Fer Biopys Hs%41 13:79. 98 European Biophysics Journal Springer-Verlag 1994 Cytoskeletal involvement in neuronal learning: a review Judith Dayhoff", Swart Hameroff*, Rafael Lahoz-Beltra’, Charles E, Swenberg * * nstinnte for Systeme Kesearch. University of Maryland, College Pack. MD 20742. USA, + Agianced Biotecinolass Laboraiary. Devsariment Anesinesiolngy. Lniversits of Arizona, College of Medicine, Tasso. AZ 85724 us Picaltad Ciencas Biologicas, Universidad Complutense. £28040 Madrid. Spain © Armed Fores Raudinbiology Research Insucute, Bothecda, MDD 20KKO-S145, USA Reosived: 21 May 1993 Abstract. This paper introduves the ideas of neural nei Works in the context of curreatly recognized cellular structures within neurons, Neural network models and Paradigms require adaptation of synapses tor learning to occur in the network. Some models of jearniny Paradigms require information to move from axon to dendrite. This motivated Us t0 exarmine the possibility of intracellular signaling t0 mediate such signals. The ey- toskeleton forms a substrate for inttaceliuiar signaling vig material transport and other putative mechanisms Furthermore, many experimental results suggest 2 link hetween the cytoskeleton and cognitive processing. In this paper we review reseatch on intracellular signaling in the context of reural network earning Key words: Neural network — Cytoskeleton adaptation Synaptic 1. Introduction A fundamental challenge of current research is to explain the possible mechanisms for learning and cogmtive pro- cessing it the brain. Such an eaplanation will entail con- cepis at multiple levels. the systems level (network of neu rors}, the neuron level processing and adapration of the neuron and is synapses}. and the molecular level isynuap- vic and intracellular proteins and trafficking. Results from systems modeling of neural networks show us effec- Vive and important aspects for the individual netirons oF Processing elements that enable the neural networks to Jearn to recognize paiterns, Biophysical mechanisms for these functional capabilities can then be examined at the Jevel of protein molecules Neural network iearning paradigms, « focus of receat attention. show how networks of neuron-like compa- APbreviatione: NIV, mice live resonance theory RCP. restricted coulomb et crotuhule assoctsted protsin, KO. sine onde Correspondence te. L Dash abyic. MTs. microtubules; ART, adap so MAB, ow Accepted in ee¥ised forrs 21 December 1993 emis can leur through adapting synaptic strengths Learning is done in response to patterns that stimulate the network, which Itams to classify chose patterns. Pow erful learning capabsiities have been discovered for neural network paradigms such as back propagation, counter propagation. siema-pi networks. time-delay neural net- works adaptive resonance theory (ART), and networks with competitive layers, Many of these paradigms coquise reciprocal communications between pairs of neuronctike components aad require information flow to go buck- wards through the network as welt as forwards. But real neurons have membrane mediated signaling that travels inthe “forwards” direction dendrite to axon terminus a8 post-syaapte potentials signal from dendrites to the spike initiation zone and action potentials travel down axons, away from the soma, Two-way communication could occur by netve impulses but world require pairs of neurons 10 synapse onto euch others dendrites, at sites already rezeiving synapses from other neurons, This reeu Jar and intricae serangement has not been observed. Ree ciprocal communication could mote plausibly he mediat ced by intracellular signals going in the reverse direction, with the neuronal cytoskeleton mediating such intracellu, jar signals. This arrangement requires anatomical struc- tures that are readily observed In this paper we introduce the ideas of neural networks: in the context of eytoskeletal structures withia neurons and the role such structures could ples with respect to intracellulat signaling, We review the neural network ap- proach, ia which synupric strengths are adapted £0 pro- duce learning im the network, We identify specific neural network Jearmng parsdigans that would require “bach wards” propagation of signals — axon to dendrite — and other intracellular signals. The structure of the neuronal- eyteskeletion is then described! with emphasis on candi- date mechanisms for supporting intragellulat signals. Al- though one mechavism, material transport, iaas been ver- ified by extensive experimentation, such transport is re tively slow. Faster mechaniams would enable learning to take phice more quickly. Thus we summarize additional mechanisms for faster signaling along the cytoskeleton2s om sn ye asp sores Fort Pes ! ont te oon wae Fig, J. Foodslorwsrd layered nator with three layers of process sup URES: input. hidden, aad ouput Lasers. Bach lay ee fall 1a teonmected te the next Faget ma forwards direetion A weight assacnted wth vach interesonection and an actvativn level ass coated With cach processing amt b Indsidual processing unite per Sorm a weighted suey arising amputs and they eam pute a sues Hee Made 44 where 7 is the Iearnme cite parameter, Although the zht adjustment sn 41 is specified precisely in a mathe matical equation. approximations will suffice. For exam ple, quantized inecements or decrements sn weights are Sufficient for learning: kater iterations can compensitte for on-optimal weights, as {raiaing is completed, Alsw, the tng funstion, usually a semoid. © Network weights are saved by tsaek-error propagation i bic Genny ety balay tor eae nt ae ceoesputed