Computational neuroethology

Post-publication activity

Figure 1: Animal behavior arises from the interaction between neural dynamics, peripheral dynamics and ecological dynamics

Computational neuroethology is concerned with the computational modeling of the neural basis of animal behavior. As a consequence, it combines the methodology of computational neuroscience with the perspective of neuroethology to construct models of behavior in which neural activity, peripheral biomechanics and ecological context are treated on an equal footing ( Figure 1).

History

The term "computational neuroethology" was coined simultaneously by Beer (1989/1990), Cliff (1990/1991a) and Achacoso & Yamaoto (1990). However, there were a number of historical precedents. For example, an early attempt to explore the neural basis of behavior through models was W. Grey Walter's (1953/1963) use of electronic analogues of neural circuits to control the behavior of a series of robotic "turtles" that he constructed (with such fanciful names as Machina speculatrix and Machina docilis). Likewise, Braitenberg (1984) described a classic series of thought experiments about how simple neural circuits could generate sophisticated behavior when coupled appropriately to a body and environment. Finally, Arbib constructed a series of models of the neural basis of visually-guided prey-catching behavior in frogs and toads that he referred to as Rana computatrix (Arbib, 1987). These models were grounded in the neuroanatomy of the optic tectum and pretectum. They not only reproduced the observed approach and avoidance responses (including the effect of pretectal lesions), but also the ability of frogs and toads to detour around obstacles before striking at targets behind semi-transparent barriers. More recently, he has been extending this work to spatial navigation in rats, saccades and grasping in monkeys, and even the role of the mirror system in human language (Arbib, 2003).

In his dissertation and subsequent book, Beer (1989/1990) described a simulated insect that he termed Periplaneta computatrix ( Figure 2). The body of this model insect included six legs that could swing back and forth relative to the body and feet that could grasp and release the substrate. It also had antennae containing tactile and long-range chemical sensors, a mouth capable of biting that also contained tactile and short-range chemical sensors, and an internal energy store that was depleted at a fixed rate by a simple metabolism and had to be replenished. The model insect existed in an environment containing a variety of obstacles and food sources that emitted a chemical odor. The behavioral repertoire of the model insect included walking at a range of speeds with a variety of different gaits, wandering, edge-following, and searching for and consuming food, all of which were organized into a behavioral hierarchy. These behaviors were coordinated by a model nervous system. The design of the locomotion circuitry was based on Pearson's flexor burst-generator model of cockroach walking (Pearson, 1976) and an extensive lesion study of this circuit revealed a rich interplay between central and peripheral mechanisms in the generation of different gaits. In addition, the feeding circuitry was based on work on feeding behavior in Aplysia (Weiss, Chiel, Koch and Kupfermann, 1986; Weiss, Chiel and Kupfermann, 1986) and the neural basis of the model insect's behavioral hierarchy was based on ideas about decision-making in insect nervous systems (Altman and Kien, 1989).

Figure 2: The body and nervous system of P. computatrix. The neurons are color-coded according to the behaviors in which they participate.

In a series of papers and his dissertation, Cliff (1991b, 1992) described a model of visual tracking in a simulated fly that he termed Syritta computatrix. The specific focus of this work was on modeling the visually-guided pursuit of conspecifics by male Syritta pipiens hoverflies (Collett and Land, 1975). More generally, however, this work was motivated by Cliff's interest in so-called active or animate vision, in which a vision system has active control over its own motion. The model included a sophisticated 3D compound eye model with realistic nonuniform foveal sampling of the optic array. Using significant extensions of schematic processing models proposed by Collett and Land, Cliff implemented neural circuits that could control the position and orientation of S. computatrix within its simulated environment so as to select and fixate a particular conspecific and then track it at a constant distance.

Key Principles

The most obvious difference between computational neuroethology and computational neuroscience is that the former includes models not only of neural circuitry, but also peripheral biomechanics and environmental context. Of course, as in computational neuroscience, these models may range from biologically realistic to highly simplified. However, computational neuroethology models are better positioned to make the bridge from neural activity to behavior. Consider, for example, a swimming animal. On the one hand, because muscles, inertia, skin and water can act as low-pass filters, a large change in neural activity may have only a negligible effect on behavior in some conditions. On the other hand, due to nonlinearities in muscle activation, skeletal geometry and hydrodynamics, a very small change in neural activity may have a disproportionally large effect on behavior in other conditions. Thus, only models that include all three components are in a position to properly evaluate the behavioral significance of a given feature of neural activity.

However, there is a deeper difference between computational neuroethology and computational neuroscience. When peripheral biomechanics and environmental dynamics are taken seriously, one begins to question the accuracy of the very notion of "the neural control of behavior". Indeed, it can be very misleading to think of the body as a mere puppet whose only task is to respond faithfully to the commands of the nervous system as it performs on the stage of the environment. Instead, computational neuroethology recognizes behavior as arising from the interactions between neural activity, peripheral biomechanics, and environmental dynamics rather than as the product solely of any one of these components (Beer, 1995; Chiel and Beer, 1997). For example, it is now widely accepted that putting sensory systems in closed-loop interaction with an environment can completely change the nature and difficulty of perceptual problems. Distributing behavioral mechanisms across the brain-body-environment boundaries certainly makes more sense from an evolutionary perspective, since it is the behavioral efficacy of the entire brain-body-environment system that is selected for, and it seems likely that evolution would take full advantage of any available freedom.

