Juan José investigates POMDP-based HVAC control framework that explicitly accounts for uncertainty in indoor environmental conditions. Combining POUCT online planning with particle-filter belief estimation, the approach balances thermal comfort and energy consumption. Co-simulation experiments show substantial energy savings and more stable operation compared with a greedy control strategy.

• J. Barriga, C. Burgos and M. Araya, Probabilistic Lookahead HVAC Control for Thermal Comfort. Energy and Buildings. 2026 

Miguel investigates Bayesian Reinforcement Learning for continual robotic learning, using Vision-Language-Action models to provide behavioural priors that guide exploration. It develops methods for updating these priors from experience while managing epistemic uncertainty, aiming to enable robots to learn new tasks efficiently without catastrophic forgetting or excessive interaction.

• M. Rojas and M. Araya. iCub-LeRobot: A VR Teleoperation Framework for Scalable VLA Dataset Generation. IEEE International Conference on Developmental Learning. 2026 

José proposes integrating nonverbal facial behavior directly into Large Language Models using Facial Action Units (AUs). It constructs word-level AU-annotated conversational data and compares in-context learning with vocabulary extension and fine-tuning, evaluating whether explicit AU representations improve the model’s ability to understand and generate contextually appropriate nonverbal behavior.

• J. Guillen, V. Barriere, and M. Araya. AU-TextAlign: A Pipeline for Word-Level Facial Expression Tagging in LLMs. ICPRS 2025

Proposal Pending...

Proposal Pending...

Proposal Pending...

Summary: This thesis investigates hyperbolic embeddings for image representation, exploiting their ability to encode implicit hierarchical structures. Applied to Matrioshka Embeddings and Deep Clustering on CIFAR10/100 and Imagenette, hyperbolic representations outperform Euclidean counterparts particularly in low-dimensional settings, although stability issues limit their effectiveness for Deep Clustering.

Contribution: It demonstrates that hyperbolic geometry can produce more effective compact image representations than conventional Euclidean embeddings, particularly when representation dimensionality is strongly constrained, while identifying stability as a key limitation for unsupervised deep clustering.

• I. Roberts, R. Ñanculef, M. Araya and M. Mallea. Recovering Latent Hierarchical Relationships in Image Datasets Through Hyperbolic Embeddings. 27th Iberoamerican Congress on Pattern Recognition (CIARP), 2024. 

• M. Solar, M. Araya, R. Ñanculef, L. Dombrovskaia, V. Castañeda, I. Roberts. ALPACS: Interoperable Repository of Medical Images. To appear in CLEI 2024.

Summary: This thesis develops a 3D convolutional neural network for reconstructing cosmic filaments from incomplete observations, using only dark-matter halo positions and masses. Trained on cosmological simulations, the approach reconstructs complex filamentary structures more accurately and substantially faster than the Minimum Spanning Tree baseline, despite the observational bias toward visible massive galaxies. 

Contribution: It demonstrates that deep learning can recover otherwise hidden large-scale cosmic structures from sparse, observationally biased information, outperforming the conventional MST approach in reconstruction quality and inference time. 

• G. Aichele, R. Smith and M. Araya. Towards Reconstructing the Filamentary Structure of an Only Partially Visible Cosmic Web with Deep Learning and 3D Cosmological Simulations. 11th International Conference on Soft Computing & Machine Intelligence (ISCMI), 2025

Summary: This thesis develops a Transformer-based approach for Music Emotion Recognition, estimating how valence and arousal evolve over time directly from audio features. Using the DEAM dataset, several Transformer configurations are evaluated and compared, and the resulting emotional trajectories are incorporated into a prototype emotion-aware music recommendation system. r

Contribution: It demonstrates the use of Transformers to model the temporal evolution of perceived emotion in music from audio alone, moving beyond static emotion classification toward continuous emotional trajectories that can support emotion-aware recommendation systems. 

Keywords: Music Emotion Recognition; affective computing; Transformers; emotional trajectory; valence; arousal; Music Information Retrieval; music recommendation systems; DEAM dataset; deep learning.

• P. A. Stoller and M. Araya. A Transformer Approach for Predicting Transitions in Emotional Trajectory Directly from Music Audio. Neural Computing and Applications. Volume 37, pages 16427–16444. 2025 

Summary: This thesis improves cross-modal retrieval by addressing inconsistent intra-modal similarities produced by triplet-loss training with hard-negative mining. It proposes two novel loss functions incorporating intra-modal constraints and evaluates them on MS-COCO and Flickr30K. Results demonstrate improvements in cross- and intra-modal retrieval, including unsupervised image retrieval without category labels.

Contribution: It identifies and theoretically and experimentally characterizes an overlooked limitation of cross-modal triplet learning, and introduces two novel loss functions that explicitly exploit intra-modal similarity constraints, enabling more accurate and novel retrieval while providing a new framework for evaluating unlabeled image-to-image retrieval.

• M Mallea, R Ñanculef, M Araya. Intramodal consistency in triplet-based cross-modal learning for image retrieval. Machine Learning 114 (4), 110. 2025

• I. Roberts, R. Ñanculef, M. Araya and M. Mallea. Recovering Latent Hierarchical Relationships in Image Datasets Through Hyperbolic Embeddings. 27th Iberoamerican Congress on Pattern Recognition (CIARP), 2024. 

