neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (04/11/91)
Neuron Digest Wednesday, 10 Apr 1991 Volume 7 : Issue 19 Today's Topics: TR - On planning and exploartion in non-discrete worlds Paper announcements NN application in molecular biology TR - Segmentation, Binding, and illusory conjunctions preprint of paper on visual binding Two papers on information transfer / problem decomposition Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" Use "ftp" to get old issues from hplpm.hpl.hp.com (15.255.176.205). ------------------------------------------------------------ Subject: TR - On planning and exploartion in non-discrete worlds From: Sebastian Thrun <thrun@gmdzi.uucp> Date: Tue, 19 Mar 91 03:39:54 -0100 Well, there is a new TR available on the neuroprose archieve which is more or less an extended version of the NIPS paper I announced some weeks ago: ON PLANNING AND EXPLORATION IN NON-DISCRETE WORLDS Sebastian Thrun Knut Moeller German National Research Center Bonn University for Computer Science St. Augustin, FRG Bonn, FRG The application of reinforcement learning to control problems has received considerable attention in the last few years [Anderson86,Barto89,Sutton84]. In general there are two principles to solve reinforcement learning problems: direct and indirect techniques, both having their advantages and disadvantages. We present a system that combines both methods. By interaction with an unknown environment a world model is progressively constructed using the backpropagation algorithm. For optimizing actions with respect to future reinforcement planning is applied in two steps: An experience network proposes a plan, which is subsequently optimized by gradient descent with a chain of model networks. While operating in a goal-oriented manner due to the planning process the experience network is trained. Its accumulating experience is fed back into the planning process in form of initial plans, such that planning can be gradually reduced. In order to ensure complete system identification, a competence network is trained to predict the accuracy of the model. This network enables purposeful exploration of the world. The appropriateness of this approach to reinforcement learning is demonstrated by three different control experiments, namely a target tracking, a robotics and a pole balancing task. Keywords: backpropagation, connectionist networks, control, exploration, planning, pole balancing, reinforcement learning, robotics, neural networks, and, and, and... =------------------------------------------------------------------------- The TR can be retrieved by ftp: unix> ftp cheops.cis.ohio-state.edu Name: anonymous Guest Login ok, send ident as password Password: neuron ftp> binary ftp> cd pub ftp> cd neuroprose ftp> get thrun.plan-explor.ps.Z ftp> bye unix> uncompress thrun.plan-explor.ps unix> lpr thrun.plan-explor.ps = ------------------------------------------------------------------------- If you have trouble in ftping the files, do not hesitate to contact me. --- Sebastian Thrun (st@gmdzi.uucp, st@gmdzi.gmd.de) ------------------------------ Subject: Paper announcements From: thanasis kehagias <ST401843@brownvm.brown.edu> Date: Tue, 19 Mar 91 18:29:54 -0500 I have just placed two papers of mine in the ohio-state archive. The first one is in the file kehagias.srn1.ps.Z and the relevant figures in the companion file kehagias.srn1fig.ps.Z. The second one is in the file kehagias.srn2.ps.Z and the relevant figures in the companion file kehagias.srn2fig.ps.Z. Detailed instructions for getting and printing these files will be included in the end of this message. Some of you have received versions of these files in email previously. In that case read a postscript at the end of this message. =----------------------------------------------------------------------- Stochastic Recurrent Network training by the Local Backward-Forward Algorithm Ath. Kehagias Brown University Div. of Applied Mathematics We introduce Stochastic Recurrent Networks, which are collections of interconnected finite state units. At every discrete time step, each unit goes into a new state, following a probability law that is conditional on the state of neighboring units at the previous time step. A network of this type can learn a stochastic process, where ``learning'' means maximizing the probability Likelihood function of the model. A new learning (i.e. Likelihood maximization) algorithm is introduced, the Local Backward-Forward Algorithm. The new algorithm is based on the Baum Backward-Forward Algorithm (for Hidden Markov Models) and improves speed of learning substantially. Essentially, the local Backward-Forward Algorithm is a version of Baum's algorithm which estimates local transition probabilities rather than the global transition probability matrix. Using the local BF algorithm, we train SRN's that solve the 8-3-8 encoder problem and the phoneme modelling problem. This is the paper kehagias.srn1.ps.Z, kehagias.srn1fig.ps.Z . The paper srn1 has undergone significant revision. It had too many typos, bad notation and also needed reorganization . All of these have been done. Thanks to N. Chater, S. Nowlan and A.T. Tsoi and M. Perrone for many useful suggestions along these lines. =-------------------------------------------------------------------- Stochastic Recurrent Network training Prediction and Classification of Time Series Ath. Kehagias Brown University Div. of Applied Mathematics We use Stochastic Recurrent Networks of the type introduced in [Keh91a] as models of finite-alphabet time series. We develop the Maximum Likelihood Prediction Algorithm and the Maximum