626 Population coding and behavioral choice William B Kristan Jr* and Brian K Shawl Many individual behavioral acts are produced by the The idea of motor commands being coded by population combined activity of large populations of broadly tuned activity is not new. For example, neuronal schemes for neurons, and the neuronal populations for different behaviors population coding of movements have been proposed for can overlap. Recent experiments monitoring and manipulating saccadic eye movements [S], arm movements [6], bending neuronal activity during behavioral decisions have begun in leeches [7], and orientation in insects [S]. However, to shed light on the mechanisms that enable overlapping each of these cases consists of singular behaviors with one populations of neurons to generate choices between parameter-the direction of movement-varying over a categorically distinct behaviors. continuous range of movements. The issue we address here is whether the term ‘population coding’ can also be applied when the motor outputs are categorically distinct Addresses behaviors, rather than continuously varying responses. We ‘Biology Department, University of California at San Diego, 9500 Gilman Drive, La Jolla, California 92093-0357, USA; will not attempt to address this issue definitively, but will e-mail: [email protected] review results that bear on it in the hope of fueling (and tThe Neurosciences Institute, 10640 John Jay Hopkins Drive, San constraining) further discussion and debate. Diego, California 92121, USA; e-mail: [email protected] Current Opinion in Neurobiology 1997, 7:826-831 Behavioral choice also contains cognitive aspects, as exemplified by discussions of this topic in two different http:llbiomednet.comlelecreflO959438800700826 reviews in a recent issue of Current Opinion in Neurobi- 0 Current Biology Ltd ISSN 0959-4388 ology on Cognitive neuroscience [9,10]. This review will Abbreviations complement, rather than reiterate, the issues discussed in LIP lateral intraparietal (area) those reviews. MT middle temporal (area) STG stomatogastric ganglion The neural basis of behavioral choice: the classical view Perhaps the most influential scheme for how the behav- ioral repertoire of an animal is organized in the nervous Introduction system was proposed by Tinbergen [ 11.A modified version To behave adaptively, an animal must be able to choose of this scheme is presented in Figure 1, in which neural the appropriate behavior for its current circumstances and ‘centers’ are arrayed in a hierarchy, with the highest must coordinate the various behaviors in its repertoire. levels concerned with more abstract behavioral functions Proposals for how nervous systems might implement such (i.e. drives) and lower levels with more specific aspects decision-making processes date back to classical ethology of motor output (in modern terminology, these would be [l-3]. With a small number of exceptions, however, it is central pattern generators, motor pools, etc.). At a middle only recently that the neural mechanisms that underlie level are centers that control particular whole-animal behavioral choice have come under experimental scrutiny. behaviors, with each center dedicated to the production of that behavior. In this model, the basis for choices between Here, we review some recent work that has begun to behaviors is mutual inhibition between the centers at shed light on how nervous systems produce choices the same level, although Tinbergen recognized that between behaviors, focusing on papers published over there could be other, more complex neural mechanisms the past year. A recurring theme to be found in this involved. work is the suggestion that single neurons can contribute to multiple behaviors (i.e. they are ‘multiplexed’) and, In the ensuing decades, the discovery of neurons that because individual neurons respond to a broad class could, when stimulated, single-handedly initiate whole- of stimuli (i.e. they are ‘coarsely coded’), that large animal behaviors in invertebrates [ll], and the discovery populations of neurons may act to determine what in vertebrates of discrete nuclei that could do the same action an animal takes. In fact, these two concepts are [12], seemed to fit nicely with this hierarchical scheme. often discussed under the general notion of ‘distributed These ‘command-like’ neurons or systems appeared well processing’ [4]. Coarse coding and multiplexing represent suited to fill the role of the behavior-dedicated centers a challenge to classical proposals for the neural basis proposed by Tinbergen (see (131). This view received of decision-making (described below) and suggest the further support from reports of a number of examples possibility that categorically distinct behavioral outcomes of inhibition between command-like neurons [14,15]. A may be coded by the profile of activity across a population simple model also provided support: a simulation em- of neurons, any one of which could contribute to multiple ploying mutually inhibitory connections among command behaviors. neurons for several different behaviors was able to produce Population coding and behavioral choice Kristan and Shaw 827 Cir.