Causal AI: Understanding Causality in AI with Dr. Robert Usazuwa Ness

The Essence of Causal AI

• 💡 Causal AI moves beyond correlation to understand the 'why' behind data, focusing on the data generating process rather than just the data itself.
• 🧠 Humans and animals intuitively grasp causality, a capability AI systems have historically lacked due to reliance on correlation-based learning.
• 🤖 While traditional causal inference uses statistical methods, AI is increasingly incorporating causal reasoning by borrowing ideas from other sciences, like cognitive science.

Bridging Causal Inference and AI

• 🌉 A key development is the connection between graphical causal inference (using Directed Acyclic Graphs - DAGs) and probabilistic graphical models, including deep probabilistic models.
• 🛠️ Libraries like Pyro (for PyTorch), NumPyro, Stan, and PMC (with its do operator) facilitate causal modeling by handling complex inference, allowing focus on causal assumptions.
• 📈 These tools enable simulating interventions and understanding cause-and-effect relationships, even when direct experimentation is infeasible or unethical.

Causal AI in Practice

• 🎯 Causal AI is crucial when needing to infer what would happen if you intervened in a system, such as testing a new medicine or understanding the impact of a feature in a video game.
• 📊 The workflow typically involves defining a causal question, specifying causal assumptions (often via a DAG), checking data feasibility, selecting an estimation method, and performing sensitivity analysis.
• ⚠️ While AI can automate parts of the process, it's essential to ground assumptions in an honest understanding of the data generating process, not just what's convenient in the dataset.

Generative AI and Causal Reasoning

• 🚀 Large Language Models (LLMs) like GPT-40 show promise in causal reasoning by acting as a knowledge base, synthesizing information from text to propose causal relationships and generate code for analysis.
• 💡 LLMs can help formalize causal questions and assumptions, but they still hallucinate and require careful validation, especially when suggesting estimation methods.
• 🎮 Future applications include using causality to enhance generative AI for video games, ensuring generated content adheres to underlying game mechanics and physics, potentially enabling composability of different game models.

Judea Pearl's Ladder of Causation

• 🪜 Level 1 (Association): Basic statistical correlation and observation.
• 🪜 Level 2 (Intervention): Answering 'what if' questions by simulating actions or experiments.
• 🪜 Level 3 (Counterfactuals): Imagining 'what might have been' by reasoning about hypothetical scenarios conditioned on observed evidence, which is more challenging and requires stronger assumptions.