Connectionism and models of memory and amnesia

1. Connectionism and models of memory and amnesia Jaap Murre University of Amsterdam jaap@murre.com http://www.neuromod.org/courses/kernthema2004

2. The French neurologist Ribot discovered more than 100 years ago that in retrograde amnesia one tends to loose recent memoriesMemory loss gradients in RA are called Ribot gradients

3. Overview • Catastrophic interference and hypertransfer • Brief review of neuroanatomy • Outline of the TraceLink model • Some simulation results of neural network model, focussing on retrograde amnesia • Recent work: • Mathematical point-process model • Detailed, more biological, neural network model • Concluding remarks

4. Catastrophic interference • Learning new patterns in backpropation will overwrite all existing patterns • Rehearsal is necessary • McCloskey and Cohen (1989), Ratcliff (1990) • This is not psychologically plausible

5. Osgood surface (1949) • Paired-associates in lists A and B will interfere strongly if the stimuli are similar but the responses vary • If stimuli are different, little interference (i.e., forgetting) occurs • Backpropagation also shows odd behavior if stimuli vary but responses are similar in lists A and B (hypertransfer)

6. Hypertransfer Learned responses Stimuli Target responses (after three learning trials) Phase 1: Learning list A rist munk twup gork gomp toup wemp twub twup Phase 2: Learning interfering list B (after five learning trials) yupe munk muup maws gomp twup drin twub twub Phase 3: Retesting on list A rist munk goub gork gomp tomp wemp twub twub

7. Problems with sequential learning in backpropagation • Reason 1: Strongly overlapping hidden-layer representations • Remedy 1: reduce the hidden-layer representations • French, Murre (semi-distributed representations)

8. Problems with sequential learning in backpropagation • Reason 2: Satisfying only immediate learning constraints • Remedy 2: Rehearse some old patterns, when learning new ones • Murre (1992): random rehearsal • McClelland, McNaughton and O’Reilly (1995): interleaved learning

9. Final remarks on sequential learning • Two-layer ‘backpropagation’ networks do show plausible forgetting • Other learning networks do not exhibit catastrophic interference: ART, CALM, Kohonen Maps, etc. • It is not a necessary condition of learning neural networks; it mainly affects backpropagation • The brain does not do backpropagation and therefore does not suffer from this problem

10. Models of amnesia and memory in the brain • TraceLink • Point-process model • Chain-development model

11. Neuroanatomy of amnesia • Hippocampus • Adjacent areas such as entorhinal cortex and parahippocampal cortex • Basal forebrain nuclei • Diencephalon

12. The position of the hippocampus in the brain

13. Hippocampal connections

14. Hippocampus has an excellent overview of the entire cortex

15. Trace-Link model: structure

16. System 1: Trace system • Function: Substrate for bulk storage of memories, ‘association machine’ • Corresponds roughly to neocortex

17. System 2: Link system • Function: Initial ‘scaffold’ for episodes • Corresponds roughly to hippocampus and certain temporal and perhaps frontal areas

18. System 3: Modulatory system • Function: Control of plasticity • Involves at least parts of the hippocampus, amygdala, fornix, and certain nuclei in the basal forebrain and in the brain stem

19. Stages in episodic learning

20. Dreaming and consolidation of memory “We dream in order to forget” • Theory by Francis Crick and Graeme Mitchison (1983) • Main problem: Overloading of memory • Solution: Reverse learning leads to removal of ‘obsessions’

21. Dreaming and memory consolidation • When should this reverse learning take place? • During REM sleep • Normal input is deactivated • Semi-random activations from the brain stem • REM sleep may have lively hallucinations

22. Consolidation may also strengthen memory • This may occur during deep sleep (as opposed to REM sleep) • Both hypothetical processes may work together to achieve an increase in the definition of representations in the cortex

23. Recent data by Matt Wilson and Bruce McNaughton (1994) • 120 neurons in rat hippocampus • PRE: Slow-wave sleep before being in the experimental environment (cage) • RUN: During experimental environment • POST: Slow-wave sleep after having been in the experimental environment

