Practical Applications of Deep Learning (Sommersemester 2023)

Dozent: Dr. Haojin Yang (Internet-Technologien und -Systeme) , Jona Otholt (Internet-Technologien und -Systeme) , Gregor Nickel , Ting Hu (Internet-Technologien und -Systeme)

Allgemeine Information

• Semesterwochenstunden: 4
• ECTS: 6
• Benotet: Ja
• Einschreibefrist: 01.04.2023 - 07.05.2023
• Prüfungszeitpunkt §9 (4) BAMA-O: 24.07.2023
• Lehrform: Seminar / Übung
• Belegungsart: Wahlpflichtmodul
• Lehrsprache: Englisch
• Maximale Teilnehmerzahl: 15

Studiengänge, Modulgruppen & Module

Beschreibung

Artificial intelligence (AI) is the intelligence exhibited by computer. This term is applied when a machine mimics "cognitive" functions that humans associate with other human minds, such as "learning" and "problem solving". In the past five years, there have been significant advances in the field of AI. Some of the notable breakthroughs include the development of deep learning algorithms that have enabled AI systems to surpass human performance in a range of tasks, including image and speech recognition, natural language processing, and game playing. There has also been a rapid growth in the availability of data and computing power, which has fueled progress in AI research.

Looking ahead to the next decade, the prospects for AI are even more exciting. Many experts predict that AI will continue to transform a wide range of industries, from healthcare and transportation to finance and education. However, there are also concerns about the potential impact of AI on employment, privacy, and security, and it will be important to address these issues as AI becomes more integrated into our lives. For young university students, the developments in AI mean that there will be many exciting career opportunities in the field, both in research and in industry. AI will likely play an increasingly important role in shaping our world, and those who understand its potential and limitations will be well positioned to make a difference. At the same time, it will be important for students to consider the ethical implications of AI and to develop a nuanced understanding of how it can be used to benefit society.

In this seminar, the project topics we provide are closely related to our current research work. We will help students understand the topics and learn relevant knowledge during the course. Typically, after attending the seminar, our students become familiar in using deep learning frameworks and have a good understanding of the related research topics.

Course language: English and German

Topics in this seminar:

• Build your games using generative AI tools:
  The development of generative Artificial Intelligence (AI) models has had a significant impact on many creativity-related industries. These AI tools have enabled the generation of numerous ready-to-use materials in related areas. Additionally, the emergence of AI copilot has accelerated the process of developing and troubleshooting, allowing individuals without prior knowledge to build their applications more easily. In this topic, we will work on building our own games from scratch using generative AI models. First, a simple text adventure game using chatbots like ChatGPT will be created, with the addition of images generated by tools like Midjourney to enhance the game's vividness. Second, we will build more complex escape games, featuring 360-degree worlds and objects from tools such as Blockade labs. We will study different prompt engineering techniques and learn more about game development in this topic.
• Weather data compression with neural networks: Numerical weather data is one of the largest scientific datasets currently available to train neural networks. For example ERA5 is a grid based dataset, which provides a large number of meteorological variables, such as temperature, humidity, wind, pressure, and precipitation with a spatial resolution of 0.25 degrees (about 30 km) and temporal resolution of 1 hour. The data is covering a period of 1979 to the present day and provides high-resolution view on earth atmosphere. For this reason it is among others widely used for weather forecasting.The sheer amount of data creates a lot of difficulties in terms of storage and processing as several petabytes can neither be stored on most hard drives nor in GPU memory. Data compression can be a possible solution to this challenges. Especially AI models for data compression have shown an impressive storage reduction of several orders of magnitude. Thus, the data is no longer stored as numerical data but in an AI model with lossy compression. In this seminar topic, we will apply a new compression technique for multidimensional weather data presented on ICLR 2023. We will initially follow the implementation and apply it to our data to generate a baseline. However, one key weakness of the existing work is its tendency to smooth out extremes, such as hurricanes. As achieving an accurate forecast of these extreme weather events is vitally important, we will try to adapt the method to address this weakness. In the end, we will be able to use the generated data for an AI weather model.
• Class Discovery using Language-Image Pretrainings: Language-Image pretrainings such as CLIP have achieved strong results in zero-shot image classification, often competitive with supervised models. However, the zero-shot setting still requires knowledge of the classes that are present in the data, which is not always available. A more realistic task for this setting is class discovery, which aims to both discover classes in an unlabeled dataset, and assign the samples to these classes.Existing methods approach this task either fully unsupervised (e.g. image clustering), or by using a set of labeled images as a reference (e.g. novel class discovery). Given their strong performance on other tasks, it seems promising to apply language-image pretrainings to class discovery as well, but it is not obvious how to best use them in this setting.In this seminar topic, we will try to find a way to solve the class discovery task using language-image pretrainings. Since pretrained models of CLIP and similar methods are freely available, our focus will be on how to leverage them for the class discovery task, for example by developing a way to automatically generate fitting text prompts for each class.

Voraussetzungen

• Strong interests in video/image processing, machine learning (Deep Learning) and/or computer vision
• Software development in C/C++ or Python
• Experience with OpenCV and machine learning applications as a plus

Literatur

Books

• Dive into deep learning
• Ian J. Goodfellow and Yoshua Bengio and Aaron Courville, "Deep Learning", online version

 

Deep Learning frameworks:

• PyTorch

Leistungserfassung

The final evaluation will be based on:

• Initial implementation / idea presentation, 10%

• Final presentation, 20%

• Report (research paper), 12-18 pages, 30%

• Implementation , 40%

• Participation in the seminar (bonus)

Termine

(apart from the presentations, there will be no regular meetings in our seminar room!)

17.04.2023 (15:15-16:30), Room: K 1.02 Zoom link here, code: 517025	Introduction and QA session for seminar topics. (Hybrid) [Slides]
until 24.04.2023	Registration: HPI students: via moodle Students from UP: an Studienreferat (Studienreferat(at)hpi.uni-potsdam.de) Sending your preferred and secondary topic to: mml-team@hpi.de
until 28.04.2023	Topics and Teams finalized Arranging individual meeting
weekly	Individual meeting with your tutor
12.06.2023 (15:15-16:30), G1-E.15/16 Zoom link here, code: 517025	Mid-Term presentation (15+5min),
24.07.2023 (15:15-16:30), G1-E.15/16 Zoom link here, code: 517025	Final presentation (15+5min)
31.08.2023	Hand in code and paper (Latex template)
until 30.09.2023	Grading finished

Zurück