The Embodied AI

Have you ever thought about how computers learn to do things like recognize your face or suggest the next song you might like? It’s not magic—it’s machine learning, and it’s somewhat similar to how we humans learn from experience. Yann LeCun, the Chief AI Scientist at Facebook and a Silver Professor at the Courant Institute, New York University, offers an insightful “cake analogy” to explain the complexities of machine learning. Let’s slice into this cake to uncover what each layer reveals about AI, and perhaps, about our own learning processes as well.

The Big Base: Self-Supervised Learning

Imagine the bottom and biggest layer of our cake as self-supervised learning. This innovative approach builds on what used to be known as unsupervised learning. Picture yourself watching what happens when you push a swing; no one needs to tell you the result—you see and learn from the action directly. Self-supervised learning operates similarly, letting machines discover patterns and predict outcomes from data that hasn’t been explicitly labeled.

This self-driven way of learning resembles how we, as humans, intuitively grasp the world around us. We don’t rely on neatly labeled datasets to understand life’s complexities. Instead, we use our innate abilities to predict and reason, inferring what may happen next based on the information we observe. For instance, even with incomplete information, such as a partially hidden sign or an overheard conversation, we can often piece together the full story.

Yann LeCun emphasizes the critical role of what he terms “predictive learning” in breaking through the current limitations of AI development. While today’s AI can recognize images or voices with high accuracy, it struggles with tasks that require deeper reasoning, like understanding relationships between different objects or predicting human movements—skills that come naturally to humans through predictive inference.

LeCun describes this advanced form of learning as the ability to predict any aspect of the past, present, or future based merely on available cues. The richness and volume of data a machine can handle for prediction define the potential complexity and capability of the learning system. In essence, the more scenarios a machine can predict, the more intelligent it can become.

The Sweet Middle: Supervised Learning

The middle layer, the icing of the cake which is a bit thinner, is supervised learning. This is like learning to cook by following a recipe book. You get specific instructions on what to do, and you learn by following them closely. In AI, supervised learning happens when we feed the computer data that’s already labeled. For example, we show it many pictures of cats and dogs, telling it which ones are which. The computer uses these examples to learn how to identify cats and dogs in new pictures.

The Cherry on Top: Reinforcement Learning

And now, for the cherry—or rather, cherries—on top of our cake: reinforcement learning (RL). This layer of learning is about experimenting with different strategies and learning from the outcomes, much like playing a video game. You explore various tactics, see what works, and adapt based on success or failure. In the AI world, reinforcement learning involves machines making decisions, trying different actions, and adjusting based on rewards or feedback to ultimately achieve their objectives.

However, not everyone in the AI community agrees on the size of this cherry. Many in the reinforcement learning field see their method not just as a small garnish but as a key ingredient, potentially overflowing with value. At NIPS 2017, UC Berkeley Professor Pieter Abbeel added a playful twist to the ongoing debate by responding to LeCun’s analogy with their own version of the cake, one topped with not just one cherry but many, emphasizing the significant potential they believe RL holds.

Why This Matters

This cake analogy isn’t just cute—it helps us understand how layered and complex AI learning is, and it even mirrors some aspects of how we learn. For instance, all types of learning, whether human or machine, involve making predictions based on past experiences. By connecting these methods—self-supervised learning laying the groundwork, supervised learning refining the details, and reinforcement learning perfecting the strategies—AI is developing in a way that’s increasingly sophisticated and effective, much like how we learn to move from simple tasks to complex problem-solving as we grow.

As AI continues to evolve, understanding these layers helps us appreciate not just how far technology has come, but also how it parallels our own learning processes. And who knows? Maybe understanding this can help us teach both computers and humans better.