USC Information Sciences Institute Courses

The Surprising Effectiveness of Membership Inference with Simple N-Gram Coverage

Discover how N-Gram Coverage Attack enables membership inference on black-box language models, outperforming existing methods while requiring only text outputs, with implications for privacy and copyright protection.

• YouTube
• 1 hour 3 minutes
• Self-Paced
• Free Video

Making Security Useful Without Being a Burden

Explore cybersecurity challenges with Erik Kline as he discusses making security user-friendly, debunks media misconceptions, and shares his journey from gaming to protecting people online.

• YouTube
• 21 minutes
• Self-Paced
• Free Video

Experiments in Scaling Reinforcement Learning with Verifiable Rewards

Explore the latest advancements in reinforcement learning with verifiable rewards, focusing on language models like DeepSeek's R1 and AI2's OLMo, and gain insights into how RL is reshaping the AI landscape.

• YouTube
• 1 hour 1 minute
• Self-Paced
• Free Video

Texera: An Open-Source System for Cloud-Based Collaborative Data Science and AI/ML Using Workflows

Explore Texera, an open-source cloud platform for collaborative data science and AI/ML using workflows, enabling users with various backgrounds to conduct AI-centric data science with Google Docs-like collaboration features.

• YouTube
• 1 hour 1 minute
• Self-Paced
• Free Video

You Can Only Design a System to the Limit of What You Can Plan For

Explore systems thinking with Joe Barnes as he discusses how competition drives innovation, the importance of building, and the future of human-AI collaboration.

• YouTube
• 20 minutes
• Self-Paced
• Free Video

LLMs Do Not Have Human-Like Working Memories

Explore the limitations of Large Language Models in replicating human working memory through three experimental games, revealing key challenges for achieving artificial general intelligence.

• YouTube
• 54 minutes
• Self-Paced
• Free Video

Ushering Agents to an Open Social World

Explore three foundational approaches for building AI agents that learn through interaction: web exploration, social norm internalization, and human feedback-based evaluation metrics for better integration into our social world.

• YouTube
• 1 hour 7 minutes
• Self-Paced
• Free Video

How I Learned to Stop Worrying and Love AI

Explore Dr. William Regli's perspective on AI, challenging both optimistic and pessimistic views while offering a socio-technical framework to address emerging challenges in human-AI integration.

• YouTube
• 1 hour
• Self-Paced
• Free Video

Contextualization for Human-AI Interactions

Explore how enhancing AI's contextual understanding is crucial for effective human-AI interactions, covering speech interfaces, personalization, and holistic evaluation approaches for complex tasks.

• YouTube
• 1 hour 3 minutes
• Self-Paced
• Free Video

Can We Blame the Chatbot if it Goes Wrong?

Explore the complex issue of AI chatbot accountability through the Air Canada case, examining responsibility frameworks, transparency requirements, and comparing regulatory approaches in the EU and Canada.

• YouTube
• 53 minutes
• Self-Paced
• Free Video

Evaluating Sparse Autoencoders with Board Game Models

Explore novel evaluation methods for Sparse Autoencoders using board game models, focusing on coverage metrics, state reconstruction, and the innovative p-annealing approach.

• YouTube
• 41 minutes
• Self-Paced
• Free Video

Towards Socially Aware and Safe AI Agents

Explore groundbreaking research on developing socially intelligent and safe AI systems, focusing on LLM evaluation frameworks, information asymmetry challenges, and safety considerations in human-AI interactions.

• YouTube
• 1 hour 1 minute
• Self-Paced
• Free Video

We Over Rely on Technology but Lack the Resilience We Need to Depend on It

Explore cybersecurity challenges with David Balenson as he discusses autonomous vehicle safety, AI lessons from cybersecurity, and the paradox of our technological dependence versus resilience.

• YouTube
• 20 minutes
• Self-Paced
• Free Video

Designing Priors for Bayesian Neural Nets to Enhance Probabilistic Predictive Models in Engineering

Explore how Bayesian neural networks can enhance engineering predictive models by integrating physics-based knowledge and providing robust uncertainty quantification for high-consequence decision-making applications.

• YouTube
• 1 hour
• Self-Paced
• Free Video

MrT5: Dynamic Token Merging for Efficient Byte-level Language Models

Explore dynamic token merging in MrT5, a more efficient byte-level language model that reduces sequence lengths by up to 75% while maintaining performance comparable to ByT5 across multilingual tasks.

• YouTube
• 1 hour 4 minutes
• Self-Paced
• Free Video