PhD Research Seminar: DNA secondary structures and the attention mechanism, social network's critical phases and stochastic repost activity

First talk: Recognizing DNA Secondary Structures Using the Attention Mechanism  
Speaker: Seungmin Jin, second-year PhD student, Faculty of Computer Science

Although deep learning models show states-of-the-art performance in many domains, there is still a strong issue to accept this model in several domains where reliability is critical, such as medical, finance, and law. This is mainly because currently proposed models are black-box and it is difficult to understand or interpret their inference. Furthermore, model developers may not point out in which case and why the model could fail even.

To solve this issue, many proposals have been submitted since 2016, especially related to the DARPA project, USA, and this domain is now called Explainable Artificial Intelligence (XAI). The current solutions of  this problem can be classified into three categories: deep explanation, interpretable models, and model induction. Deep explanation approaches disassemble complex neural networks and build explanation methods for it. Interpretable models use statistical models, such as Bayesian, to avoid issues from complex neural networks. Lastly, model induction creates comparable models to collect evidence of inferences and make inductive logics to interpret the black-box deep-learning models.

In the NLP domain, attention mechanism is introduced, which is between deep explanation and interpretable models approach. The key idea of this approach is using a similarity function to let models decide which input to focus on. This model still uses complex neural networks; however, the focusing value is probability which may help humans understand the inference. We believe that text data processing (NLP) and DNA data processing (Bioinformatics) have a lot in common in terms of languages, so we decided to use this model to build interpretable attention-based deep learning for our topic.

Our main topic is recognizing DNA secondary structures using the attention mechanism. DNA secondary structure is the set of interactions between bases in which parts of strands are bound to each other. In DNA double helix, the two strands of DNA are held together by hydrogen bonds. The nucleotides on one strand base pairs with the nucleotide on the other strand. Understanding this structure is a crucial part to understanding the functions of DNA. However, the process from DNA strands to DNA secondary structures is complex; therefore, computational help is required to understand. Along with our main topic, we also apply the model in the different domains, especially for spatio-temporal space, such as traffic and finance, to see how generally the model is interpretable.

Second talk: Identification and Early Detection of Critical Phases and States of a Microblogging Social Network Based on the Analysis of Its Stochastic Repost Activity
Speaker: Viktor Dmitriev, second-year PhD student, Department of Business Informatics

Critical phenomena in microblogging social networks (MBS), which occur as a result of phase transitions, are characterized by an avalanche-like spread of microposts inside it. Most studies of stochastic avalanche dynamics are at the level of catastrophic changes in the networks’ structure as a result of the insignificant external impact corresponding to the critical value of the control parameter. As a rule, such studies are carried out using the formalism of statistical physics and are quite effective in the presence of detailed information about the interaction of network users. Despite this, the main drawback limiting the possibilities of this approach is the lack, and, in some cases, the complete absence of the necessary detailed information.

The presentation is devoted to the study of the self-organization of MBS into the critical state at the level of stochastic repost activities’ analysis. The proposed approach does not require detailed information on the interaction of users, and, in spite of this, it gives reliable results on the identification of critical phases and states of the MBS using the criticality spectrum. In addition, the presentation will describe results on the critical states’ early detection for model systems that generate avalanche dynamics.