spa-question-answering-control

A spiking SPA model that decides — using a basal-ganglia/thalamus action-selection loop — whether an incoming visual input is something to *remember* (a STATEMENT) or something to *answer* (a QUESTION), and acts accordingly.

model advanced NengoGUI-ready spa semantic-pointers binding question-answering working-memory action-selection basal-ganglia thalamus cognitive hbb-tutorial

spa-question-answering-control

A spiking SPA model that decides — using a basal-ganglia/thalamus action-selection loop — whether an incoming visual input is something to remember (a STATEMENT) or something to answer (a QUESTION), and acts accordingly.

Description

This is example 6 of chapter 5 of How to Build a Brain (Eliasmith, 2013), distributed with nengo-gui as 6-spa_question-control.py. It's the most elaborate of the chapter's question-answering models: instead of always-on cortical bindings, the right thing to do is selected dynamically based on the structure of the visual input.

There are three states:

visual — receives the current input, e.g. STATEMENT+RED*CIRCLE or QUESTION+BLUE
memory — recurrent SPA state, holds the sum of seen bound pairs
motor — read-out used to "answer" questions

And two competing actions, picked by spa.BasalGanglia / spa.Thalamus:

dot(visual, STATEMENT) --> memory = visual          # store the statement
dot(visual, QUESTION)  --> motor = memory * ~visual # answer the question

The utility of each action is the dot product of visual with one of the role pointers (STATEMENT, QUESTION). Whichever role is present wins, the basal ganglia inhibits the loser, the thalamus releases the winner, and the corresponding routing happens. Statements load memory; questions read it out.

A toy schedule of inputs runs:

0.1 – 0.3 s: STATEMENT + RED*CIRCLE
0.35 – 0.5 s: STATEMENT + BLUE*SQUARE
0.55 – 0.7 s: QUESTION + BLUE → motor should read out SQUARE
0.75 – 0.9 s: QUESTION + CIRCLE → motor should read out RED

This example is sensitive to dimensionality: 32 dimensions is small enough that you may need a couple of runs (or to bump dim higher) to see clean answers in motor. The tutorial keeps it at 32 to stay fast in the GUI.

A note on `nengo.spa` vs `nengo_spa`

This script uses the legacy nengo.spa module that still ships inside core Nengo. The newer, separately maintained nengo_spa package has a redesigned action-selection API. Both run today, but new submissions are generally encouraged to prefer nengo_spa; this submission preserves the original tutorial verbatim.

Run it

In NengoGUI (the intended way to view it):

pip install nengo-gui
nengo spa_question_answering_control.py

Or just at the command line:

python spa_question_answering_control.py

How it works

spa.BasalGanglia builds a spiking neural model of the direct/indirect striatal pathway: a population per action whose firing rate scales with the action's utility (here, dot(visual, STATEMENT) or dot(visual, QUESTION)). The pathway competitively inhibits all but the highest-utility action's GPi output. spa.Thalamus reads GPi, gates the corresponding routing connection (here, either "copy visual into memory" or "compute memory * ~visual into motor"), and lets the winning action execute.

The result is a closed-loop control system: stimuli arrive in visual, the BG/thalamus pick which routing to enable based on the role pointer present, and memory/motor update accordingly. No clock, no explicit gating signals from outside — selection emerges from the dynamics of the action utilities.

memory uses feedback=1, feedback_synapse=0.1 to integrate inputs over a 100 ms timescale, giving the statement enough time to be written in cleanly without smearing into the subsequent question.

Citation

@book{eliasmith2013htb,
  author    = {Eliasmith, Chris},
  title     = {How to Build a Brain: A Neural Architecture for Biological Cognition},
  publisher = {Oxford University Press},
  year      = {2013}
}