Researchers have developed a new machine learning model that learns like the brain, sidestepping the need for error backpropagation. The Meta-Representational Predictive Coding (MPC) model, proposed by Alexander Ororbia, Karl Friston, and Rajesh Rao, combines predictive coding and active inference to learn representations through "sensory glimpsing" — actively sampling informative parts of the input.
How It Works
Traditional self-supervised learning relies on backpropagation, a biologically implausible method. In contrast, MPC uses a hierarchical predictive coding framework where each layer predicts the activity of the layer below. Instead of predicting raw pixels, MPC predicts representations across parallel streams, reducing computational complexity. The model uses active inference to decide where to "glimpse" next, driving representation learning through sequences of decisions.
Why It Matters
This work bridges neuroscience and AI, offering a more efficient and plausible way to learn without human annotations. For your own brain, it highlights the importance of active exploration—seeking out new information—for learning. Just as MPC learns by sampling informative glimpses, you can enhance your learning by actively searching for challenging problems and diverse experiences.
What You Can Do
- Embrace active learning: Instead of passive reading, quiz yourself, solve puzzles, and seek feedback.
- Diversify input: Expose your brain to varied domains—science, art, music—to build richer mental representations.
- Try brain training: Engage with adaptive challenges that push your cognitive boundaries, much like MPC's glimpsing strategy.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
