Abstract
Language-guided generative models are increasingly used in audiovisual production. Image diffusion allows for the development of video sequences and some of its coordination can be established by text prompts. This research automates a video production pipeline leveraging CLIP-guidance with longform text inputs and a separate text-to-speech system. We introduce a method for producing frame-accurate video and audio summaries using a virtual timeline and document a set of video outputs with diverging parameters. Our approach was applied in the production of the film Irreplaceable Biography and contributes to a future where multimodal generative architectures are set as underlying mechanisms to establish visual sequences in time. We contribute to a practice where language modelling is part of a shared and learned representation which can support professional video production, specifically used as a vehicle throughout the composition process as potential videography in physical space.

Abstract
Distributed Machine Learning, in which data and learning tasks are scattered across a cluster of computers, is one of the answers of the field to the challenges posed by Big Data. Still, in an era in which data abounds, decisions must still be made regarding which specific data to use on the training of the model, either because the amount of available data is simply too large, or because the training time or complexity of the model must be kept low. Typical approaches include, for example, selection based on data freshness. However, old data are not necessarily outdated and might still contain relevant patterns. Likewise, relying only on recent data may significantly decrease data diversity and representativity, and decrease model quality. The goal of this paper is to compare different heuristics for selecting data in a distributed Machine Learning scenario. Specifically, we ascertain whether selecting data based on their characteristics (meta-features), and optimizing for maximum diversity, improves model quality while, eventually, allowing to reduce model complexity. This will allow to develop more informed data selection strategies in distributed settings, in which the criteria are not only the location of the data or the state of each node in the cluster, but also include intrinsic and relevant characteristics of the data.

Abstract
Fires in large buildings can have tragic consequences, including the loss of human lives. Despite the advancements in building construction and fire safety technologies, the unpredictable nature of fires, particularly in large buildings, remains an enormous challenge. Acknowledging the paramount importance of prioritising human safety, the academic community has been focusing consistently on enhancing the efficiency of building evacuation. While previous studies have integrated evacuation simulation models, aiding in aspects such as the design of evacuation routes and emergency signalling, modelling human behaviour during a fire emergency remains challenging due to cognitive complexities. Moreover, behavioural differences from country to country add another layer of complexity, hindering the creation of a universal behaviour model. Instead of centring on modelling the occupant behaviour, this paper proposes an innovative approach aimed at enhancing the occupants' behaviour predictability by providing real-time information to the occupants regarding the most suitable evacuation routes. The proposed models use a building's environmental conditions to generate contextual information, aiding in developing solutions to make the occupants' behaviour more predictable by providing them with real-time information on the most appropriate and efficient evacuation routes at each moment, guiding the occupants to safety during a fire emergency. The models were incorporated into a context-aware recommender system for testing purposes. The simulation results indicate that such a system, coupled with hazard and congestion models, positively influences the occupants' behaviour, fostering faster adaptation to the environmental conditions and ultimately enhancing the efficiency of building evacuations.

Abstract

Abstract

Abstract
Power transformers are one of the main elements of a power grid, and their downtime impacts the entire network. Repairing their failures can be very costly, so sophisticated maintenance techniques are necessary. To attempt to solve this problem, we developed a mixed-integer nonlinear optimization model that, focusing on a single power transformer, both schedules this maintenance and also decides how much of the hourly demand it will satisfy. A high level of load on a power transformer increases its temperature, which increases its degradation, and so these two decisions have to be carefully balanced. We also consider that power transformers have several components that degrade differently. Our model becomes very difficult to solve even in reasonably sized instances, so we also present an iterative refinement heuristic.