"I need help getting into shape. Can I use AI as a personal trainer?" The most popular AI tools, like OpenAI's ChatGPT and Anthropic's Claude, can't actually help you trim down that waistline or make your pecs look plump. Sure, Claude can draft a decent nutrition plan, and ChatGPT can tell you which dumbbell exercises to start with, but basic Google searches could already do that a decade ago. Yet here we both are, in worse shape than we'd like.

The emergence of Deep Learning has marked a profound shift in machine learning, driven by numerous breakthroughs achieved in recent years. However, as Deep Learning becomes increasingly present in everyday tools and applications, there is a growing need to address unresolved challenges related to its efficiency and sustainability. This dissertation delves into the role of inductive biases -- particularly, continuous modeling and symmetry preservation -- as strategies to enhance the efficiency of Deep Learning. It is structured in two main parts. The first part investigates continuous modeling as a tool to improve the efficiency of Deep Learning algorithms. Continuous modeling involves the idea of parameterizing neural operations in a continuous space. The research presented here demonstrates substantial benefits for the (i) computational efficiency -- in time and memory, (ii) the parameter efficiency, and (iii) design efficiency -- the complexity of designing neural architectures for new datasets and tasks. The second focuses on the role of symmetry preservation on Deep Learning efficiency. Symmetry preservation involves designing neural operations that align with the inherent symmetries of data. The research presented in this part highlights significant gains both in data and parameter efficiency through the use of symmetry preservation. However, it also acknowledges a resulting trade-off of increased computational costs. The dissertation concludes with a critical evaluation of these findings, openly discussing their limitations and proposing strategies to address them, informed by literature and the author insights. It ends by identifying promising future research avenues in the exploration of inductive biases for efficiency, and their wider implications for Deep Learning.

Vision-Language adaptation (VL adaptation) transforms Large Language Models (LLMs) into Large Vision-Language Models (LVLMs) for multimodal tasks, but this process often compromises the inherent safety capabilities embedded in the original LLMs. Despite potential harmfulness due to weakened safety measures, in-depth analysis on the effects of VL adaptation on safety remains under-explored. This study examines how VL adaptation influences safety and evaluates the impact of safety fine-tuning methods. Our analysis reveals that safety degradation occurs during VL adaptation, even when the training data is safe. While safety tuning techniques like supervised fine-tuning with safety datasets or reinforcement learning from human feedback mitigate some risks, they still lead to safety degradation and a reduction in helpfulness due to over-rejection issues. Further analysis of internal model weights suggests that VL adaptation may impact certain safety-related layers, potentially lowering overall safety levels. Additionally, our findings demonstrate that the objectives of VL adaptation and safety tuning are divergent, which often results in their simultaneous application being suboptimal. To address this, we suggest the weight merging approach as an optimal solution effectively reducing safety degradation while maintaining helpfulness. These insights help guide the development of more reliable and secure LVLMs for real-world applications.

The systemic risks posed by general-purpose AI models are a growing concern, yet the effectiveness of mitigations remains underexplored. Previous research has proposed frameworks for risk mitigation, but has left gaps in our understanding of the perceived effectiveness of measures for mitigating systemic risks. Our study addresses this gap by evaluating how experts perceive different mitigations that aim to reduce the systemic risks of general-purpose AI models. We surveyed 76 experts whose expertise spans AI safety; critical infrastructure; democratic processes; chemical, biological, radiological, and nuclear risks (CBRN); and discrimination and bias. Among 27 mitigations identified through a literature review, we find that a broad range of risk mitigation measures are perceived as effective in reducing various systemic risks and technically feasible by domain experts. In particular, three mitigation measures stand out: safety incident reports and security information sharing, third-party pre-deployment model audits, and pre-deployment risk assessments. These measures show both the highest expert agreement ratings (>60\%) across all four risk areas and are most frequently selected in experts' preferred combinations of measures (>40\%). The surveyed experts highlighted that external scrutiny, proactive evaluation and transparency are key principles for effective mitigation of systemic risks. We provide policy recommendations for implementing the most promising measures, incorporating the qualitative contributions from experts. These insights should inform regulatory frameworks and industry practices for mitigating the systemic risks associated with general-purpose AI.

Environmental and individualistic variables affect the rate of human decomposition in complex ways. These effects complicate the estimation of the postmortem interval (PMI) based on observed decomposition characteristics. In this work, we develop a generative probabilistic model for decomposing human remains based on PMI and a wide range of environmental and individualistic variables. This model explicitly represents the effect of each variable, including PMI, on the appearance of each decomposition characteristic, allowing for direct interpretation of model effects and enabling the use of the model for PMI inference and optimal experimental design. In addition, the probabilistic nature of the model allows for the integration of expert knowledge in the form of prior distributions. We fit this model to a diverse set of 2,529 cases from the GeoFOR dataset. We demonstrate that the model accurately predicts 24 decomposition characteristics with an ROC AUC score of 0.85. Using Bayesian inference techniques, we invert the decomposition model to predict PMI as a function of the observed decomposition characteristics and environmental and individualistic variables, producing an R-squared measure of 71%. Finally, we demonstrate how to use the fitted model to design future experiments that maximize the expected amount of new information about the mechanisms of decomposition using the Expected Information Gain formalism.

