1 of 27

Distributed Representations and Superposition

Aaron Schumacher

Tuesday 2025-06-10

Data Science DC

2 of 27

3 of 27

4 of 27

5 of 27

“ strawberry”

6 of 27

The unit/value hypothesis

7 of 27

“ strawberry”

?

?

?

8 of 27

2012

2017

Interpretable neurons!

9 of 27

“ strawberry”

?

?

?

50k nodes

12k nodes

10 of 27

11 of 27

1986

course coding etc.

12 of 27

Combinatorial approaches don’t compose well

  • Unit/value: “strawberry” is 000001
  • Combinatorial: “strawberry” is 101010
    • Or is that “dog” 101000 + “big” 000010?
  • Problem: representations aren’t orthogonal
    • Only as many orthogonal directions as nodes

13 of 27

1,0,0

0,1,0

0,0,1

14 of 27

15 of 27

Quasi-

orthogonal dimensions (within 10 degrees of 90 vs. all others) based on drawing random (Gaussian) unit vectors

actual dimensions

exponential!

16 of 27

17 of 27

“Superposition”

  • More than one of these quasi-orthogonal features active at the same time

18 of 27

1986

binding problem:

red, blue, dog, cat means what?

19 of 27

2017

20 of 27

2003

21 of 27

22 of 27

2017

23 of 27

query

key

resulting value

24 of 27

2024

25 of 27

2025

26 of 27

27 of 27

Discussion?