Opening Ceremony
| [10:25 – 10:30] Opening the Workshop Gabriela Czibula (Babeș-Bolyai University), Sorin Burcea (Head of the Weather Radar Network Coordination Department, Romanian National Meteorological Administration) |
Applied Deep Learning
[10:30 – 10:55] Quantum Neural Network Design via Quantum Deep Reinforcement Learning
Anca Muscalagiu (Babeș-Bolyai University)
Quantum neural networks (QNNs) represent an important advancement at the intersection of quantum computing and artificial intelligence, offering notable improvements over classical models. However, designing optimal QNN architectures for specific problems is challenging due to the requirement for deep quantum mechanics knowledge and the time-consuming complexity of manual design. Addressing these issues, this paper introduces a novel approach to QNN design through quantum deep reinforcement learning. Our novel methodology extends beyond simple quantum circuits, applying quantum reinforcement learning to parameterised quantum circuits, fully integrating these into trainable QNNs. As far as we know, we introduce one of the pioneering methods that employ quantum reinforcement learning to autonomously generate optimal QNN architectures. These architectures undergo evaluation on specific datasets from various machine learning problems in order to assess their accuracy, thereby selecting the most effective ones on the task at hand, advancing towards more efficient and accessible quantum computing solutions.
[10:55 – 11:20] Prompt-based image editing for interior-design images
Victor Zarzu (Babeș-Bolyai University)
The topic of instruction-based image editing has got a lot of attention in recent years with a lot of research conducted due to its immense potential in various applications such as removing unwanted details present in existent images or improving them. However, one of the main problems in addressing this problem is acquiring a dataset for model training. Several methods and variations were proposed, but all of them rely on already existent data. We propose a method to address this problem by creating a context-specific dataset for interior-design with no previously available information by leveraging the knowledge of large language models (LLM). Furthermore, we test and prove the efficiency of the generated dataset on InstructPix2Pix which starts to compute better results for interior-design setting after the fine-tuning. Moreover, we propose a method for enhancing the localisation of the edit region through cross-attention map regularisation based on a text-based segmentation mask.
[11:20 – 11:45] Collaborative Task Scheduling Using Multi-Agent Deep Reinforcement Learning
Imre Mali (Babeș-Bolyai University)
Managing resources effectively, scalably, and affordably is a multi-faceted online decision-making challenge increasingly encountered in networking and cloud computing. Specifically, task scheduling is a complex challenge essential for the optimal functioning of today’s systems. Traditional heuristic approaches to scheduling are labor-intensive to design and particularly difficult to tune, leading to the proposal of various machine-learning-based methods. Reinforcement Learning (RL) showed great results in similar decision making problems, and many existing approaches employ RL to solve task scheduling problems. Most of these studies either focus on single-agent scenarios, which inherently suffer from scalability issues, or on highly specialised multi-agent applications. We propose a general-purpose multi-agent RL framework that can successfully learn collaborative optimal scheduling policies, making one step further towards clouds and networks that are both scalable and autonomous. Our experiments demonstrate that these agents can collaboratively learn optimal scheduling policies for dynamic workloads.
[11:45 – 12:10] Generating code from natural language using Abstract Syntax Graphs
Sîrbu Alexandru-Gabriel (Babeș-Bolyai University)
Code generation from natural language is the process in which a software system automatically generates executable code based on human-readable descriptions. Over the years, Sequence-to-Sequence models have been tasked with this problem and achieved notable results by generating, instead of code, an Abstract Syntax Tree (AST). ASTs are a representation of code in the form of a tree, which are built by parsers during source code translation and used to validate the semantics of a section of code. Although the generation of AST has improved the generation results as opposed to generating the code as simple plain text, traditional ASTs often suffer from redundancy, as entities and their corresponding values are stored separately, leading to inefficiencies in the generation process. Thus, we introduce the concept of Abstract Syntax Graph, which is derived from the AST, but merges the leaf nodes that represent either named entities, such as variables and functions, as well as their atomic values.
