Ofertes de projectes

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).

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

This project will be carried out in collaboration with GusanoFilms
(https://www.gusano.org) audiovisual producers.  Its main goal is to
develop a tool, based on Machine Learning state-of-the-art methods,
for semi-automatic segmentation in multivariate temporal series of
images according to similarities among the elements in the image.

The project could be divided into different subtasks:

1. Identification of the state-of-the-art methodologies for the problem
2. Identification of suitable segmentation criteria to address the problem
3. Collection of a database adequate to these criteria
4. Implementation of a prototype

This project will be carried out in collaboration with GusanoFilms
(https://www.gusano.org) audiovisual producers. Its main goal is to
develop a tool, based on Machine Learning state-of-the-art methods,
for semi-automatic detection of relevant transitions in multivariate
temporal series of images.

The project could be divided into different subtasks:

1. Identification of the state-of-the-art methodologies for the problem
2. Identification of the suitable criteria to detect relevant transitions in multivariate temporal series of images
3. Collection of a database adequate to these criteria
4. Implementation of a prototype