*-r 4 ray”,= I on the behavior, it may be driven by reflex pathways or a central pattern generator, and may play either a critically important role or a modest one. Another example is provided by our study of the choice Level of the between swimming and shortening in the medicinal major instinct (reproductive) leech [ZO”]. We found that most of the command-like neurons of the swim circuit, which can elicit swimming 2nd level when artificially stimulated, were excited by stimuli that (fighting, nesting, etc.) produced shortening, as well as by stimuli that produced swimming (Figure 2). Only one of the command-like neurons, cell 204, was inhibited during shortening. These 3rd level results raise the possibility that most of the cells (consummatory act) previously identified as command-like components of the swim circuit are, in fact, multifunctional neurons that 4th level (fins) contribute to more than one behavior. While inhibition is probably important for the choice between swimming 5th level and shortening, this inhibition is not targeted to all (fin rays) command-like neurons, nor exclusively to command-like neurons. 6th level (muscles) 5’ B A conclusion to be drawn from these two studies 7th level 3 6 (motor units) [ 19**,20**] is that neurons with command-like properties 8 2 need not act as behavior-dedicated decision points. To express this another way, knowing what a single command- like cell is doing will not necessarily allow one to predict A hierarchical scheme proposed by Tinbergen Ill, based largely what the animal is doing. If CC5 is active, for instance, upon ethological studies, to explain the nature of decision-making a sea hare may be turning its head, or undergoing local in animals. He proposed that animals make high-level decisions first (e.g. whether to reproduce or to feed), then make a sequence of withdrawal, or performing some other behavior [18’,19”]. more and more specific decisions, down to the level of the choice of If cell Trl is active, a leech may be about to swim or specific motor patterns. In this scheme, decisions are made at ‘nodes’ about to shorten [ZO”] (Figure 2). As a further example in the brain, represented by circles in the diagram. Once activated, of the complex roles that such cells may play, stimulating a node inactivates all the other nodes at the same level, either by inhibiting them directly or by removing their excitation from the higher a particular high-level interneuron in the leech, cell TX, level. Modified from [l]. can produce swimming in a quiescent nervous system or terminate an ongoing swimming bout [Zl]. a variety of surprisingly complex and life-like behavioral Overlapping sets of neurons can be activated sequences (161. In a more complex modeling effort, during categorically distinct behaviors however, a more distributed use of sensory information One consequence of using population coding for contin- over many hierarchical layers, rather than a process of uous behaviors is that the same neuron will be active mutual inhibition at each level, produced more advanta- during a number of different responses (Figure 3). Recent geous behavioral choices [ 171. Recent neurophysiological work shows that this can also be true for categorically investigations also suggest that behavioral choices involve distinct, incompatible behaviors. In one example already complex interactions among neuronal populations. We will considered (Figure Z), many interneurons initially iden- now review some of the points raised by this work. tified as part of the leech swim circuit are also active during shortening [ZO”]. In an additional example, most Higher-order neurons need not be dedicated neurons in the Ap!ysia abdominal ganglion fire during to single behaviors the performance of two or three different behaviors: the An implication of the scheme of Figure 1 is that each gill withdrawal reflex, spontaneous gill contractions, and command-like neuronal system should be dedicated to the respiratory pumping [ZZ]. Another system with extensive production of a single behavior. A number of recent studies neuronal overlap is the spinal circuitry generating the have shown that this is not always the case. For instance, turtle scratch reflex. Turtles show three categorically CCS, an identified neuron in Ap&sia that is both sufficient different forms of scratching, each having a distinct and necessary