Deep Learning Applied in Meteorology
[12:30 – 13:00] Predicting weather radar data using an ensemble of neural network models
Andrei Mihai (Babeș-Bolyai University)
The topic of predicting weather radar echo is of major interest in both operational and research meteorology. Weather radar measurements are an important data source used by operational meteorologists for weather analysis, radar reflectivity having a significant influence on short-term heavy rainfall prediction. To manage the prediction of multiple timesteps ahead we proposed a novel approach: combining the results of multiple models that each predict at a different time step in the future – an ensemble model we named SepConv-ens. SepConv-ens uses three separable convolution-based deep learning models trained on real radar data from the Romanian National Meteorological Administration (NMA). Experiments reveal a good performance of the model in predicting radar data up to more than 40 minutes ahead and a good correlation between the radar measurements and the predictions in terms of spatial and intensity evolution of the radar echoes. SepConv-ens is integrated in the operational visualisation software utilised by the Romanian NMA and is the first attempt, at the national level, to offer an artificial intelligence-based automated assistance for operational meteorologists.
[13:00 – 13:30] ConvSNow: A tailored Conv-LSTM architecture for weather nowcasting based on satellite imagery
Eugen Mihuleț (Romanian National Meteorological Administration)
The objective of the paper is to improve upon current nowcasting methods by applying a Deep Learning model that uses Convolutional Long-Short Term Memory Networks on a combination of satellite data. It proposes the ConvS Now model for short-term prediction of satellite images that would be useful for precipitation nowcasting. The proposed model was trained and evaluated on satellite imagery collected by EUMESAT’s Meteosat-11 satellite utilising the Severe Storms RGB product. The experimental results performed a subset of the Meteosat-11 data spanning Europe demonstrate that this model can enhance weather short-term forecasting, reduce costs and time, and improve the general quality of predictions, as a normalised mean of absolute errors of 1.6% was attained, outperforming every other baseline approaches considered for comparison.
[13:30 – 14:00] System for Identification of Maize Ideotypes, optimal sowing dates and nitrogen fertilization under climate change – PREPCLIM
Mihaela Caian (Romanian National Meteorological Administration)
We present the implementation and validation of a new integrated climate / phenology / decision support (CPD) modeling system (based on CORDEX models / DSSAT model), a TLR4 system, developed under the frame of PREPCLIM project, and the results of system simulations carried out to identify the ideotypes for maize in the near future climate (2050) for SE Europe/ Romania. Ideotyping with the CPD system was carried out for an ensemble of genotyping codes (a genotyping code involved ~ 2000 simulations), for each testing alternate 12 treatments (4 planting dates and 3 fertilisation levels). Projected changes in ideotype were simulated for two scenarios: RCP45 and RCP85 against historical simulations using an ensemble of three CORDEX models. The results for this ensemble indicate, for the control genotype, a mean decrease in production in both scenarios for all sowing dates and fertilisation levels tested, changes that were analysed in relation to several factors.
[14:00 – 14:25] Semi-supervised classification using generative diffusion models
Paul-Dumitru Orășan (Babeș-Bolyai University)
Diffusion models have revolutionized the field of generative machine learning due to their effectiveness in capturing complex, multimodal data distributions. Semi-supervised learning represents a technique that allows us to extract information from a large corpus of unlabeled data, assuming that a small sample of labeled data is provided. Many generative methods have been previously adapted to semi-supervised learning tasks. In this work, we pioneer adapting state-of-the-art generative diffusion models to the problem of semi-supervised image classification. We propose a self-supervised, pseudo-labelling pipeline which uses a diffusion model to learn the conditional probability distribution of neighboring data points. Preliminary experiments reveal strong performance even when the model is exposed to a very small percentage of labeled data (1%), validating the extraction of information from the unlabeled data. We conclude by conducting a study on the application of diffusion models in the problem of rainfall nowcasting, a precipitation event classification task based on remote sensing data.