dynamical system s beha vior in recurren t and non
play

Dynamical-System s Beha vior in Recurren t and Non-Recurren - PDF document

Dynamical-System s Beha vior in Recurren t and Non-Recurren t Connectionist Nets An Honors Thesis Presen ted b y Jason M. Eisner to The Departmen t of Psyc hology in P artial F ulllmen t of the Requiremen ts


  1. Dynamical-System s Beha vior in Recurren t and Non-Recurren t Connectionist Nets An Honors Thesis Presen ted b y Jason M. Eisner to The Departmen t of Psyc hology in P artial F ul�llmen t of the Requiremen ts for the Degree of Bac helor of Arts with Honors in the Sub ject of Psyc holgy Harv ard-Radcli�e Colleges Cam bridge, Massac h usetts April 2, 1990

  2. Abstract A broad approac h is dev elop ed for training dynamical b eha viors in connectionist net w orks. General recurren t net w orks are p o w erful computational devices, necessary for di�cult tasks lik e constrain t sat- isfaction and temp oral pro cessing. These tasks are discussed here in some detail. F rom b oth theoretical and empirical considerations, it is concluded that suc h tasks are b est addressed b y recurren t net w orks that op erate con tin uously in time|and further, that e�ectiv e learn- ing rules for these con tin uous-time net w orks m ust b e able to prescrib e their dynamic al prop erties. A general class of suc h learning rules is deriv ed and tested on simple problems. Where existing learning algo- rithms for recurren t and non-recurren t net w orks only attempt to train a net w ork's p osition in activ ation space, the mo dels presen ted here can also explicitly and successfully prescrib e the nature of its movement thr ough activ ation space. I am indebted to Ja y Ruec kl, m y advisor, b oth for his suggestions and for his supp ort. Ja y Ruec kl and Greg Galp erin pro vided com- putational facilities that pro v ed indisp ensable. I w ould also lik e to thank m y family and the man y friends whose encouragemen t sa w this pro ject through its �nal stages. Con ten ts 1 In tro duction 3 1.1 The uses of recurren t net w orks : : : : : : : : : : : : : : : : : : 3 1.1.1 Recurren t net w orks and constrain t satisfaction : : : : : 3 1.1.2 Recurren t net w orks and temp oral problems : : : : : : : 4 1.1.3 The computational p o w er of recurren t net w orks : : : : 5 1.2 Recurren t net w orks in practice : : : : : : : : : : : : : : : : : : 6 1.2.1 Existing mo dels of constrain t satisfaction : : : : : : : : 6 1.2.2 Existing temp oral mo dels : : : : : : : : : : : : : : : : 8 1.3 Summary : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 10 2 Some Theoretical Observ ations 11 2.1 Though ts on temp oral pattern pro cessing : : : : : : : : : : : : 12 1

  3. 2.1.1 The usefulness of gradual-resp onse nets : : : : : : : : : 12 2.1.2 The dynamical systems approac h : : : : : : : : : : : : 14 2.1.3 Dynamics of small clusters : : : : : : : : : : : : : : : : 16 2.2 Though ts on constrain t satisfaction : : : : : : : : : : : : : : : 19 2.2.1 The need for hidden units : : : : : : : : : : : : : : : : 19 2.2.2 A �rst attempt at a satisfaction mo del : : : : : : : : : 20 2.2.3 A learning rule for the non-resonan t case : : : : : : : : 21 2.2.4 Extending our rule to the resonan t case : : : : : : : : : 23 2.2.5 The problem with this approac h : : : : : : : : : : : : : 25 2.3 Summary of Theoretical Observ ations : : : : : : : : : : : : : : 26 3 A General Mo del and its Learning Rule 26 3.1 Conception of the general mo del : : : : : : : : : : : : : : : : : 26 3.1.1 Molding a dynamical system : : : : : : : : : : : : : : : 26 3.1.2 The role of input : : : : : : : : : : : : : : : : : : : : : 27 3.1.3 Ho w to use the error measure : : : : : : : : : : : : : : 28 3.1.4 Ho w w eigh t c hanges shift the tra jectory : : : : : : : : 31 3.2 F ormal deriv ation of the general mo del : : : : : : : : : : : : : 32 3.2.1 Notation : : : : : : : : : : : : : : : : : : : : : : : : : : 33 3.2.2 Calculating the gradien t in w eigh t space : : : : : : : : 34 3.2.3 An algorithm : : : : : : : : : : : : : : : : : : : : : : : 35 3.3 Summary of the general mo del : : : : : : : : : : : : : : : : : : 37 4 P articular Mo dels 38 4.1 Some mo dels of p oten tial in terest : : : : : : : : : : : : : : : : 38 4.2 Some top ologies of p oten tial in terest : : : : : : : : : : : : : : 41 4.3 Detailed deriv ation of particular error measures : : : : : : : : 42 4.3.1 Mapping mo del I: No des to w ard targets : : : : : : : : 42 4.3.2 Mapping mo del I I: System to w ard target : : : : : : : : 44 4.3.3 General gradien t-descen t mo del : : : : : : : : : : : : : 45 4.3.4 Con ten t-addressable memory mo del : : : : : : : : : : : 46 5 Sim ulation Results 49 5.1 Results for feedforw ard X OR : : : : : : : : : : : : : : : : : : : 51 5.2 Other tasks : : : : : : : : : : : : : : : : : : : : : : : : : : : : 53 6 Conclusions 54 2

  4. 1 In tro duction In the 1960's, Minsky and P ap ert p oin ted to hidden units as a p oten tial remedy for some of connectionism's problems. Recurren t connections ha v e lately b een attracting the same kind of in terest. Muc h as hidden units extend the computational p o w er of p erceptrons, recurren t connections extend the computational p o w er of feedforw ard net w orks. The w ork rep orted here is ultimately concerned with b oth recurren t and non-recurren t net w orks. Ho w ev er, it fo cuses on net w ork prop erties that are most eviden t (and most useful) in the presence of recurrence. These are dynamic al prop erties of net w orks|prop erties describing ho w net w orks' states c hange or remain stable o v er time. The pap er has three ma jor aims, as follo ws. First, to highligh t the fea- tures of recurrence that mak e it useful. Second, to demonstrate that certain net w ork arc hitectures exhibit esp ecially ric h kinds of b eha vior. Finally , to dev elop a training algorithm that can pro duce the desired b eha viors in net- w orks that use these arc hitectures. 1.1 The uses of recurren t net w orks Since it is useful to fo cus on actual problems, the early sections of this pa- p er will pa y sp ecial atten tion to t w o domains in whic h recurren t net w orks ha v e pro v ed esp ecially useful. These are the c onstr aint satisfaction domain and the temp or al domain. In a constrain t satisfaction problem, the net w ork is supp osed to disco v er an y regularities that hold among v arious static in- puts. The temp oral domain includes all those tasks where a net w ork's inputs and/or outputs are to c hange o v er time in a principled w a y . 1.1.1 Recurren t net w orks and constrain t satisfaction The general constrain t satisfaction task is simple. V arious p atterns (v ectors of n um b ers) are sho wn to a net w ork. The net w ork is supp osed to disco v er regularities in the set of patterns it sees. When it is sho wn only part of a pattern, it should correctly �ll in the missing elemen ts. The purest form of constrain t satisfaction mak es no distinction b et w een input and output no des of the net w ork. There is simply a set of visible no des , whic h hold the patterns. A partial pattern can b e \clamp ed" on to some of 3

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend