بازنمایی متن در فضای برداری با حفظ ویژگی‌های معنایی متن یکی از مهم‌ترین بخش‌ها در پردازش زبان طبیعی می‌باشد. با انقلاب مدل‌های ترنسفورمر در حوزه پردازش متن مدل‌های گسترده‌ای جهت بازنمایی متن ارائه شده‌اند. همچنین به موازات آن مدل‌های پیشرفته‌ای برای تولید متن ارائه شده است که توانسته‌اند نقش به‌سزایی در بالابردن کیفیت سیستم‌های مبتنی بر هوش مصنوعی ایفا نمایند. از جمله این موارد می‌توان به مدل ChatGPT از OpenAI و یا مدل BARD از گوگل اشاره نمود. در این سخنرانی مروری بر تاریخچه کاربرد مدل‌های مبتنی بر ترنسفور در بازنمایی و تولید متن خواهیم داشت.

According to the rapid developments of machine learning techniques, it is essential to ensure the security and robustness of the deployed algorithms. Recent studies have showed that many classification methods, especially deep neural networks (DNNs), are vulnerable to adversarial examples designed to deceive a classification models in calculating the correct output with a slight change in input. Hence, adversarial attack and defense techniques have attracted increasing attention and have become a hot research topic in recent years. In this talk, we introduce this topic and discuss about applications, methods and challenges.

Lots of learning tasks require dealing with graph data which contains rich relation information among elements. Graph neural networks (GNNs), as a branch of deep learning in non-Euclidean space, perform particularly well in various tasks that process graph structure data. GNNs are connectionist models that capture the dependence of graphs via message passing between the nodes of graphs. Unlike standard neural networks, graph neural networks retain a state that can represent information from its neighborhood with an arbitrary depth. In this talk, we first review the state-of-the-art GNN models and their basic principles. Based on the analysis, we provide an outlook on the latest studies and point out their developing prospect.

With the increasing efficiency and applications of deep learning in recent years, the desire to create deeper networks with fewer parameters and increase the speed of basic operations has become a fundamental requirement. One of the important approaches to this end is the use of tensor-based methods. In this lecture, in addition to introducing several architectures based on tensor decompositions in deep learning, we examine the alpha-tensor method, which has recently developed a fast method for performing basic matrix calculations based on the Tensor structure using the reinforcement learning method.

In conventional learning techniques, a large labeled training data was required, which prevented progress in various applications for many low resource languages. In recent years, a considerable progress has been made in self-supervised learning to solve this problem, which is the case in the speech processing fields. In this presentation, several self-supervised learning models in the speech processing will be investigated and their applications in the speech and speaker recognition will be explained.

Image and video generation models have experienced significant growth in the past few years. These models have the ability to produce images/videos based on an input noise, a text, another image/video, or any other input. Various techniques and models have been proposed for this purpose, which have greatly improved the visual quality and semantic coherency of the outputs. In this presentation, different generative models such as Variational Auto-Encoders (VAEs), Generative Adversarial Networks (GANs), Normalizing Flows, and Diffusion Models are presented and the most recent image and video generation models are analyzed.