2. Bernard Ans, Stéphane Rousset, Robert M. French & Serban Musca (European Commission grant HPRN-CT-1999-00065) Preventing Catastrophic Interference in Multiple-Sequence Learning Using Coupled Reverberating Elman Networks

3. The Problem of Multiple-Sequence Learning Real cognition requires the ability to learn sequences of patterns (or actions). (This is why SRN’s – Elman Networks – were originally developed.) But learning sequences really means being able to learn multiple sequences without the most recently learned ones erasing the previously learned ones. Catastrophic interference is a serious problem for the sequential learning of individual patterns. It is far worse when multiple sequences of patterns have to be learned consecutively.

4. The Solution We have developed a “dual-network” system using coupled Elman networks that completely solves this problem. These two separate networks exchange information by means of “reverberated pseudopatterns.”

5. Pseudopatterns f(x) Assume a network-in-a-box learns a series of patterns produced by a function f(x). These original patterns are no longer available. How can you approximate f(x)?

6. 1 0 0 1 1Random Input

7. 1 0 0 1 1Random Input 1 1 0Associated output

8. 1 0 0 1 1Random Input 1 1 0Associated output This creates a pseudopattern:  1 : 1 0 0 1 1  1 1 0

9. A large enough collection of these pseudopatterns:  1 : 1 0 0 1 1  1 1 0  2 : 1 1 0 0 0  0 1 1  3 : 0 0 0 1 0  1 0 0  4 : 0 1 1 1 1  0 0 0 Etc will approximate the originally learned function.

10. Transferring information from Net 1 to Net 2 with pseudopatterns 1 0 0 1 1 1 1 0 1 0 0 1 1 1 1 0 target input Random input Associated output Net 1 Net 2

11. Information transfer by pseudopatterns in dual-network systems New information is presented to one network (Net 1). Pseudopatterns are generated by Net 2 where previously learned information is stored. Net 1 then trains not only on the new pattern(s) to be learned, but also on the pseudopatterns produced by Net 2. Once Net 1 has learned the new information, it generates (lots of) pseudopatterns that train Net 2 This is why we say that information is continually transferred between the two networks by means of pseudopatterns.

12. Are all pseudopatterns created equal? No. Even though the simple dual-network system (i.e., new learning in one network; long-term storage in the other) using simple pseudopatterns does eliminate catastrophic interference, we can do better using “reverberated” pseudopatterns.

13. Building a Network that uses “reverberated” pseudopatterns. Input layer Hidden layer Output layer Start with a standard backpropagation network

14. Input layer Hidden layer Output layer Add an autoassociator

15. A new pattern to be learned, P: Input  Target, will be learned as shown below. Input Target Input

16. What are “reverberated pseudopatterns” and how are they generated?

17. We start with a random input î 0, feed it through the network and collect the output on the autoassociative side of the network.. This output is fed back into the input layer (“reverberated”) and, again, the output on the autoassociative side is collected. This is done R times.

22. After R reverberations, we associate the reverberated input and the “target” output. This forms the reverberated pseudopattern:

23. This dual-network approach using reverberated pseudopattern information transfer between the two networks effectively overcomes catastrophic interference in multiple-pattern learning Net 2 Storage network Net 1 New-learning network

24. But what about multiple-sequence learning? Elman networks are designed to learn sequences of patterns. But they forget catastrophically when they attempt to learn multiple sequences. Can we generalize the dual-network, reverberated pseudopattern technique to dual Elman networks and eliminate catastrophic interference in multiple-sequence learning? Yes

25. Elman networks (a.k.a. Simple Recurrent Networks) Copy hidden unit activations from previous time-step Standard input S(t) Context H(t-1) Hidden H(t) S(t+1) Learning a sequence S(1), S(2), …, S(n).

