|
|
Machine Learning Concepts for Categorization of Objects in Images
Hubený, Marek ; Honec, Peter (referee) ; Horák, Karel (advisor)
This work is focused on objects and scenes recognition using machine learning and computer vision tools. Before the solution of this problem has been studied basic phases of the machine learning concept and statistical models with accent on their division into discriminative and generative method. Further, the Bag-of-words method and its modification have been investigated and described. In the practical part of this work, the implementation of the Bag-of-words method with the SVM classifier was created in the Matlab environment and the model was tested on various sets of publicly available images.
|
|
|
The Application of Variational Autoencoders for Ancestral Sequence Reconstruction
Kohout, Pavel ; Martínek, Tomáš (referee) ; Musil, Miloš (advisor)
Protein engineering is an interdisciplinary science concerned with the design of improved proteins. A successful method used to design more stable and active proteins is ancestral sequence reconstruction. This method explores the evolutionary relationships between existing proteins and uses phylogenetic trees to generate their evolutionary ancestors, which often exhibit the desired improved properties. Therefore, new and more robust methods using mathematical models together with huge amounts of sequence data could become a powerful tool for protein engineering. This thesis explores the use of variational autoencoders as an alternative approach to ancestral sequence design compared to conventional methods using phylogenetic trees. Experiments were performed to optimize the architecture and statistical methods were proposed to evaluate the quality of the models and the sequences generated. At the same time, robustness tests of the whole method were performed and strategies for ancestral sequence generation were proposed and implemented.
|
|
|
A neural network for reconstruction of extinct animals
Pešek, David ; Bilík, Šimon (referee) ; Jirsík, Václav (advisor)
This work was focused on designing, learning and evaluating an artificial neural network for reconstructing extinct species. First, the main element of the proposed artificial neural network, i.e., the generative model, was selected. Given their excellent performance in the field of image generation, the class of diffusion models reasonably seemed to be the right choice. Specifically, the Stable diffusion model was chosen. One of the initial steps of the work was to create a training set for the proposed model. The animal images needed to be paired with some labels that could be used to identify the animal. For this purpose, the cytochrome c oxidase subunit I genes of the given animals were used. Furthermore, the sequential transformer model GPT-2, which is learned on the training set of human natural language, was used. This model was used to encode the DNA sequences into a vector form in which the semantics and context between the different parts of the DNA sequence were captured. The models would be very difficult to learn from scratch due to the large training set size required and the computational and time requirements. Thus, the GPT-2 model was only learned on the training set of DNA sequences of the passeriformes order, and the diffusion model itself was learned on pairs of images of these animals and DNA sequences encoded by the GPT-2 model. To generate the images, the original DNA sequences that resembled the sequences from the training set were generated using GPT-2. The encoding of these sequences was then passed to the diffusion model, which generated the images itself. The method of generating new DNA sequences using the GPT-2 model is based on the idea that the generated DNA sequence partially resembles the DNA sequences from the training set. Such experimentally generated DNA sequences may resemble DNA sequences of extinct ancestors or relatives of the passeriformes order. The model was in some cases able to generate images that could be considered as animal species , but it should be noted that often the generated images could not be considered as animal reconstructions. The success rate of generating a decent animal image was approximately 10%. The functionality of the model was also tested on a test set of DNA sequences of animals of several orders that fall under the class of birds as well as the order of passeriformes. The success rate of generating a reconstruction that could be compared to a photograph was around 5%.
|
|
|
The Application of Variational Autoencoders for Ancestral Sequence Reconstruction
Kohout, Pavel ; Martínek, Tomáš (referee) ; Musil, Miloš (advisor)
Protein engineering is an interdisciplinary science concerned with the design of improved proteins. A successful method used to design more stable and active proteins is ancestral sequence reconstruction. This method explores the evolutionary relationships between existing proteins and uses phylogenetic trees to generate their evolutionary ancestors, which often exhibit the desired improved properties. Therefore, new and more robust methods using mathematical models together with huge amounts of sequence data could become a powerful tool for protein engineering. This thesis explores the use of variational autoencoders as an alternative approach to ancestral sequence design compared to conventional methods using phylogenetic trees. Experiments were performed to optimize the architecture and statistical methods were proposed to evaluate the quality of the models and the sequences generated. At the same time, robustness tests of the whole method were performed and strategies for ancestral sequence generation were proposed and implemented.
|
|
|
Machine Learning Concepts for Categorization of Objects in Images
Hubený, Marek ; Honec, Peter (referee) ; Horák, Karel (advisor)
This work is focused on objects and scenes recognition using machine learning and computer vision tools. Before the solution of this problem has been studied basic phases of the machine learning concept and statistical models with accent on their division into discriminative and generative method. Further, the Bag-of-words method and its modification have been investigated and described. In the practical part of this work, the implementation of the Bag-of-words method with the SVM classifier was created in the Matlab environment and the model was tested on various sets of publicly available images.
|
guest ::