• M Mallea, R Nanculef, M Araya. Enhancing Intra-modal Similarity in a Cross-Modal Triplet Loss. International Conference on Discovery Science, 249-264. 2023.

Summary: This thesis develops a semi-supervised deep-learning approach for gamma/hadron classification in the Cherenkov Telescope Array Observatory. Using convolutional neural networks and pseudo-labeling, it enables models trained on labeled simulations to continue learning from unlabeled events. Experiments demonstrate promising classification improvements and explore learned representations through comparisons with conventional Hillas parameters.

Contribution: It proposes a semi-supervised classification strategy that exploits unlabeled Cherenkov telescope observations, reducing dependence on labeled simulations while investigating the physical interpretability of CNN representations through their relationship with Hillas parameters.

• D. Riquelme, M. Araya, M, S. Borquez, B. Panes and E. Carquin. (2023). Deep Learning Semi-Supervised Strategy for Gamma/Hadron Classification of Imaging Atmospheric Cherenkov Telescope Events. In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods - ICPRAM, ISBN 978-989-758-626-2; ISSN 2184-4313, pages 725-732. 2023

Summary: This thesis proposes Deep Ideal Scene, a deep-learning approach to Multiple Object Tracking (MOT) that transforms video into a simplified representation where objects retain identity through time. Combining convolutional, recurrent, and generative adversarial networks, the method is evaluated on MOT16, Stanford Drone, and UA-DETRAC, achieving competitive tracking performance across challenging scenarios. 

Contribution: It introduces the Ideal Scene representation as an intermediate learned representation for MOT, simplifying object–identity association by encoding tracked objects with persistent identities. The resulting architecture demonstrates that this representation can provide competitive tracking accuracy while operating substantially faster than several contemporary deep-learning trackers.

• D. Avila and M. Araya, “Multiple object tracking through deep ideal scene and recurrent neural networks,” unpublished manuscript, 2022.

Summary: This thesis develops a deep-learning content-based medical image retrieval system designed for largely unlabeled clinical datasets. It combines representations learned by segmentation and classification networks to construct richer image embeddings. Experiments on COVID-19 and pulmonary-nodule images demonstrate effective retrieval and information transfer, while near-duplicate removal improves robustness.

Contribution: It proposes a medical image retrieval architecture that combines segmentation and classification embeddings, enabling content-based retrieval in realistic clinical scenarios with limited labeled data, while demonstrating the regularizing effect of near-duplicate removal.

• G. Molina, M. Mendoza, I. Loayza, C. Nuñez, M. Araya, V. Castañeda and M. Solar. A new content-based image retrieval system for SARS-CoV-2 computer-aided diagnosis.  In International Conference on Medical Imaging and Computer-Aided Diagnosis, MICAD 2021, 2021. 

• F. Mena, P. Olivares, M. Bugueño, G. Molina and M. Araya, On the Quality of Deep Representations for Kepler Light Curves Using Variational Auto-Encoders. Signals, 2, 706-728. MDPI 2021. 

• C.G. Sotomayor, M. Mendoza, V. Castañeda, H. Farías, G. Molina, G. Pereira, S. Härtel, M. Solar and M. Araya. Content-Based Medical Image Retrieval and Intelligent Interactive Visual Browser for Medical Education, Research and Care. In Diagnostics, Vol. 11(8):1470-, MDPI 2021. 

• M. Bugueño, G. Molina, F. Mena, and M. Araya. Harnessing the power of CNNs for Light curves using Markov Transition Fields. Astronomy and Computing, 35: 100461, Elsevier 2021. 

Summary: This thesis develops a scalable method for identifying and hierarchically decomposing structures in astronomical molecular clouds. It represents observational data using Gaussian mixtures and applies hierarchical clustering through Gaussian Mixture Reduction. Evaluations on ALMA and synthetic data show accurate recovery of overlapping, multi-scale cloud structures with substantially reduced computational cost.

Contribution: It introduces a computationally scalable hierarchical representation of molecular clouds, improving the method through divide-and-conquer decomposition, faster Gaussian-mixture algorithms, and parallelization, while preserving, and in some cases improving, structural reconstruction accuracy.

• M. Villanueva, C. Torres, M. Araya and P. Amigo. Hdmclouds: A hierarchical decomposition of molecular clouds based on gaussian mixtures. Monthly Notices of the Royal Astronomical Society, 482(3):2878–2892, OUP 2019. 

• M. Villanueva and M. Araya. Multiresolution and hierarchical analysis of astronomical spectroscopic cubes using 3d discrete wavelet transform. In 8th International Conference of Pattern Recognition Systems, Madrid, Spain, 2017.

• M. Villanueva, C. Valenzuela, M. Sánchez, and M. Araya. Wrapping and deploying legacy astronomical code into python environments: An applied case study. In Astronomical Data Analysis Software & Systems. ADASS, 2017.

• M. Araya, M. Osorio, C. Ponce, M. Villanueva, C. Valenzuela, and M. Solar. Jovial: Notebook-based astronomical data analysis in the cloud. Astronomy and Computing, 25:110–117, Elsevier 2018.

• M. Araya, M. Solar, D. Mardones, L. Arevalo, M. Mendoza, C. Valenzuela, T. Hochfärber, M. Villanueva, M. Jara, and A. Simonsen. The ChiVO library: advanced computational methods for astronomy. In Astronomical Data Analysis Software & Systems. ADASS, 2015.