Thesis offers

Boltzmann Machines are probabilistic models developed in 1985 by
D.H. Ackley, G.E. Hinton and T.J. Sejnowski. In 2006, Restricted
Boltzmann Machines (RBMs) were used in the pre-training step of
several successful deep learning models, leading to a new renaissance
of neural networks and artificial intelligence.

In spite of their nice mathematical formulation, there are a number of
issues that are hard to compute:

1. The computation of the partition function is NP-hard, involving an  exponential sum of terms
2. The exact computation of the derivative of the log-likelihood is  also NP-hard, since it contains the derivative of the partition function

Therefore, in practice we have to approximate both the computation of
the probabilities and several components of the learning process
itself. These drawbacks have prevented RBMs to show their real
potential as truly probabilistic models.

Currently, we are working on trying to improve several of the unsolved
issues related to RBMs:

1. - Mechanisms to control the learning process: www.lsi.upc.edu/~eromero/Publications/Downloads/2018-tnnls-stopcritRBM.pdf
2. Better approximations of the derivative of the log-likelihood: http://www.lsi.upc.edu/~eromero/Publications/Downloads/2019-nn-weightedCD.pdf
3. Efficient approximation of the partition function (work in progress)

These works have opened new lines of research, some of which can be
the topic of a Master's Thesis. The scope and degree of depth of the
work can be adapted to the estimated times to complete the Thesis. For
further details, contact Enrique Romero (eromero@cs.upc.edu).

Accurate prediction of these conditions for a sailing race is invaluable for sailors as it heavily influences their tactics during the race. In order to develop a decision support system applicable to the Olympic Classes Sailing Venues, the following components are being developed and integrated into a single web-based platform: Wind component, Waves component, Oceanic current component, and Boat performance component.

The current master's thesis proposal is focused on the wind component.

The goal of the previous theses was to develop a methodology/procedure for analyzing wind datasets to automatically identify significant wind patterns. In other words, the aim was to recognize characteristic features of wind speed and direction related to other weather parameters of the day (such as air pressure, air and water temperature, cloud coverage, etc.) and the geographical position of the specific racing area. The analysis was conducted on both recorded wind data and wind data predicted by numerical prediction models. A highly promising correlation was found between the predicted and recorded data, which is believed to provide valuable insights for validating the weather model that will be used daily during the Olympic Games.

For instance, while traditional classification schemes for wind patterns are based on meteorological experience and manual analysis of synoptic weather charts, our current approach based on clustering can automatically induce wind patterns, including characteristic features, its evolution throughout the day, and specific characteristics related to the geographical position of the racing area. This allows for:

- A detailed analysis to determine the representativeness of the wind fields encountered during the measurement period, their frequency of occurrence, timing, rate of evolution, and transition probabilities.

- Consequently, a more accurate prediction of the expected conditions before a sailing race, which is highly valuable information for the sailors.

During the previous master projects, the clustering module was fully developed and applied to data from weather prediction models as well as data collected at sea. Traditional clustering techniques were employed for this purpose. Additionally, since the recorded data are collected daily and exhibit a time-series behavior, advanced and novel sequential clustering techniques were applied to extract more complex sequential wind patterns. Building on these results, the current project aims to expand and enhance the achievements thus far by:

1.  - Applying additional machine learning techniques to incorporate key parameters such as geolocation and weather parameters other than specific wind data.

2.  - Developing a prediction module that suggests the forecasted pattern and behavior of the wind for a racing day based on weather model data and, if available, actual data from the previous days and/or early on the current day.

3.  - Generalizing the environment so that it can be used not only in Marseille (the Olympic venue for Paris 2024) but also in any specific location where wind patterns are of interest and data is available.

The sea data is collected in the format provided by the Austrian Sailing Federation, which owns the on-board wind measurement system. The weather prediction models dataset follows the format used by operational meteorological centers, as suggested by the World Meteorological Organization, for the storage and exchange of gridded weather and climate model fields (grib2 or NetCDF).

Although this is a Modality A thesis, interaction and collaboration with TriM company are highly likely depending on the progress of the work.

This is precisely why companies such as Google have invested a lot of resources on exploring applications of GNN, some of which are as mediatic as Alpha Fold (recently solving the Protein-Folding problem), or used by Goolge Maps to make predictions of the expected traffic of a road. In this regard, the goal of this thesis is to continue investigating about the potential of GNNs when applied to the field of Computer Networks. Our goal is to provide GNN-based solutions that make computer networks run more autonomously and in a more efficient way.

 The goal of this thesis is to enable the application of Deep Reinforcement Learning techniques for optimization problems in very large and complex scenarios, in this case Data Center Networks. Inspired by the last related work from OpenAI (https://openai.com/blog/evolution-strategies/), we will explore time-efficient alternatives to train a Graph Neural Network based DRL agent on large optimization scenarios in a reasonable amount of time.