A Posteriori Classification Algorithm (which can both be implemented in recurrent PDP form). The prediction problem is: given the output up to the present time: Y^1,...,Y^t and the input up to the immediate future: U^1,...,U^t+1, predict with Maximum Likelihood the output Y^t+1 that the SRN will produce in the immediate future. The classification problem is: given the output up to the present time: Y^1,...,Y^t and the input up to the present time: U^1,...,U^t, as well as a number of candidate SRN's: M_1, M_2, .., M_K, find the network that has Maximum Posterior Probability of producing Y^1,...,Y^t. We apply our algorithms to prediction and classification of speech waveforms. This is the paper kehagias.srn2.ps.Z, kehagias.srn2fig.ps.Z . =----------------------------------------------------------------------- To get these files, do the following: gvax> ftp cheops.cis.ohio-state.edu 220 cheops.cis.ohio-state.edu FTP server ready. Name: anonymous 331 Guest login ok, send ident as password. Password:neuron ftp> Guest login ok, access restrictions apply. ftp> cd pub/neuroprose ftp> binary 200 Type set to I. ftp> get kehagias.srn1.ps.Z ftp> get kehagias.srn1fig.ps.Z ftp> get kehagias.srn2.ps.Z ftp> get kehagias.srn2fig.ps.Z ftp> quit gvax> uncompress kehagias.srn1.ps.Z gvax> uncompress kehagias.srn1fig.ps.Z gvax> uncompress kehagias.srn2.ps.Z gvax> uncompress kehagias.srn2fig.ps.Z gvax> lqp kehagias.srn1.ps gvax> lqp kehagias.srn1fig.ps gvax> lqp kehagias.srn2.ps gvax> lqp kehagias.srn2fig.ps POSTSCRIPT: All of the people that sent a request (about a month ago) for srn1 in its original form are in my mailing list and most got copies of new versions of srn1,srn2 in email. Some of these files did not make it through internet, because of size restrictions etc. so you may want to fpt them now. Incidentally, if you want to be removed from the mailing list (for when the next paper in the series comes by) send me mail. Thanasis Kehagias ------------------------------ Subject: NN application in molecular biology From: BRUNAK@nbivax.nbi.dk Date: Wed, 20 Mar 91 13:06:00 +0100 The following preprint is available in postscript form by anonymous ftp "Prediction of human mRNA donor and acceptor sites from the DNA sequence". S. Brunak, J. Engelbrecht and S. Knudsen. Journal of Molecular Biology, to appear. Abstract: Artificial neural networks have been applied to the prediction of splice site location in human pre-mRNA. A joint prediction scheme where prediction of transition regions between introns and exons regulates a cutoff level for splice site assignment was able to predict splice site locations with confidence levels far better than previously reported in the literature. The problem of predicting donor and acceptor sites in human genes is hampered by the presence of numerous amounts of false positives --- in the paper the distribution of these false splice sites is examined and linked to a possible scenario for the splicing mechanism in vivo. When the presented method detects 95% of the true donor and acceptor sites it makes less than 0.1% false donor site assignments and less than 0.4% false acceptor site assignments. For the large data set used in this study this means that on the average there are one and a half false donor sites per true donor site and six false acceptor sites per true acceptor site. With the joint assignment method more than a fifth of the true donor sites and around one fourth of the true acceptor sites could be detected without accompaniment of any false positive predictions. Highly confident splice sites could not be isolated with a widely used weight matrix method or by separate splice site networks. A complementary relation between the confidence levels of the coding/non-coding and the separate splice site networks was observed, with many weak splice sites having sharp transitions in the coding/non-coding signal and many stronger splice sites having more ill-defined transitions between coding and non-coding. Subject category: Genes, under the sub--headings: expression, sequence and structure. Keywords: Intron--splicing, human genes, exon selection, neural network, computer--prediction. =----------------------------------------------------------------------- You will need a POSTSCRIPT printer to print the file. To obtain a copy of the preprint, use anonymous ftp from cheops.cis.ohio-state.edu (here is what the transaction looks like): unix> ftp ftp> open cheops.cis.ohio-state.edu Connected to cheops.cis.ohio-state.edu. 220 cheops.cis.ohio-state.edu FTP server (Version blah blah) ready. Name (cheops.cis.ohio-state.edu:yourname): anonymous 331 Guest login ok, send ident as password. Password: anything 230 Guest login ok, access restrictions apply. ftp> cd pub/neuroprose 250 CWD command successful. ftp> bin 200 Type set to I. ftp> get brunak.netgene.ps.Z 200 PORT command successful. 