha eller överträffa objektcentrerade-modeller i prestanda, kräver de större nedströmsmodeller och mer beräkningskraft på grund av deras mindre explicita representationer. I kontrast erbjuder objektcentrerade-modeller mer tolkningsbara representationer men möter utmaningar på mer komplexa datamängder. Att kombinera objektcentrerade-representationer med grundmodeller framstår som en lovande lösning, eftersom det minskar beräkningskostnaderna samtidigt som hög prestanda bibehålls. Dessutom presenterar vi flera ytterligare insikter, såsom sambandet mellan segmenteringsprestanda och nedströmsprestanda samt effekten av faktorer som datasetstorlek och frågetyper, för att ytterligare förbättra vår förståelse av dessa modeller.

Representing uncertainty in causal discovery is a crucial component for experimental design, and more broadly, for safe and reliable causal decision making. Bayesian Causal Discovery (BCD) offers a principled approach to encapsulating this uncertainty. Unlike non-Bayesian causal discovery, which relies on a single estimated causal graph and model parameters for assessment, evaluating BCD presents challenges due to the nature of its inferred quantity - the posterior distribution. As a result, the research community has proposed various metrics to assess the quality of the approximate posterior. However, there is, to date, no consensus on the most suitable metric(s) for evaluation. In this work, we reexamine this question by dissecting various metrics and understanding their limitations. Through extensive empirical evaluation, we find that many existing metrics fail to exhibit a strong correlation with the quality of approximation to the true posterior, especially in scenarios with low sample sizes where BCD is most desirable. We highlight the suitability (or lack thereof) of these metrics under two distinct factors: the identifiability of the underlying causal model and the quantity of available data. Both factors affect the entropy of the true posterior, indicating that the current metrics are less fitting in settings of higher entropy. Our findings underline the importance of a more nuanced evaluation of new methods by taking into account the nature of the true posterior, as well as guide and motivate the development of new evaluation procedures for this challenge.

Causal reasoning can be considered a cornerstone of intelligent systems. Having access to an underlying causal graph comes with the promise of cause–effect estimation and the identification of efficient and safe interventions. However, learning causal representations remains a major challenge, due to the complexity of many real-world systems. Previous works on causal representation learning have mostly focused on Variational Auto-Encoders (VAEs). These methods only provide representations from a point estimate, and they are less effective at handling high dimensions. To overcome these problems, we propose a Diffusion-based Causal Representation Learning (DCRL) framework which uses diffusion-based representations for causal discovery in the latent space. DCRL provides access to both single-dimensional and infinite-dimensional latent codes, which encode different levels of information. In a first proof of principle, we investigate the use of DCRL for causal representation learning in a weakly supervised setting. We further demonstrate experimentally that this approach performs comparably well in identifying the latent causal structure and causal variables.

Object-centric (OC) representations, which represent the state of a visual scene by modeling it as a composition of objects, have the potential to be used in various downstream tasks to achieve systematic compositional generalization and facilitate reasoning. However, these claims have not been thoroughly analyzed yet.Recently, foundation models have demonstrated unparalleled capabilities across diverse domains from language to computer vision, marking them as a potential cornerstone of future research for a multitude of computational tasks.In this paper, we conduct an extensive empirical study on representation learning for downstream Visual Question Answering (VQA), which requires an accurate compositional understanding of the scene. We thoroughly investigate the benefits and trade-offs of OC models and alternative approaches including large pre-trained foundation models on both synthetic and real-world data, and demonstrate a viable way to achieve the best of both worlds. The extensiveness of our study, encompassing over 600 downstream VQA models and 15 different types of upstream representations, also provides several additional insights that we believe will be of interest to the community at large.