"The Power of Demonstrations in Machine Learning"
Date: May 11, 2023
About lecturer: Pavel Sulimov, Ph.D. in Computer Science, Research Associate in Institute of Applied Information Technology (Zurich University of Applied Sciences)
Abstract: Classical supervised learning might suffer from the lack of labeled data available for training. In unsupervised learning, where no labels/targets are given, learning process can be quite expensive due to need of patterns understanding from scratch. As an attempt to find a "golden middle" between supervised and unsupervised learning, weak supervision (semi-supervision) has come around with combining a small amount of labeled data with a large amount of unlabeled data during training.
The other approach that tries to get rid of problems with data is a reinforcement learning, where it's suggested to collect the data along with training via running through the episodes of the task e.g. the game playing. However, not every problem could be formulated through the principles of reinforcement learning - and more to say, reinforcement learning itself occurred to be more powerful when introducing the demonstrations (aka weak supervision) - examples of how to "play the game correctly".
At the lecture we will touch the theoretical background of the weak supervision and demonstrations, study the recent cases from different fields (mathematics, bioinformatics, information sciences etc.), and discuss the "chicken-egg problem" consequences of semi-supervision.
A Visual Analytics System for Improving Attention-based Traffic Forecasting Models
Date: May 15, 2023
The speaker Seungmin Jin, 3rd year postgraduate student of the Department of Big Data and Information Retrieval of the Faculty of Computer Science.
Abstract: With deep learning (DL) outperforming conventional methods for different tasks, much effort has been devoted to utilizing DL in various domains. Researchers and developers in the traffic domain have also designed and improved DL models for forecasting tasks such as estimation of traffic speed and time of arrival. However, there exist many challenges in analyzing DL models due to the black-box property of DL models and complexity of traffic data (i.e., spatio-temporal dependencies). Collaborating with domain experts, we design a visual analytics system, AttnAnalyzer, that enables users to explore how DL models make predictions by allowing effective spatio-temporal dependency analysis. The system incorporates dynamic time warping (DTW) and Granger causality tests for computational spatio-temporal dependency analysis while providing map, table, line chart, and pixel views to assist users to perform dependency and model behavior analysis. For the evaluation, we present three case studies showing how AttnAnalyzer can effectively explore model behaviors and improve model performance in two different road networks. We also provide domain expert feedback.
“Utilizing empirical p-values in False Discovery Rate control and examination of the reasoning capacity of the deep net based METDR method”.
Date: May 26, 2023
Speaker: Borevskiy Andrey, Research assistant at the Laboratory on AI for computational biology
Abstract: Artificial Intelligence has been demonstrated as an incredibly useful instrument for a broad range of tasks. One of them — Bioinformatics — proposed fundamentally novel techniques at the intersection of machine learning and statistics. Despite their potential, these methods have not been utilized for more general tasks yet. Accordingly, in our work, we elaborate a drastically new approach called empirical p-values (EPV). Assuming negative training data of classification task to be the null hypothesis distribution, we calculate the corresponding p-values for test samples. Later, we expand the BH procedure to control FDR, making it possible both to regulate the interrelation of train and test data distributions, as well as to predict new labels based on those already investigated. The major goal is to accurately predict number of accepted discoveries at each level without true labels.
Speaker: Latypov Insan-Aleksandr, 2nd year Master's student on Data Science
Abstract: Visual Question Answering (VQA) is an important task in Artificial Intelligence which utilizes a combination of visual and textual data to answer questions. Recent advancements in deep learning, including transformer-based models, have allowed VQA systems to achieve results that are comparable to those of humans. However, questions remain regarding their ability to accurately reason and understand real-world concepts and relationships. Progress in the VQA task was helped by the implementation of large-scale datasets which often necessitated the use of complex visual and textual reasoning. A central challenge for deep learning models is the lack of reliable metrics when evaluating their reasoning abilities. To investigate the “Clever Hans” phenomenon, we have designed a new dataset which uses limited visual and textual concepts but requires complex reasoning skills and understanding. This study provides a methodology for building such datasets, along with a description of Blender and Python scripts and tasks. We also present the performance of pre-trained MDETR model on our dataset.
Have you spotted a typo?
Highlight it, click Ctrl+Enter and send us a message. Thank you for your help!
To be used only for spelling or punctuation mistakes.