MLx Representation Learning & Generative AI 2026

This module examines how AI is moving beyond narrow automation toward open-ended systems that can generate novel artifacts, learn in rich environments, and support scientific discovery. It traces the shift from early automation ideas to foundation models, explains why game-like simulation success does not transfer directly to the real world, and introduces world models as controllable simulators for embodied learning. It also explores how scaling, curriculum learning, and latent actions can expand capability, and how foundation models may help automate innovation through iterative generate-test-refine loops.

1. 1. Open Endedness, World Models and The Automation of Innovation

This module extends the benchmarking unit by deepening how model performance is measured, compared, and interpreted in practice. Learners examine the strengths and blind spots of static and dynamic evaluation, including saturation, task mismatch, noisy signals, and cost-aware tradeoffs. The module also explores how ranking systems, agent-mode metrics, and real-world user feedback shape more reliable assessments of AI capability, especially for research, judgment, and efficiency.

1. 1. Benchmarking

This module introduces multi-robot and multi-agent learning for solving coordination problems that are hard for centralized planning, such as pathfinding, logistics, and collective transport. Learners examine why joint state spaces grow exponentially, how deadlocks arise, and how graph neural networks enable scalable decentralized decision-making. The module also covers imitation learning, communication-based coordination, hybrid neural-classical solvers, and key challenges in sim-to-real transfer, robustness, and trust for real-world robotic systems.

1. 1. Part 1: Multi-Robot and Multi-Agent Learning
2. 2. Part 2: Multi-Robot and Multi-Agent Learning

This module introduces intelligent data gathering as a principled approach to experimental design. Learners explore how entropy and expected information gain quantify uncertainty and guide decisions in settings such as adaptive questioning and source localization. The module then connects these ideas to Bayesian decision theory, showing how beliefs, losses, and future actions shape optimal data collection. It concludes with scalable methods like deep adaptive design and ActionBED for training end-to-end systems that optimize data gathering for task-specific objectives in real-world applications.

1. 1. Intelligent Data Gathering

This module introduces causal learning as a framework for answering intervention and counterfactual questions beyond prediction. Learners will examine confounding, randomized trials, potential outcomes, and graphical models to understand how causal assumptions support robust decision-making under uncertainty. The module also connects causal reasoning to distribution shift, transfer learning, and targeted machine learning methods for estimating treatment effects, with an emphasis on discovery and extrapolation in real-world settings.

1. 1. On Causal Learning: Discovery and Extrapolation

This module introduces multimodal AI as the study of learning from heterogeneous data such as text, vision, audio, touch, and smell. Building on the foundations of representation learning and generative AI, it explores how modalities share information, interact, and complement one another. Learners will examine key challenges in alignment, reasoning, generation, transfer, and quantification, along with modern approaches such as multimodal foundation models, adapter-based conditioning, and contrastive learning. The module also highlights emerging applications in tactile sensing and smell recognition and generation.

1. 1. Multimodal AI

This module explores human-centric cooperative AI in multi-agent systems, focusing on how autonomous agents coordinate, compete, and align with human goals in shared environments. Learners will examine multi-agent reinforcement learning, social dilemmas, collective intelligence benchmarks, and the practical challenges of scalability, partial observability, and incentive misalignment. The module also covers human-in-the-loop alignment using imperfect feedback and translating natural language into safety constraints for safer deployment in tasks such as cooperative games, cooking, traffic, and navigation.

1. 1. Human-Centric Cooperative AI: Evaluating, Aligning & Governing Multi-Agent Systems

This module introduces continuous-time generative modeling through flow matching, diffusion, and optimal transport. Learners will connect probability paths, vector fields, and continuity equations to the process of transforming noise into data, and compare these methods with autoregressive models, GANs, and VAEs. The module emphasizes simulation-free conditional training, bridges, and guidance strategies, while highlighting how optimal transport can produce straighter, more efficient sampling paths. Practical implementation choices and their impact on stability, inference speed, and flexibility across domains are also explored.

1. 1. Continuous-Time Generative Modeling: Matching and Optimal Transport

This module examines the rapid rise of frontier AI in mathematical reasoning, from elementary benchmarks to research-level problem solving. Learners will explore how supervised learning, chain-of-thought prompting, self-taught reasoning, reinforcement learning, and inference-time search improve performance. The module also covers agentic verification workflows and recent autonomous mathematics results, highlighting both breakthroughs and limitations. By the end, learners will understand the evolving capabilities of AI systems in math, the role of breadth and counterexamples, and where formal methods still matter.

1. 1. The Rise and Ramifications of Artifical Mathematical Intelligence

This module introduces geometric deep learning through symmetry, group theory, and non-Euclidean geometry. Learners explore invariance and equivariance, then see how these ideas shape supervised learning, CNNs, graph neural networks, transformers, spherical CNNs, and geometric graph methods. The module connects architectural priors to real-world applications in vision, language, football analytics, and protein folding, showing how exploiting data symmetries can improve model design and performance.

1. 1. Geometric Deep Learning

This module explains why mathematics is formalized and how formalization supports reliable reasoning in AI and theorem proving. Building on the shift from manual proof checking to Lean and AI-assisted mathematics, it shows how formal systems make proofs precise, machine-verifiable, and reusable. Learners will see why formalization matters for automation, library growth, and trust in advanced mathematical results, while also understanding the practical challenges of compatibility, technical debt, and verification in modern mathematical research.

1. 1. Why Formalize Mathematics?

This module introduces probing neural networks as a way to inspect hidden activations for debugging, safety, and interpretability. Learners will explore how frozen classifiers can read internal representations, the linear representation hypothesis, and the differences between linear and nonlinear probes. The module also highlights common pitfalls, including misleading probe accuracy, weak baselines, and the gap between correlation and causation. Finally, it presents activation patching as a stronger method for testing whether a model actually uses the information encoded in its internal states.

1. 1. Probing Neural Networks