Welcome to the first lesson of your course. We're beginning our exploration of the human system's architecture at the most fundamental level of cortical computation. Your goal is to build a comprehensive, mechanistic understanding of what makes activities energizing or depleting, and to do that, we must first understand the algorithm the brain uses to model the world and itself.
This first module, "The Predictive Brain," will lay that foundation. In this lesson, we will address how the abstract theory of predictive coding is physically implemented in the brain's hardware. You will learn how the brain's cortical pyramidal neurons are organized to generate predictions and calculate errors, and how the system dynamically modulates the influence of these error signals. This is the implementation level—the bridge from abstract algorithm to the nuts and bolts of neural circuitry.
The Universal Algorithm of the Neocortex
For over a century, neuroanatomists have known that the neocortex, the seat of our higher cognitive functions, possesses a remarkably uniform structure. It's composed of repeating units called cortical columns. This uniformity strongly suggests that the cortex employs a single, universal algorithm for processing information, regardless of whether it's handling vision, sound, or abstract thought. The "Thousand Brains Theory" is a modern expression of this idea.
To begin, let's watch a short video that introduces this concept of a canonical computation.
Your Brain Is 150,000 Mini-Brains
This video, "Your Brain Is 150,000 Mini-Brains," provides an excellent visual introduction to the columnar hypothesis—the idea that the neocortex is built from a vast number of identical processing modules.
Watch the first three minutes, from the beginning to the introduction. Focus on how the shared architecture of the neocortex implies a universal underlying algorithm.
The video sets up a central premise you are already familiar with: the brain's primary function is to build a predictive model of the world. Predictive coding is the leading framework for how it achieves this. The core of this framework is a continuous, hierarchical process of generating predictions about the causes of sensory input and updating those predictions based on the error between what was predicted and what was sensed.
The Components: Predictions, Errors, and Precision
To understand the neural implementation, we first need to be precise about the computational components. A predictive coding system operates through a reciprocal exchange of signals between hierarchical levels.
- Predictions are top-down signals. A higher level of the hierarchy generates a prediction about the activity it expects to see in the level below it.
- Prediction Errors are bottom-up signals. A lower level computes the discrepancy (the error) between the top-down prediction and its own activity, sending this error signal to the level above.
The goal of the entire system is to minimize prediction error at all levels, which is equivalent to maximizing the accuracy of its generative model of the world. This is the operational core of Friston's Free Energy Principle.
A key paper by Shipp, Adams, and Friston outlines the principles of this process. Please read the following sections to formalize these concepts.
Reflections on agranular architecture: predictive coding in the motor cortex - PMC
This paper details the principles of predictive coding and its proposed implementation. While its focus is the motor cortex, the foundational principles it lays out are universal.
First, read the initial paragraphs of the section "The principles of predictive coding" to solidify the concept of hierarchical generative models and the interplay of prediction and error. Please read from this passage. Next, read the subsection titled "Expectations: causes, states, and precisions." This introduces the three distinct types of representations the system must maintain. Focus on the definition of each: Causes: Invariant, categorical aspects of the world (e.g., an object's identity). States: The dynamic, moment-to-moment properties of causes (e.g., an object's motion). Precisions: The reliability or confidence assigned to predictions and prediction errors. Read the text from this part.
The concept of precision is critical. It is the system's estimate of the reliability of the prediction error signal. By modulating precision, the brain can control the "volume" or "gain" of specific error signals, effectively directing attention. An error signal with high precision will have a strong influence on updating the model's predictions, while a low-precision signal will be largely ignored.
Mapping Computation to Cortex
Now we arrive at the core of today's outcome: How does the six-layered neocortex implement this message-passing scheme? The leading hypothesis, often called the "neocortical proposal," maps these computational units onto specific types of neurons in specific cortical layers.
- Prediction units are thought to be implemented by deep pyramidal cells, primarily in layers V and VI. These neurons are the source of the top-down/backward projections that convey predictions to lower cortical areas.
- Prediction error units are thought to be implemented by superficial pyramidal cells, primarily in layers II and III. These neurons are the source of the bottom-up/forward projections that convey prediction errors to higher cortical areas.
The following PDF provides a clear, rigorous explanation of this mapping. It is a draft paper by Mark Sprevak, a philosopher of cognitive science, and it treats the topic with the intellectual precision you appreciate.
