While many brain simulations focus on scale (number of neurons), few focus on function (what the neurons actually do). In 2012, Chris Eliasmith and colleagues at the University of Waterloo broke this trend with Spaun (Semantic Pointer Architecture Unified Network). Unlike its predecessors, which were often silent networks of randomly connected neurons, Spaun could “see,” “think,” and “act.”

The Semantic Pointer Architecture (SPA)

The core innovation of Spaun is the Semantic Pointer Architecture. This theoretical framework explains how biological neurons can represent complex symbols.

• Representation: High-dimensional vectors are encoded into the spiking patterns of neuron populations.
• Transformation: Neural connections perform mathematical operations (like addition or convolution) on these vectors.
• Semantic Pointers: Compressed representations that carry information between different brain areas, allowing for efficient communication without sending raw data.

Functional Capabilities

Spaun consists of 2.5 million spiking neurons organized into subsystems representing the visual cortex, basal ganglia, thalamus, and motor cortex. It can perform eight distinct tasks:

1. Copy drawing: Reproducing a visual input.
2. Image recognition: Identifying handwritten digits.
3. Reinforcement learning: Learning bandit tasks.
4. Serial working memory: Remembering a list of numbers.
5. Counting: Moving a semantic pointer along a number line.
6. Question answering: Simple logical reasoning.
7. Fluid intelligence: Completing patterns (Raven’s matrices).
8. Rapid variable binding: Assigning values to variables syntax.

Biological Realism

Spaun is not an artificial neural network (ANN) in the traditional sense of machine learning. It uses spiking integrate-and-fire neurons.

• Neurotransmitters: It models the effects of GABA and AMPA/NMDA receptors.
• Basal Ganglia: It implements a biologically plausible action-selection mechanism where the basal ganglia disinhibit specific thalamic loops to “select” an action.
• Limitations: It is still a simplified model. It does not simulate dendritic computation or detailed ion channel dynamics, focusing instead on the population code level.

Path Forward

Spaun demonstrated that we do not need to simulate every molecule to get functional behavior. It suggests that if we can identify the “code” (like Semantic Pointers) used by the brain, we can build functional emulations at a higher level of abstraction than molecular simulation.

Official Sources

• Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., & Rasmussen, D. (2012). A large-scale model of the functioning brain. Science, 338(6111), 1202-1205.
• Nengo Neural Simulator