Module 5 — Compositionality and the symbolic / subsymbolic divide

dreamy bubblegum dance, reggae k-pop, electro-jungle, p-funk samba · 3:36

Listen on 93

Lyrics

[Verse 1]
When vectors collide in the neural space
Two patterns merge but they lose their place
The superposition creates a mess
Which feature belongs where becomes anyone's guess
Smolensky found the tensor trick
Outer products make the binding stick
But dimensions explode beyond control
Mathematical beauty takes its toll

[Chorus]
Bind the symbols, make them dance
In hyperdimensional expanse
Circular convolution holds the key
Neural symbols living free
Spikes and vectors intertwined
Structured thoughts in neural mind
From subsymbolic to the clear
Symbol manipulation's here

[Verse 2]
Plate discovered the golden path
Holographic reduced with convolution math
Fixed dimensions, approximate but true
Vector symbolic breakthrough
Random patterns quasi-orthogonal
High dimensions make recovery possible
Kanerva showed us why it works
In hyperdimensional neural networks

[Chorus]
Bind the symbols, make them dance
In hyperdimensional expanse
Circular convolution holds the key
Neural symbols living free
Spikes and vectors intertwined
Structured thoughts in neural mind
From subsymbolic to the clear
Symbol manipulation's here

[Bridge]
Eliasmith built the bridge complete
Neural Engineering Framework neat
Spaun perceives and reasons whole
Biological constraints under control
Semantic pointers guide the way
Working memory holds and plays
Probabilistic programs run
Neural substrate, symbolic fun

[Verse 3]
The binding problem meets its match
Role-filler pairs no longer catch
Variable binding in the brain
Distributed patterns break the chain
From connectionist to symbolic thought
The hard problem finally fought
Neurosymbolic architectures rise
Bridging the computational divide

[Verse 4]
Transformers learn to bind and parse
Attention weights cross semantic space
Large language models show the way
Vector embeddings come to play
But consciousness still hides its code
In distributed neural episodes
The symbol grounding problem waits
While AI dreams at silicon gates

[Chorus]
Bind the symbols, make them dance
In hyperdimensional expanse
Circular convolution holds the key
Neural symbols living free
Spikes and vectors intertwined
Structured thoughts in neural mind
From subsymbolic to the clear
Symbol manipulation's here

[Outro]
In the space of thousand dimensions
Symbol processing finds extensions
Brain encodes with vector grace
Structured reasoning finds its place

← Module 4 — The structural substrate: connectomes and integration | Module 6 — Cognitive maps: reasoning as navigation →