26. A “Reverberated Simple Recurrent Network” (RSRN): an Elman network with an autoassociative part

27. RSRN technique for sequentially learning two sequences A(t) and B(t). Net 1 learns A(t) completely. Reverberated pseudopattern transfer to Net 2. Net 1 makes one weight-change pass through B(t). Net 2 generates a few “static” reverberated pseudopatterns Net 1 does one learning epoch on these pseudopatterns from Net 2. Continue until Net 1 has learned B(t). Test how well Net 1 has retained A(t).

28. Two sequences to be learned: A(0), A(1), … A(10) and B(0), B(1), … B(10) Net 1 Net 2 Net 1 learns (completely) sequence A(0), A(1), …, A(10)

29. Net 1 Net 2 010110100110010 1110010011010 Net 1 produces 10,000 pseudopatterns, : 010110100110010 1110010011010 010110100110010 Input 1110010011010 Teacher Transferring the learning to Net 2

30. Net 1 Net 2 feedforward 1110010011010 Teacher 010110100110010 Input Transferring the learning to Net 2

31. Net 1 Net 2 Backprop weight change For each of the 10,000 pseudopatterns produced by Net 1, Net 2 makes 1 FF-BP pass. 010110100110010 Input 1110010011010 Teacher Transferring the learning to Net 2

32. Learning B(0), B(1), … B(10) by NET 1 Net 1 Net 2 1. Net 1 does ONE learning epoch on sequence B(0), B(1), …, B(10) 2. Net 2 generates a few pseudopatterns  NET 2 3. Net 1 does one FF-BP pass on each  NET 2

33. Net 1 Net 2 1. Net 1 does ONE learning epoch on sequence B(0), B(1), …, B(10) 2. Net 2 generates a few pseudopatterns  NET 2 Learning B(0), B(1), … B(10) by NET 1 3. Net 1 does one FF-BP pass on each  NET 2 Continue until Net 1 has learned B(0), B(1), …, B(10)

34. Sequences chosen Twenty-two distinct random binary vectors of length 100 are created. Half of these vectors are used to produce the first ordered sequence of items, A, denoted by A(0), A(1), …, A(10). The remaining 11 vectors are used to create a second sequence of items, B, denoted by B(0), B(1), …, B(10). In order to introduce a degree of ambiguity into each sequence (so that a simple BP network would not be able to learn them), we modify each sequence so that A(8) = A(5) and B(5) = B(1).

35. Test method First, sequence A is completely learned by the network. Then sequence B is learned. During the course of learning, we monitor at regular intervals how much of sequence A has been forgotten by the network.

36. Normal Elman networks: Catastrophic forgetting (a): Learning of sequence B (after having previously learned sequence A). By 450 epochs (an epoch corresponds to one pass through the entire sequence), sequence B has been completely learned. (b): The number of incorrect units (out of 100) for each serial position of sequence A during learning of sequence B. After 450 epochs, the SRN has, for all intents and purposes, completely forgotten the previously learned sequence A

37. Dual-RSRN’s: Catastrophic forgetting is eliminated Recall performance for sequences B and A during learning of sequence B by a dual-network RSRN. (a): By 400 epochs, the second sequence B has been completely learned. (b): The previously learned sequence A shows virtually no forgetting. Catastrophic forgetting of the previously learned sequence A has been completely overcome.

38. Normal Elman Network: Massive forgetting % Error on Sequence A Dual RSRN: No forgetting of Sequence A Seq. B being learned

39. Cognitive/Neurobiological plausibility? The brain, somehow, does not forget catastrophically. Separating new learning from previously learned information seems necessary. McClelland, McNaughton, O’Reilly (1995) have suggested the hippocampal-neocortical separation may be Nature’s way of solving this problem. Pseudopattern transfer is not so far-fetched if we accept results that claim that neo-cortical memory consolidation, is due, at least in part, to REM sleep.

40. Conclusions The RSRN reverberating dual-network architecture (Ans & Rousset, 1997, 2000) can be generalized to sequential learning of multiple temporal sequences. When learning multiple sequences of patterns, interleaving simple reverberated input-output pseudopatterns, each of which reflect the entire previously learned sequence(s), reduces (or eliminates entirely) forgetting of the initially learned sequence(s).