Trabajos de fin de máster (TFM)

URI permanente para esta colección

hdl:123456789/37

Examinar

Examinando Trabajos de fin de máster (TFM) por Palabra clave "12 Matemáticas::1203 Ciencia de los ordenadores ::1203.04 Inteligencia artificial"

Mostrando 1 - 4 de 4
  • Miniatura
    Publicación
    Bayesian Inference on the Pleiades Open Cluster
    (Universidad de Educación a Distancia (UNED), 2023-09) Palmero Moya, Francisco José; Olivares Romero, Javier; Sarro Baro, Luis Manuel
    Context: Age is one of the fundamental parameters of any astrophysical object, being it a galaxy, a star, or a planet. The main three stellar dating techniques in astrophysics are dynamical ages, isochrone ages, and the chemical abundance of the Lithium isotopes. However, these methods produce ages that differ by up to 50 % for stars younger than the Sun. The main culprit for these differences is the lack of a consistent and robust age calibration due to the fact that the only sufficiently precise age available is that of the Sun. Aims: Our aim is to define and implement a Bayesian hierarchical model able to determine the ages of star clusters and associations through two age dating techniques: isochrones and Lithium abundance. Methods: The resulting model combines existent photometric, parallax, and chemical abundance of Lithium data sets of stars belonging to stellar open clusters to infer its age distribution through modern and robust artificial intelligence methods. A Neural Network is trained given a grid of pre-calculated BT-Settl models to interpolate the spectral energy distributions of stars, working as a black-box interpolator in the model. The Bayesian hierarchical model not only facilitates simultaneous inference of star-level parameters but also offers an elegant framework for effectively pooling open cluster information and propagating uncertainty. Markov Chain Monte Carlo techniques allow us to sample the posterior distribution using the Hamiltonian Monte Carlo algorithm. Results: Our model’s robust performance on a synthetic dataset with known parameters, coupled with its successful age estimation of the Pleiades Open Cluster (116.8 ± 1.9 Myr), represents a significant advancement in the field by overcoming key challenges that have hindered previous attempts mixing artifical intelligence paradigms. The resulting model signifies a new methodology for age estimation that can be applied to a wide range of open clusters, with the Pleiades serving as the initial test benchmark.
  • Miniatura
    Publicación
    MIDI-Conditional Text-to-Audio Synthesis Using ControlNet on AudioLDM
    (2023-09) Ibáñez Martínez, Laura; Cuadra Troncoso, José Manuel
    Text-to-audio systems have gained attention in recent months, achieving impressive results in general audio synthesis. However, they often lack fine-grained control over the musical output, as note-level adjustments cannot be determined by text. In this work, we present MIDI-AudioLDM, which implements MIDI conditioning into AudioLDM with the use of ControlNet. This enables MIDI-conditional text-to-audio synthesis, which adds up to AudioLDM’s previous capacities, including direct text-to-audio synthesis as well as audio style transfer and inpainting. Like AudioLDM, the model uses contrastive language-audio pretraining (CLAP) latents and is trained on audio embeddings, while using text embeddings for inference. In contrast to unconditional audio synthesis, MIDI-AudioLDM offers detailed control over various musical aspects such as notes, genre, mood, and timbre, which makes it a more valuable tool for the music production process. A demo is available at https://huggingface.co/spaces/lauraibnz/midi-audioldm.
  • Miniatura
    Publicación
    Optimización de Parámetros en Extreme Learning Machine mediante Algoritmos Evolutivos
    (Universidad de Educación a Distancia (UNED), 2023-06) Pinto Lozano, José Manuel; Carmona Suárez, Enrique J.
    Extreme Learning Machine es un paradigma relativamente reciente. Específicamente, se trata de una técnica que utiliza redes neuronales en las que los pesos de entrada son generados aleatoriamente y los pesos de salida son obtenidos mediante la matriz inversa generalizada de Moore-Penrose de la matriz de salida de la capa oculta. El tiempo de entrenamiento de este algoritmo es mucho menor que el de las redes neuronales típicas y su rendimiento no se ve afectado por ello, de ahí que los últimos años hayan sido el centro de atención de numerosas investigaciones. Algunos de estos trabajos se benefician de este bajo coste computacional en el entrenamiento para aplicar algoritmos de computación evolutiva que sintonicen adecuadamente los hiperparámetros de este tipo de redes con el n mejorar su potencia predictiva o de reducir su complejidad. Este trabajo se enmarca en esta última línea de investigación, teniendo como principal objetivo el uso de un algoritmo de computación evolutiva, denominado Covariance Matrix Adaptation Evolution Strategy, para caracterizar la función de activación más adecuada para cada conjunto de datos utilizado, con el n último de mejorar (aumentando la potencia predictiva o reduciendo la complejidad) el modelo obtenido mediante Extreme Learning Machine.
  • Miniatura
    Publicación
    A Time-Aware Approach to Detect Help-Seeking Behaviour from Student-Platform Interaction
    (Universidad de Educación a Distancia (UNED), 2023-09) Horta Bartomeu, Raquel; Santos, Olga C.; Arevalillo Herráez, Miguel
    Seeking help when needed is a crucial skill, especially in unsupervised learning scenarios. It is known that some students do not ask for assistance when they need it, which can lead to counterproductive learning sessions. This work addresses the challenge of detecting help-seeking behaviour by learning from student-platform interaction events using deep learning models. This is the first work to predict help-seeking by considering the temporal nature of student behaviour while being independent of the topic being taught and the task at hand. We depict student-platform interaction as sequences of actions and evaluate five distinct approaches alongside various data representation techniques. Our research yields a model for detecting help-seeking behaviour solely from action sequences. We hypothesise that this approach has the potential for further improvement, especially when combined with pedagogical data and personalised features. Furthermore, we introduce a novel knowledge representation technique for categorical sequence analysis.