Munich AI Day 2024

1. Can we reconstruct images that a person saw, directly from his/her fMRI brain recordings?

2. Can we reconstruct the training data that a deep-network trained on, directly from the parameters of the network?

The answer to both of these intriguing questions is “Yes!”

In this talk I will show our work in both of these domains. Furthermore, exploring the two in tandem, under a single computational framework, can lead to significant breakthroughs in both fields. I will show how combining the power of Brains & Machines can potentially be used to bridge the gap between those two domains.

In today's rapidly evolving world, technology is profoundly transforming healthcare. This talk will explore five key elements driving this change, focusing on how Artificial Intelligence (AI) can shape the future of medicine and healthcare.

1. Proactive and Personalized Care: We are moving away from traditional reactive models to a proactive approach. AI-driven analytics and predictive modeling enable us to anticipate health issues before they occur. This shift not only improves patient outcomes but also revolutionizes our approach to wellness and disease prevention.

2. Tailored Treatments for Individuals: The era of one-size-fits-all is ending. By using AI to analyze vast clinical records, we can determine the best treatment options for each patient, considering their unique characteristics. This personalized approach introduces a new age of precision medicine.

3. Streamlining Healthcare Systems: Efficiency and resource allocation in healthcare systems need to undergo significant improvements to better support an aging population. AI can help optimize clinical resources, ensuring that the right care is delivered at the right time, reducing costs, and improving patient satisfaction.

4. Enhanced Healthcare Data: The quality, continuity, and accessibility of data are crucial in order for AI to be able to transform healthcare. AI can enhance data quality, integration and management, enabling seamless information flow across different platforms, systems and even countries.

5. Educating Clinicians on AI: For AI to reach its full potential, clinicians must understand its applications and implications. This talk will highlight the importance of AI education for healthcare professionals, ensuring not only that they are prepared to use these technologies effectively and ethically, but also that they are able to drive innovation in healthcare using AI.

Humans can be effective autodidacts once they have a strong foundation in a particular area. The same turns out to be true for Large Language Models (LLMs) - even though the idea of an LLM training itself may seem suspect at first.

However, a naive approach results in so-called model collapse. We will discuss the general principles of successful "self-training" and apply them to the case of improving an LLM's capability to produce long form content. Our evaluation shows that our synthetically generated dataset is superior to existing manually constructed instruction tuning datasets. This approach is promising for many skills and domains for which training data are sparse and too expensive to generate manually.

Data in form of graphs occurs in many applications, including network analysis, recommendation, physics, chemistry, biology and drug design, and AI for optimization. With the success in many applications come further open questions, about modeling, stability and transferability of such models. This talk will summarize some recent research questions and results in these directions.

The use of neural representations in computer vision research has recently grown tremendously. While Neural representations are now making their way to real products, some significant gaps still exist that limit their applicability in real world scenarios. In this talk I will illustrate some our recent research work aimed to fill such gaps, with a focus on 3D scene understanding and for common tasks such as novel view synthesis, 3D semantic segmentation and 3D asset generation.

For each of these three tasks, I will first highlight some important practical limitations of current neural representations. I will then show solutions leveraging neural representations to overcome such limitations, which include high fps novel view synthesis, open set 3D semantic segmentation and realistic 3D asset generation from text prompts.