Gender bias in artificial intelligence has become an important issue, particularly in the context of language models used in communication-oriented applications. This study examines the extent to which Large Language Models (LLMs) exhibit gender bias in pronoun selection in occupational contexts. The analysis evaluates the models GPT-4, GPT-4o, PaLM 2 Text Bison and Gemini 1.0 Pro using a self-generated dataset. The jobs considered include a range of occupations, from those with a significant male presence to those with a notable female concentration, as well as jobs with a relatively equal gender distribution. Three different sentence processing methods were used to assess potential gender bias: masked tokens, unmasked sentences, and sentence completion. In addition, the LLMs suggested names of individuals in specific occupations, which were then examined for gender distribution. The results show a positive correlation between the models' pronoun choices and the gender distribution present in U.S. labor force data. Female pronouns were more often associated with female-dominated occupations, while male pronouns were more often associated with male-dominated occupations. Sentence completion showed the strongest correlation with actual gender distribution, while name generation resulted in a more balanced 'politically correct' gender distribution, albeit with notable variations in predominantly male or female occupations. Overall, the prompting method had a greater impact on gender distribution than the model selection itself, highlighting the complexity of addressing gender bias in LLMs. The findings highlight the importance of prompting in gender mapping.

Trusted Execution Environments (TEE) are used to safeguard on-device models. However, directly employing TEEs to secure the entire DNN model is challenging due to the limited computational speed. Utilizing GPU can accelerate DNN's computation speed but commercial widely-available GPUs usually lack security protection. To this end, scholars introduce TSDP, a method that protects privacy-sensitive weights within TEEs and offloads insensitive weights to GPUs. Nevertheless, current methods do not consider the presence of a knowledgeable adversary who can access abundant publicly available pre-trained models and datasets. This paper investigates the security of existing methods against such a knowledgeable adversary and reveals their inability to fulfill their security promises. Consequently, we introduce a novel partition before training strategy, which effectively separates privacy-sensitive weights from other components of the model. Our evaluation demonstrates that our approach can offer full model protection with a computational cost reduced by a factor of 10. In addition to traditional CNN models, we also demonstrate the scalability to large language models. Our approach can compress the private functionalities of the large language model to lightweight slices and achieve the same level of protection as the shielding-whole-model baseline.

With the continuous development of large language models (LLMs), transformer-based models have made groundbreaking advances in numerous natural language processing (NLP) tasks, leading to the emergence of a series of agents that use LLMs as their control hub. While LLMs have achieved success in various tasks, they face numerous security and privacy threats, which become even more severe in the agent scenarios. To enhance the reliability of LLM-based applications, a range of research has emerged to assess and mitigate these risks from different perspectives. To help researchers gain a comprehensive understanding of various risks, this survey collects and analyzes the different threats faced by these agents. To address the challenges posed by previous taxonomies in handling cross-module and cross-stage threats, we propose a novel taxonomy framework based on the sources and impacts. Additionally, we identify six key features of LLM-based agents, based on which we summarize the current research progress and analyze their limitations. Subsequently, we select four representative agents as case studies to analyze the risks they may face in practical use. Finally, based on the aforementioned analyses, we propose future research directions from the perspectives of data, methodology, and policy, respectively.

Understanding the text in legal documents can be challenging due to their complex structure and the inclusion of domain-specific jargon. Laws and regulations are often crafted in such a manner that engagement with them requires formal training, potentially leading to vastly different interpretations of the same texts. Linguistic complexity is an important contributor to the difficulties experienced by readers. Simplifying texts could enhance comprehension across a broader audience, not just among trained professionals. Various metrics have been developed to measure document readability. Therefore, we adopted a systematic review approach to examine the linguistic and readability metrics currently employed for legal and regulatory texts. A total of 3566 initial papers were screened, with 34 relevant studies found and further assessed. Our primary objective was to identify which current metrics were applied for evaluating readability within the legal field. Sixteen different metrics were identified, with the Flesch-Kincaid Grade Level being the most frequently used method. The majority of studies (73.5%) were found in the domain of "informed consent forms". From the analysis, it is clear that not all legal domains are well represented in terms of readability metrics and that there is a further need to develop more consensus on which metrics should be applied for legal documents.

As automated decision-making solutions are increasingly applied to all aspects of everyday life, capabilities to generate meaningful explanations for a variety of stakeholders (i.e., decision-makers, recipients of decisions, auditors, regulators...) become crucial. In this paper, we present a taxonomy of explanations that was developed as part of a holistic 'Explainability-by-Design' approach for the purposes of the project PLEAD. The taxonomy was built with a view to produce explanations for a wide range of requirements stemming from a variety of regulatory frameworks or policies set at the organizational level either to translate high-level compliance requirements or to meet business needs. The taxonomy comprises nine dimensions. It is used as a stand-alone classifier of explanations conceived as detective controls, in order to aid supportive automated compliance strategies. A machinereadable format of the taxonomy is provided in the form of a light ontology and the benefits of starting the Explainability-by-Design journey with such a taxonomy are demonstrated through a series of examples.