24. Wilson en McNaughton Data • PRE: Slow-wave sleep before being in the experimental environment (cage) • RUN: During experimental environment • POST: Slow-wave sleep after having been in the experimental environment

25. Some important characteristics of amnesia • Anterograde amnesia (AA) • Implicit memory preserved • Retrograde amnesia (RA) • Ribot gradients • Pattern of correlations between AA and RA • No perfect correlation between AA and RA

26. Normal forgetting anterograde amnesia retrograde amnesia x past lesion present

27. An example of retrograde amnesia patient data Kopelman (1989) News events test

28. Retrograde amnesia • Primary cause: loss of links • Ribot gradients • Shrinkage

29. Anterograde amnesia • Primary cause: loss of modulatory system • Secondary cause: loss of links • Preserved implicit memory

30. Connectionist implementationof the TraceLink model With Martijn Meeter from the University of Amsterdam

31. Some details of the model • 42 link nodes, 200 trace nodes • for each pattern • 7 nodes are active in the link system • 10 nodes in the trace system • Trace system has lower learning rate that the link system

32. How the simulations work: One simulated ‘day’ • A new pattern is activated • The pattern is learned • Because of low learning rate, the pattern is not well encoded at first in the trace system • A period of ‘simulated dreaming’ follows • Nodes are activated randomly by the model • This random activity causes recall of a pattern • A recalled pattern is than learned extra

33. (Patient data) Kopelman (1989) News events test

34. A simulation with TraceLink

35. Frequency of consolidation of patterns over time

36. Strongly and weakly encoded patterns • Mixture of weak, middle and strong patterns • Strong patterns had a higher learning parameter (cf. longer learning time)

38. Other simulations • Focal retrograde amnesia • Levels of processing • Transient Global Amnesia (TGA) • Semantic dementia • Implicit memory • More subtle lesions (e.g., only within-link connections, cf. CA1 lesions)

39. The Memory Chain Model: a very abstract neural network With Antonio Chessa from the University of Amsterdam

40. Abstracting TraceLink (level 1) • Model formulated within the mathematical framework of point processes • Generalizes TraceLink’s two-store approach to multiple neural ‘stores’ • trace system • link system • working memory, short-term memory, etc. • A store corresponds to a neural process or structure

41. Learning and forgetting as a stochastic process: 1-store example • A recall cue (e.g., a face) may access different aspects of a stored memory • If a point is found in the neural cue area, the correct response (e.g., the name) can be given Forgetting Successful Recall Unsuccessful Recall Learning

42. Jo Brand Neural network interpretation

43. m a Link system Retrieval Survival probability Single-store point process • The expected number of points in the cue area after learning is called  • This  is directly increased by learning and also by more effective cueing • At each time step, points die • The probability of survival of a point is denoted by a

44. Some aspects of the point process model • Model of simultaneous learning and forgetting • Clear relationship between signal detection theory (d'), recall (p), savings (Ebbinghaus’ Q), and Crovitz-type distribution functions • Multi-trial learning and multi-trial savings • Currently applied to over 250 experiments in learning and forgetting, since 1885

45. Forgetting curve If we need to find at least one point we obtain the following curve (one-store case): m is the intensity of the process (expected number of points) and a is the decay parameter We predict a flex point when the initial recall is at least

46. Example: Single-store model fitted to short-term forgetting data R2 = 0,985

47. Multi-store generalization • Information about the current event passes through many neural ‘stores’ • The retina, for example, holds a lot of information very briefly • The cerebral cortex holds very little information (of the current event) for a very long time

48. General principles of the PPM multi-store model • A small part of the information is passed to the next store before it decays completely • Subsequent stores hold information for longer time periods: slower decay rates in ‘higher’ stores

49. Two-store model • While neural store 1 is decaying (with rate a1) it induces new points (representations) in store 2 • Induction rate is linear with the intensity in store 1 and has induction rate m2 • The points in store immediately start to decay as well (at a lower rate a2)

50. Example of two neural stores • Store 1: firing neural groups • Store 2: synaptic connections between the neural groups • Other interpretation are possible as well, e.g.: • Store 1: hippocampus • Store 2: cerebral cortex Skip