150 Opening BINARY mode data connection for brunak.netgene.ps.Z 226 Transfer complete. local: brunak.netgene.ps.Z remote: brunak.netgene.ps.Z ftp> quit 221 Goodbye. unix> uncompress brunak.netgene.ps.Z unix> lpr brunak.netgene.ps Hardcopies are also available: S. Brunak and J. Engelbrecht Department of Structural Properties of Materials Building 307 The Technical University of Denmark DK-2800 Lyngby, Denmark brunak@nbivax.nbi.dk ------------------------------ Subject: TR - Segmentation, Binding, and illusory conjunctions From: HORN%TAUNIVM.BITNET@bitnet.CC.CMU.EDU Date: Thu, 21 Mar 91 18:29:54 -0500 The following preprint is available. Requests can be sent to HORN at TAUNIVM.BITNET : Segmentation, Binding and Illusory Conjunctions D. HORN, D. SAGI and M. USHER Abstract We investigate binding within the framework of a model of excitatory and inhibitory cell assemblies which form an oscillating neural network. Our model is composed of two such networks which are connected through their inhibitory neurons. The excitatory cell assemblies represent memory patterns. The latter have different meanings in the two networks, representing two different attributes of an object, such as shape and color. The networks segment an input which contains mixtures of such pairs into staggered oscillations of the relevant activities. Moreover, the phases of the oscillating activities representing the two attributes in each pair lock with each other to demonstrate binding. The system works very well for two inputs, but displays faulty correlations when the number of objects is larger than two. In other words, the network conjoins attributes of different objects, thus showing the phenomenon of "illusory conjunctions", like in human vision. ------------------------------ Subject: preprint of paper on visual binding From: "Erik D. Lumer" <lumer@parc.xerox.com> Date: Fri, 05 Apr 91 10:27:48 -0800 The following paper is available in hardcopy form. If interested, send e-mail requests to "lumer@parc.xerox.com" =------------------------------------------------------------ "Binding Hierarchies: A Basis for Dynamic Perceptual Grouping" Erik Lumer and Bernado Huberman Dynamics of Computation Group, Xerox Palo Alto Research Center and Stanford University (submitted to Neural Computation) ----------------------------------- Since it has been argued that the brain binds its fragmentary representations of perceptual events via phase-locking of stimulated neuron oscillators, it is important to determine how extended synchronization can occur in a clustered organization of cells posessing an intrisic distribution of firing rates. In order to answer that that question, we establish the basic conditions for the existence of a binding mechanism based on phase-locked oscillations. In addition, we present a simple hierarchical architecture of feedback units which not only induces robust synchronization and segregation of perceptual groups, but serves as a generic binding machine. ------------------------------ Subject: Two papers on information transfer / problem decomposition From: "Lorien Y. Pratt" <pratt@paul.rutgers.edu> Date: Fri, 05 Apr 91 17:08:52 -0500 The following two papers are now available via FTP from the neuroprose archives. The first is for AAAI91, so written towards an AI/Machine learning audience. The second is for IJCNN91, so more neural network-oriented. There is some overlap between them: the AAAI paper reports briefly on the study describved in more detail in the IJCNN paper. Instructions for retrieval are at the end of this message. --Lori #@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@#@ Direct Transfer of Learned Information Among Neural Networks To appear: Proceedings of AAAI-91 Lorien Y. Pratt and Jack Mostow and Candace A. Kamm Abstract A touted advantage of symbolic representations is the ease of transferring learned information from one intelligent agent to another. This paper investigates an analogous problem: how to use information from one neural network to help a second network learn a related task. Rather than translate such information into symbolic form (in which it may not be readily expressible), we investigate the direct transfer of information encoded as weights. Here, we focus on how transfer can be used to address the important problem of improving neural network learning speed. First we present an exploratory study of the somewhat surprising effects of pre-setting network weights on subsequent learning. Guided by hypotheses from this study, we sped up back-propagation learning for two speech recognition tasks. By transferring weights from smaller networks trained on subtasks, we achieved speedups of up to an order of magnitude compared with training starting with random weights, even taking into account the time to train the smaller networks. We include results on how transfer scales to a large phoneme recognition problem. @%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@%@ Improving a Phoneme Classification Neural Network through Problem Decomposition To appear: Proceedings of IJCNN-91 L. Y. Pratt and C. A. Kamm Abstract In the study of neural networks, it is important to determine whether techniques that have been validated on smaller experimental tasks can be scaled to larger real-world problems. In this paper we discuss how a methodology called {\em problem decomposition} can be applied to AP-net, a neural network for mapping acoustic spectra to phoneme classes. The network's task is to recognize phonemes from a large corpus of multiple-speaker, continuously-spoken sentences. We review previous AP-net systems and present results from a decomposition study in which smaller networks trained to recognize subsets of phonemes are combined into a larger network for the full signal-to-phoneme mapping task. We show that, by using this problem decomposition methodology, comparable performance can be obtained in significantly fewer arithmetic operations. %^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^%^% To retrieve: unix> ftp cheops.cis.ohio-state.edu (or 128.146.8.62) Name: anonymous Password: neuron ftp> cd pub/neuroprose ftp> binary ftp> get pratt.aaai91.ps.Z ftp> get pratt.ijcnn91.ps.Z ftp> quit unix> uncompress pratt.aaai91.ps.Z pratt.ijcnn91.ps.Z unix> lpr pratt.aaai91.ps pratt.ijcnn91.ps ------------------------------ End of Neuron Digest [Volume 7 Issue 19] ****************************************