1. CLS, Rapid Schema Consistent Learning, and Similarity-weighted Interleaved learning
Psychology 209 Feb 26, 2019

2. Your knowledge is in your connections!
An experience is a pattern of activation over neurons in one or more brain regions.
The trace left in memory is the set of adjustments to the strengths of the connections.
Each experience leaves such a trace, but the traces are not separable or distinct.
Rather, they are superimposed in the same set of connection weights.
Recall involves the recreation of a pattern of activation, using a part or associate of it as a cue.
The reinstatement depends on the knowledge in the connection weights, which in general will reflect influences of many different experiences.
Thus, memory is always a constructive process, dependent on contributions from many different experiences.

3. Effect of a Hippocampal Lesion
Intact performance on tests of intelligence, general knowledge, language, other acquired skills
Dramatic deficits in formation of some types of new memories:
Explicit memories for episodes and events
Paired associate learning
Arbitrary new factual information
Spared priming and skill acquisition
Temporally graded retrograde amnesia:
lesion impairs recent memories leaving remote memories intact.
Note: HM’s lesion was bilateral

5. Key Points
We learn about the general pattern of experiences, not just specific things
Gradual learning in the cortex builds implicit semantic and procedural knowledge that forms much of the basis of our cognitive abilities
The Hippocampal system complements the cortex by allowing us to learn specific things without interference with existing structured knowledge
In general these systems must be thought of as working together rather than being alternative sources of information.
Much of behavior and cognition depends on both specific and general knowledge

6. Emergence of Meaning in Learned Distributed Representations through Gradual Interleaved Learning
Distributed representations (what ML calls embeddings) that capture aspects of meaning emerge through a gradual learning process
The progression of learning and the representations formed capture many aspects of cognitive development
Progressive differentiation
Sensitivity to coherent covariation across contexts
Reorganization of conceptual knowledge

8. The Rumelhart Model

9. The Training Data:
All propositions true of items at the bottom level of the tree, e.g.:
Robin can {grow, move, fly}

11. Early
Later
Later Still
E x p e r i e
n c e

12. What happens in this system if we try to learn something new?
Such as a Penguin

13. Learning Something New
Used network already trained with eight items and their properties.
Added one new input unit fully connected to the representation layer
Trained the network with the following pairs of items:
penguin-isa living thing-animal-bird
penguin-can grow-move-swim

14. Rapid Learning Leads to Catastrophic Interference

16. Avoiding Catastrophic Interference with Interleaved Learning

20. Rapid Consolidation of Schema Consistent Information
Richard Morris
Rapid Consolidation of Schema Consistent Information

21. Tse et al (Science, 2007, 2011) During training, 2 wells
uncovered on each trial

22. Schemata and Schema Consistent Information
What is a ‘schema’?
An organized knowledge structure into which existing knowledge is organized.
What is schema consistent information?
Information that can be added to a schema without disturbing it.
What about a penguin?
Partially consistent
Partially inconsistent
In contrast, consider
a trout
a cardinal

23. New Simulations
Initial training with eight items and their properties as before.
Added one new input unit fully connected to the representation layer also as before
Trained the network on one of the following pairs of items:
penguin-isa & penguin-can
trout-isa & trout-can
cardinal-isa & cardinal-can

24. New Learning of Consistent and Partially Inconsistent Information
INTERFERENCE

25. Connection Weight Changes after Simulated NPA, OPA and NM Analogs
Tse et al. 2011

26. How Does It Work?

27. How Does It Work?

28. Remaining Questions
Are all aspects of new learning integrated into cortex-like networks at the same rate?
No, some aspects are integrated much more slowly than others
Is it possible to avoid replaying everything one already knows when one wants to learn new things with arbitrary structure?
Yes, at least in some circumstances that we will explore
Perhaps the answers to these questions will allow us to make more efficient use of both cortical and hippocampal resources for learning.

29. Toward an Explicit Mathematical Theory of Interleaved Learning
Characterizing structure in a dataset to be learned
The deep linear network that can learn this structure
The dynamics of learning the structure in the dataset
Initial learning of a base data set
Subsequent learning of an additional item
Using similarity weighted interleaved learning to increase efficiency of interleaved learning
Initial thoughts on how we might use the hippocampus more efficiently

30. Hierarchical structure in a synthetic data set
Sparrow
Hawk
Salmon
Sunfish
Oak
Maple
Rose
Daisy
SparrowHawk

31. Processing and Learning in a deep linear network
Saxe et al, (2013a,b,…)

32. SVD of the dataset

33. Dynamics of Learning – one-hot inputs
SSE
a(t)
Variable discrepancy affects
takeoff point, but not shape
Solid lines are simulated values of a(t) Dashed lines are based on the equation

34. Dynamics of Learning – auto-associator
SSE
𝑠 2
a(t)
𝑠 2
Dynamics are a bit more predictable
Solid lines are simulated values of a(t) Dashed lines are based on the equation

35. Adding a new member of an existing category
Sparrow
Hawk
Salmon
Sunfish
Oak
Maple
Rose
Daisy
SparrowHawk

36. SVD of New Complete Dataset

37. The consequence of standard interleaved learning

38. SVD Analysis of Network Output for Birds
Adjusted Dimensions New Dimension

39. Similarity Weighted Interleaved Learning
Full Interleaving
Similarity- weighted Interleaving
Uniform Interleaving

40. Freezing the output weights initially
Full Interleaving
Similarity- weighted Interleaving
Uniform Interleaving

41. Discussion
Integration of fine-grained structure into a deep network may always be a slow process
Sometimes this fine-grained structure is ultimately fairly arbitrary and idiosyncratic, although other times it may be part of a deeper pattern the learner has not previously seen
One way to address such integration:
Initial reliance on sparse / item-specific representation
This could be made more efficient by storing only the ‘correction vector’ in the hippocampus
Gradual integration through interleaved learning

42. SparrowHawk Error Vector After Easy Integration Phase is Complete

43. Questions, answers and next steps
Are all aspects of new learning integrated into cortex-like networks at the same rate?
No, some aspects are integrated much more slowly than others
Is it possible to avoid replaying everything one already knows when one wants to learn new things with arbitrary structure?
Yes, at least in some circumstances that we have explored
Perhaps the answers to these questions will allow us to make more efficient use of both cortical and hippocampal resources for learning.