1. Complementary Learning Systems
McClelland, McNaughton & O’Reilly, 1995
McClelland & Goddard, 1996
Anthony Cate
March 22, 2001

2. Hippocampal Amnesic Syndrome
Patient HM
Bilateral medial temporal lobectomy
Still alive

3. Patient HM - Deficits
Explicit memory for events and episodes
New facts (arbitrary)
Paired Associate Learning
“locomotive-dishtowel”

4. Patient HM – Preserved Abilities
Motor skill acquisition
Implicit learning - priming
Memories from before damage, with a qualification…

5. Temporally Graded Retrograde Amnesia

6. The sea monster in your head:

7. Hippocampal Anatomy
Archicortex
Phylogenetically older than neocortex, fewer layers
Much smaller than neocortex, especially in humans

8. Inputs from diffuse areas of neocortex converge on hippocampus (via the Entorhinal Cortex)

9. Hippocampus forms a simple circuit:

10. Hippocampus encodes arbitrary conjunctions

11. Hippocampus reproduces patterns during sleep

12. The McC, McN & O’R model
Information processing takes place via the propagation of activation among neurons in the neocortical system

13. The McC, McN & O’R model
Implicit learning results from small changes to connections among neurons active in each episode of processing
Single processing episodes produce item-specific effects
Skills arise through accumulation of episodes

14. The McC, McN & O’R model
New arbitrary associations are based on learning that takes place in the hippocampal system that occurs alongside with processing occurs in the cortex
Bidirectional connections between neocortex and hippocampal system
Large connection changes in hippocampus for fast learning
Recall of new associations depends on pattern completion in the hippocampus

15. The McC, McN & O’R model
Consolidation of a memory occurs through the accumulation of small weight changes in neocortical connections
These changes are spread over time
They result from repeated reinstatements of the hippocampal memory to the neocortex
This may happen when a memory is retrieved from the hippocampal system, and during sleep.

16. Discovery of Structure
How the cortex learns about the world:
Discovery of Structure
via
Interleaved Learning
Rumelhart 1990
Train a multi-layer network on a set of propositions about things in the world

17. An intuitive view of conceptual structure:

18. Rumelhart 1990

19. How does the model discover this structure?
The concept representation units exploit similarities between patterns
The most efficient way to reduce error is by grouping inputs with similar outputs on the same concept unit

20. With interleaved learning at a slow rate, concept unit representations differentiate

21. These patterns of activation describe a hierarchical structure!

22. Points from Rumelhart’s model
Interleaved learning allows a network to discover a structure in an environment of patterns
A “hierarchical” relationship can be found in this structure based on the graded similarity of particular patterns.

23. Rumelhart’s model depends on:
Interleaved learning
Slow learning rate
A set of inputs that overlap in similarity
What if learning had to take place in a situation where none of these applied?

24. Failures of this architecture (Why a neocortex alone is inufficient):
One-shot learning
(=high learning rate)
Focused learning
(=non-interleaved training)

25. Catastrophic Interference!

26. Modeling Paired Associate Learning
McCloskey & Cohen 1989
AB-AC paradigm:
List A: List B: List C:
locomotive dishtowel seltzer
table pinecone headphones
weasel jacket waterfall
… … …

27. A simple model for paired associate learning:

28. The model completely fails to retain the AB associations once the AC associations are learned.

29. Catastrophic interference is a general phenomenon that occurs in complex models as well.
Training a network on a set of patterns with backprop and similar algorithms guarantees optimal weights for that set
Guarantees nothing about other patterns

30. Catastrophic interference in Rumelhart’s model:

31. Catastrophic interference can be overcome by changing a model’s architecture or the set of training patterns
Use patterns with less overlap
Weight changes for each pattern won’t affect performance of network for other patterns

32. But maybe we don’t want to make these changes.
Problem with less overlapping patterns:
No way to extract structure, because by definition no structure in pattern set
Structure = covariations

33. Another problem with less overlap:
Less generalization
A novel input will never be very similar to trained patterns, so the network cannot produce an appropriate output
We never see exactly the same thing twice

34. How arbitrary are real world exceptions and events?
“Pint:” most of the phonemes are regular
JFK assassination example: most of what we know about an event is drawn from existing associations in memory

35. The role of the hippocampus
Encodes patterns in sparse, non-overlapping manner
Trains cortex by reinstating this pattern repeatedly over time.

36. A simple model:
In this model, hippocampal & cortical representations are same
Consolidation rate “C” determines rate at which hippocampal units feed patterns to cortical units

37. The role of the hippocampus
Sparse coding allows a high learning rate, since learning one pattern won’t interfere with weights for another pattern

38. A problem with sparse coding:
There are many fewer cells in the hippocampus than in the neocortex
How can the hippocampus encode patterns in a way that is both sparse and compressed?

39. Sparse coding and hidden unit activity:
Coarse coding
Sparse coding
Really sparse coding

40. Really, really sparse coding:

41. Random, conjunctive coding
k-winners-take-all:
the weights between the k most active units and a randomly selected hidden unit are increased

42. Implementing this in a model of the hippocampal system:
Assume that cortex employs componential coding -- efficient, good for generalization:
Letters from 9x9 pixel array can be encoded with 13 feature units, instead of 81 pixel units

43. Much overlap between patterns
With componetial coding, about 34% of the 13 input units will be active to represent a given letter
With the 9x9=81, 1 pixel = 1 hidden unit system, 28% of input units active
Much overlap between patterns
This is good for generalization
Bad for pattern separation

44. Really, really sparse coding:
Have every possible combination of 3 input units correspond to 1 hidden unit
For an input consisting of 5 pixel letters, over 10 million such triples!
A 5 letter input would activate about 230,000 of these triples
Only 2% of possible input triples active – much more sparse

46. Sparse hippocampal representation
“Compressed” represention from cortex
Cortical representation

47. How this model maps onto anatomy:

48. Plausibility of this encoding scheme:
Many more cells in Dental Gyrus (hippocampus proper) than in Entorhinal Cortex
Autoassocitor (via recurrent collaterals) and Hebbian pattern associator also present in basic hippocampal circuit

49. In short:
Representations in cortex compressed via connections with Entorhinal Cortex
These coarse, componential representations are made sparse (random, conjunctive coding) via connections with hippocampus proper
First compress, then sparsify

50. Problems with these models:
Does the hippocampus really encode all kinds of arbitrary associations, or just spatial maps?
Cortical learning implies that only a prototype stored in memory -- no information about individual training events in cortex

51. In summary: Important pairs of concepts:
Interleaved/Focused (training)
Slow/Fast (learning rates)
Coarse/Sparse (representations)