Overview

Our AI safety research is organized into four main areas, focusing on enhancing AI's robustness, adaptability, transparency, and practical applications. We want to ensure that AI technologies remain reliable, fair, and beneficial for society.

 

Machine Robustness

Robust AI systems are essential for maintaining performance in dynamic and adversarial environments. Our research in this area explores methods to enhance the resilience of AI models against adversarial influences and evolving challenges.

• Uncertainty Quantification and Robustness: Measuring model confidence and ensure predictions remain reliable despite noise or unexpected inputs.
• Preventing Adversarial Attacks: Developing defenses to protect models from subtle, malicious input modifications.
• Factuality in Large Language Models: Improving mechanisms to verify that generated outputs are factually accurate and resistant to misleading information.
• Outlier and Changepoint Detection: Identifying abnormal data points or shifts in data distribution to detect anomalies and adapt models quickly.
• Visual Object Tracking: Designing tracking systems that reliably follow objects in videos, even under challenging conditions such as occlusions or rapid movements.

Generalization

For AI to be truly intelligent, it must generalize across diverse contexts rather than relying on superficial patterns. Our research focuses on advancing AI’s ability to adapt and reason across different domains.

• Artificial General Intelligence (AGI):
  Exploring methods to build AI systems that can adapt flexibly to a wide range of tasks, rather than excelling at just one specific function.
• Bias and Shortcuts:
  Investigating how models learn unintended biases and rely on shortcuts, and design strategies to reduce these issues for fairer and more robust outcomes.
• Multi-component Neuro-symbolic Generalization:
  Combining neural network learning with symbolic reasoning to boost abstraction and logical inference, enhancing AI's problem-solving capabilities across diverse domains.

Transparency and Privacy

Building AI systems that are interpretable and privacy-preserving is crucial for ethical and accountable deployment. Our research focuses on balancing explainability with user privacy and compliance with legal frameworks.

• Explainable AI: Improving transparency so users and stakeholders can understand how models make decisions.
• Privacy: Preventing attacks like membership inference by using techniques (e.g., Differential Privacy) that protect individual data contributions.
• Machine Unlearning: Developing methods to remove specific information from a model without compromising overall performance.
• Synthetic Data Generation: Creating realistic, privacy-preserving datasets to train and evaluate AI models without using sensitive real-world data.

Applications

AI has the potential to revolutionize key industries by improving efficiency, fairness, and decision-making. Our research ensures that AI applications are both impactful and responsible.

• AI for Finance: Using AI to enhance risk assessment, detect fraud, and support smarter financial decision-making.
• AI for Healthcare: Leveraging AI-driven insights to improve diagnostics, patient care, and advance medical research.
• AI in Education: Integrating AI in educational settings to personalize learning and boost data-driven learning analytics.
• AI for Ethics: Refining AI models, especially LLMs, to better align with ethical guidelines, societal values, and user expectations.