[PDF] Predictive coding IV: The implementation level - Mark Sprevak
This document explores how the abstract algorithm of predictive coding relates to the physical "nuts and bolts" of the brain.
Please read Section 4, "Implementing prediction units and error units." Focus on the distinction between superficial and deep pyramidal cells and their respective roles. Note the discussion on population coding versus one-to-one mapping, which speaks to the complexity of this implementation. Read from this section.
To visualize this circuitry in greater detail, let's return to the Shipp, Adams, and Friston paper. Their Figure 3 provides a canonical microcircuit diagram.
The key takeaway is the anatomical basis for the opposing flows of information: prediction errors ascend from superficial layers, while predictions descend from deep layers. This segregated, bidirectional signaling is the fundamental architectural motif of cortical processing.
In his lectures, Karl Friston explains how this anatomical mapping arises naturally from the mathematics of the Free Energy Principle when expressed as a process theory of neuronal dynamics.
Karl Friston - 2016 CCN Workshop: Predictive Coding
In this workshop talk, Friston provides a dense but foundational overview of predictive coding.
First, watch the segment from 21:58 to 23:11, where he explains the crucial step of turning the optimization problem into a gradient descent that can be implemented by a dynamical system like the brain. Next, watch from 34:35 to 36:47. Here, he explicitly maps the computational quantities—expected states and prediction errors—onto the message-passing between deep and superficial pyramidal cells across the cortical hierarchy.
The Implementation of Precision Weighting
The final piece of the puzzle is understanding how precision—the gain on error units—is implemented. This cannot be encoded in the fixed wiring of the cortex, as attention and confidence need to shift on a millisecond timescale.
The prevailing theory proposes two primary mechanisms for this rapid, dynamic gain control:
- Neuromodulation: Chemicals like dopamine (DA), acetylcholine (ACh), norepinephrine (NE), and serotonin (5-HT) act as broadcast signals that can rapidly change the "excitability" or synaptic gain of large populations of neurons, particularly the superficial error units.
- Fast Synchronized Firing: High-frequency oscillations (especially in the gamma band, 30-90 Hz) in the signals arriving at a neuron can increase its responsiveness. This synchronous input acts as a powerful gain amplifier. It is hypothesized that error signals are carried by gamma-band frequencies, while predictions are carried by slower beta-band frequencies, allowing for selective amplification of errors.
Let's turn back to the Sprevak paper for a detailed treatment of these mechanisms.
[PDF] Predictive coding IV: The implementation level - Mark Sprevak
This section outlines the leading candidates for the physical implementation of fast changes in precision weighting.
Please read the first few paragraphs of Section 5, "Implementing precision weighting." Focus on the distinction between slow synaptic changes (learning) and the fast changes required for attention. Read from the start of the section. Now, continue reading the following two paragraphs that detail the roles of neuromodulators and fast synchronized firing.
These mechanisms are not mutually exclusive; in fact, they are deeply interactive. For example, the release of acetylcholine is known to promote gamma-band oscillations. Together, they provide the system with a flexible toolkit for dynamically weighting sensory evidence against its internal predictions, allowing it to navigate a noisy and uncertain world.
Conclusion
In this lesson, we have established the foundational neural architecture of the predictive brain. We have moved from the abstract algorithm of predictive coding to its concrete implementation in cortical microcircuits.
Here are the key takeaways:
- The uniform, columnar structure of the neocortex implements a canonical predictive coding algorithm.
- This algorithm involves a hierarchical, reciprocal exchange of signals: deep-layer pyramidal cells send top-down predictions, while superficial-layer pyramidal cells send bottom-up prediction errors.
- The system's goal is to minimize prediction error, thereby optimizing its generative model of the world.
- Precision weighting acts as a dynamic gain control on error signals, allowing the brain to modulate its confidence in sensory evidence. This is the neurobiological basis of attention.
- Precision is implemented by fast-acting mechanisms, primarily neuromodulators (like dopamine and acetylcholine) and synchronous neural oscillations (like gamma-band rhythms).
In our next lesson, we will dissect the concept of precision weighting further. We will move beyond the general mechanism to examine the specific, distinct roles that different neuromodulators play in assigning precision to different types of information, providing a deeper understanding of the brain's chemical architecture.