umd used to coFreer incoming wert. Large Years Te tine sures for the ersar delta value computations, Replied trom Dovhoff 11990) 8h pecisston value of y can be raised or lowered during training: usnal ly lowering 9 slowly. provides etter leurnsng, The abore description assumes a layered and fully interconnected network. but the back-propagation learns ing paradigm apphes { more izeguiar configurations ‘with interconnections skippang layers, eoing to bower tes rs. and sparse interconnections. Figure 2 depicts a tay- ered fully connected nesork in & biological context Ecroradiffereneiny can take place between the output signal of the neural network and an external signal ind ating the target value, reach signal is coded hy ats firing‘ueout Laver Fig. a tf the network has more than awe kisers of neurons, the ‘enor feedbuch signats would navel 1 Re carried ack wards weross Somapses vil & reeense azgnmitler «ez. NO} b Internal sigoals ‘aveling up avons weld be merged a¢ Mhe oval teu resulting Ina summation vi signal seteasits. Reprintet trom Dayholf et 41993) error signals through the network amd cous be consid ered variations of the original back propagation paradigm, For example. the sigma-pt network multiplies groups of incoming signals belore « summation js done at the weiget processing unit /Rumelhast et al. 19K6), Ferors are then back propagated through tie network to correct the werehts. The time-delay neural network (TDNN) propa error backwards to adjuist weights ia spite of time delays along interconnections, and the ATNN back-prop error correction signals tat are used to adapt time delays along interconnections (Waibel et el. 1989. Day and Davenport 1993. Lin et ul. 1992), These paradigms allow for learning of spatiotemporal patterns Recurrent networks allow for Jaterial and backwards aterconnections in the network and adapt weights along all interconnections: the vonaccnions may be arbitrails, placed (Werbos 19881. Back propagation itcelf does nat rave to adhere to a fully-connected layered topology but can also have webitrarily placed Forwards connections These networks all depend on the buekwards flow oF in- Formation lahout errors) through the network, the con- struct that requires wxon-ww-cendrite signaling biologica ly gates Other learning paradigms have powerful functional approximation and pattern mappiny capabilities, The work of Copper and coworkers om the RCF neiwerks includes sehemes for recruiting hidden umts ac needed by sinexternally connected layer of output units (Reilly etal, 1952), Thresholds of hidden units ave adjusted aecording to information seat buck wards from the output layer: this, antormation depends on error differences. These networks map mput Patterns to desired output patterns. This also requires reverse intraceliukte signaling. possibly neediated by the cyleskeleion 1m, biological systems. In this ease however. the signal may only have 1o propagate bach- wards along the axon to the cell body where the threshold is set ‘An important class of neural network archiweetwres consists of vo layers of bidirectionally connected cells “The most notable of test axcvtectures is ART iadaptive resonance theory) (Carpenter and Grossberg. i987. 1988 The top-down and bottom-up weiglts have separate rules for adjustment during learning, Recognition of in- coming patterns occurs with “resonance”. a state in which detivanion flowing downwards reinforces that Towing up wards Since each unit in the bottoms layer is bidirection= ally connected to each anit is the top ktyer. the obsiows biological analog is forwards signaling alony axonal membranes and backwards. signaling intracellularly along the cytoskeleton (Dayhoff ct at. 19921 Other teseatchers have proposed learning mechs nisms that employ communication berwsen dendritic sites, Heterosynaplic mechanisms proposed (Finke! ét al. 1989) ullize « selectionist population approach in which other dendrites on the same neuron communicate ‘modify a piven dendritic synapse. Pribrasn (1991) has pro~ posed dendritic inflaenees acting as field effect and hav. ing mechanisms of communication other than that ch: neled along membranes 2. The neuronal cytoskeleton, The cytoskeleton, a lattice-like network of protein poly mers and associated proteins, is found in the mterior of neurons and collectively supports the cell structure and interttal coll processes. This network includes micro jubules (MT), contmoles, actin Alaments and neuroifl ments. membrane enchoniny proteins, and anerotubule sssociated proteins (MAPs! that crosslink MTs and other structures, Microtubules and other evtoskeletal compo: nents are found in victually all eukaryotic cells. and pravide physical and structural support to the cell (soc Fig. 51 In addition, microtubules provide @ communica. tion and transport channel bewween remoce eelh pats. Ln most cells. the MTs connect the cell center «centriole! to the cell petiphery in 4 radial pattern. In nerve cells. MTs, have evoived a striking adaptatien in that the miicratubu- Jaf lattice extends into the extremities of the axons and the dendrites, and some Mis may bypass the cell cotter These internal highly ordered structures allow not onl for the mechanical transport of material pacticles. but recent models have Suggested that the cytoskeleton might he capable of sransmutting fast signals and could de so internally within a nerve cell (Rasmussen et al. 19901. 4, variety of candidate mechanisms have been proposed for microtubule signal propagation. inchuding propagating conformational changes and ionic movons Mlamerol 1987: Hlameroil etal. L989: Rasmussen et al. 19900,