PhD Research Seminar: Vector semiotic architecture in visual question answering, machine learning in sequence analysis, adaptive methods in stochastic optimization
When: May 4, 18:10–21:00
First talk: Vector Semiotic Architecture for Visual Question Answering
Speaker: Aleksei Kovalev, third-year PhD student, Faculty of Computer Science
The symbol grounding problem is one of the crucial obstacles on the path of constructing an intelligent agent able to act in the real nondeterministic environment. First, we clearly state this problem and show the potential power of the semiotic approach to solve it. Then we enrich it with vector symbolic architecture means that enable efficient computation. As a model problem, the visual question answering task is chosen.
Second talk: Efficiency and Interpretability of Machine Learning Methods for Sequence Analysis
Speaker: Anna Muratova, second-year PhD student, Faculty of Computer Science
In the first part, I will present the results of the analysis of demographic sequences obtained using various methods. In particular, the accuracy of the following methods was compared: decision trees, SVM with the use of special kernels and sequences converted to features, and recurrent neural networks.
In the second part, I will talk about the task of developing a neural network model in Python using the Tensorflow / Keras library. The development process includes experiments with various options for constructing a multi-channel convolutional neural network, searching for the most efficient architecture on the Movielens dataset. In particular, to improve recommendations, it is planned to use a channel for extracting semantics from text headers. In well-known SOTA papers, the semantics of heading text is not taken into account.