A more detailed description of the project can be found in

https://www.cs.upc.edu/~eromero/Downloads/Retina-TFM-Project-01.pdf

The project is proposed in collaboration with Javier Zarranz Ventura
(Institut Clínic d'Oftalmologia, ICOF, Hospital Clínic de Barcelona, and
Institut d'Investigacions Biomèdiques August Pi I Sunyer, IDIBAPS),
which would provide a large annotated database to develop the project. For further information, please contact Alfredo Vellido (avellido@cs.
upc.edu) or Enrique Romero (eromero@cs.upc.edu).

Recently, diffusion-based image generative models have obtained amazing results in multiple tasks such as text-conditioned image generation. They have been also successfully applied to image translation tasks, in which an input image is converted into a target domain defined by, e.g., the appearance of an example image (style transfer), text condition or in other ways. Image cartoonization is a particular case of image translation in which the goal is to apply to the input image a cartoon style. While full image or video cartoonization could be seen as a threat by animation professionals, the possibility to use cartoonized image or video backgrounds (without people) would be an innovation opportunity.

Gradient descend based RBM learning is a computationally expensive procedure that requires the computation of a probability normalization constant (called the Partition function) that involves the computation of an exponentially large number of terms. In order to bypass that, different approximations exist, the most celebrated one being the Contrastive Divergence (CD) algorithm [1].

In this project we want to explore an alternative RBM model based on archetype (ARC) learning, described in [2]. The student will have to implement both the standard CD and the ARC algorithms in CUDA, and compare their performance in common benchmarking datasets.

[1] "Training restricted Boltzmann machines: An introduction",
Asja Fischer and Christian Igel, Pattern Recognition 47 (2014) 25-39

[2] "The emergence of a concept in shallow neural networks"
Elena Agliari, Francesco Alemanno, Adriano Barra, Giordano De Marzo, Neural Networks 148 (2022) 232-253

Development of an automatic plan generation alternative to

  https://arxiv.org/abs/2004.13204v1

  http://staff.ustc.edu.cn/~fuxm/projects/DeepLayout/index.html

with Reinforcement Learning

TinyML is a growing community that does machine learning on microcontrollers https://www.tinyml.org. 

Microcontrollers are sometimes the only choice when the power supply is limited, e.g. in for battery- or solar-powered applications. Application examples of TinyML are "wildlife" observation, such as: https://mybirdbuddy.com/ https://opencollar.io/, but tiny machine learning is also in domestic, healthcare and industrial applications.

While Arduino Uno boards are well known for all kinds of hobbyist microcontroller projects, for machine learning the more powerful 32 bit microcontrollers and development boards are used, such as the Arduino Portenta H7 or Arduino Nano 33 BLE Sense. These are able to run machine learning applications. To get the first information on this topic have a look at TensorFlow Lite for Microcontrollers: https://www.tensorflow.org/lite/microcontrollers.

In this project we would like to explore on-device machine learning model training specifically doing it by federated learning. We would like to use the Arduino Portenta H7, the most powerful microcontroller board from the Arduino family. We have Arduino Portenta H7 that can be used for the project. 

We have done previous work and code available where we showed that it is feasible.

https://www.mdpi.com/2079-9292/11/4/573

In the project it would be interesting to explore the capacities of the Portant H7, like using the two cores of the board, use at least one the sensors (camera, accelaration, microphone,...), and use the network connectivity (Wifi, BLE, LoRa), try with Convolutionary Neural Networks, ...

One concrete case would be to use for the communication part of federated learning the Wifi connectivity of the Arduino Portenta H7 to build a "cluster" (inspired by clusters of Raspberry Pi) where all the nodes are interconnected and can participate in federated learning. 

The project can start with already working code (in C/C++, Python) that we have from previous work on TinyML with on-device training and federated learning. https://upcommons.upc.edu/bitstream/handle/2117/353756/160036.pdf

The application of Artificial Intelligence (AI) and Machine Learning (ML) to network security (AI4SEC) is paramount against cybercrime. While AI/ML is mainstream in domains such as computer vision and natural language processing, traditional AI/ML has produced below-par results in AI4SEC. Solutions do not properly generalize, are ineffective in real deployments, and are vulnerable to adversarial attacks. A fundamental limitation is the lack of AI/ML technology specific to network security. Due to their unique ability to learn and generalize over graph-structured information, graph-learning approaches, and in particular Graph Neural Networks (GNNs), have recently enabled groundbreaking applications in multiple fields where data are generally represented as graphs. Network security data are intrinsically relational, and initial research suggests that graph-structured representations and GNNs have the potential to become foundational to AI4SEC, in the way convolutional and recursive networks were to computer vision and natural language processing.
The goal of GRAPHS4SEC is to leverage graph data representations and modern GNN technology to conceive a new breed of robust GNN-based network security methods which could radically advance the AI4SEC practice. The objectives of GRAPHS4SEC are: (a) to investigate algorithmic methods that facilitate modeling and learning from graph-based network security data; (b) to compare the benefits and overheads of GNN-based AI4SEC to traditional AI/ML in terms of detection performance, generalization, scalability, and robustness against adversarial attacks; (c) to showcase the benefits and improvements of GRAPHS4SEC technology in four critical, real-world network security applications with significant impact for society, considering (in particular) the detection and early mitigation of phishing and fake/malicious websites, a threat among the most popular and society-wide harmful in today's Internet.