Recent Examples

A striking example of the importance of the interaction of neural control, biomechanics and environmental forces for behavior was provided by a simulation of lamprey swimming created by Ekeberg, Lansner and Grillner (Ekeberg, Lansner and Grillner, 1993). As the two-dimensional simulated lamprey swam through a region of faster moving water, the feedback from stretch receptors induced a faster pattern of swimming that allowed the simulated organism to swim through the obstacle. Removal of the sensory feedback impaired the model's response, and suggested that the new swimming pattern emerged not from a discrete neural command, but from the interacting dynamics of the nervous system, body, and surrounding fluid. Further evidence for the importance of these interactions was provided by a detailed simulation of biomechanics and sensory feedback of stick insect walking. The model demonstrated that the timing of the step cycle could be understood in terms of the sequential activation of and feedback from specific sensory receptors activated by the leg movement, and the interaction of the leg with the environment (i.e., mechanical loading and unloading) (Ekeberg, Blümel and Büschges, 2004). The role of purely neuromechanical feedback for enforcing normal gaits was demonstrated by a simulation of walking in the cat ( Figure 3). Removal of load from a limb, monitored by force in the ankle extensor muscle, serves as key sensory feedback that can enforce stable patterns of walking, even in the absence of other feedback. In contrast, using hip extension to terminate stance led to unstable gaits. Thus, mechanical coupling through the body can serve as a stabilizing mechanism for locomotion (Ekeberg and Pearson, 2005; Pearson, Ekeberg, and Büschges, 2006).

Figure 3: A frame from a neuromechanical simulation of stepping in the hind legs of the cat. Figure courtesy of Örjan Ekeberg.

What mechanisms allow organisms to flexibly switch among different gaits, and how do the interplay of biomechanics and neural control contribute to this flexibility? To explore this question, Ijspeert (Ijspeert, 2001) created a two-dimensional biomechanical simulation of salamander locomotion (the model was modified from Ekeberg's simulation of lamprey swimming), and then used genetic algorithms to develop a neural circuit in three stages: segmental oscillators, intersegmental coupling, and finally development of a limb central pattern generator (CPG) coupled to the body central pattern generator. The resulting neural network generated gaits similar to those observed in salamander, responded to asymmetric inputs with turning movements, and could smoothly transition among different gaits (swimming versus trotting) as a function of the intensity of the tonic drive to the circuit. The results suggest that the intrinsic dynamical modes of the body central pattern generator, and the connections from the limb CPG to the body CPG, which induce the body CPG to produce a standing wave rather than a traveling wave, play important roles in inducing gait transitions.

Can changes in the biomechanical interactions of the structures of the periphery affect neural control? A neuromechanical model of feeding in Aplysia has recently been developed and extensively analyzed (Sutton et al., 2004; Novakovic et al., 2006; see also Aplysia feeding biomechanics). The model has clarified multiple subtle interactions of biomechanics and neural control during biting, swallowing and rejection responses. In particular, during rejection, the model predicted that the change in shape of the grasper as it closes would both lengthen and strengthen a protractor muscle. Subsequent experiments (Ye, Morton and Chiel, 2006) demonstrated that this predicted mechanical reconfiguration was observed. These results suggest the concept of neuromechanical modulation: a set of motor neurons may alter the actions of another set of motor neurons through peripheral mechanical changes (rather than through central release of modulatory compounds).

How does the volume of space that an animal can sense relate to the volume defined by its motor capabilities? Malcolm MacIver, Mark Nelson, and colleagues have addressed this question by studying prey capture in a weakly electric fish (the South American black ghost knifefish Apteronotus albifrons). The motor volume of the fish was estimated over all the initial velocities that were typical of prey capture from infrared video of actual prey capture sequences. The sensory volume that could be used for prey detection by the active electric sense was estimated both from measurements and from a detailed model of the fish's body, sensor distribution, electric field and the prey (Snyder, Nelson, Burdick and MacIver, 2007). There was a striking similarity between the two volumes. The sensing using the electric field is omnidirectional. At the same time, the fish's ability to swim backwards as well as forwards endows its motor volume for prey capture with a similar omnidirectionality. More generally, they conclude that the ratio of sensory volume to motor volume leads to different behavioral strategies. When the ratio is less than one, an animal will collide with objects in its environment. If the ratio is close to one, animals will adopt a reactive strategy, in which there is a tight coupling between sensory inputs and motor action. If the ratio is large, the animal may adopt a deliberative strategy, in which complex motion planning can be part of its response.

Related Approaches

Computational neuroethology is closely related to two other lines of work. The first, biorobotics, involves the construction of robotic models of neuroethological systems (Beer, Ritzmann and McKenna, 1993; Webb, 2001). Biorobotics is a natural extension of computational neuroethology given the latter's emphasis on embodiment and situatedness. As long as they are properly validated against the biological system of interest, robotic models can be particularly useful when good models of the physical phenomena involved are unavailable or too computationally expensive to simulate.

The other related line of work is evolutionary approaches to computational neuroethology (Beer and Gallagher, 1992; Cliff, Husbands and Harvey, 1993). In this approach, genetic algorithms are used to evolve neural (and sometimes body) properties in a model brain-body-environment system so as to exhibit some desired behavioral performance. The evolved agents can then be subjected to a detailed analysis to uncover their principles of operation. Evolutionary approaches are particularly useful for exploring spaces of possible solutions to a given behavioral task because these approaches minimize a priori assumptions about how a given behavior ought to be instantiated. They can also be useful for exploring different ways to complete a computational neuroethology model when only partial neural circuitry is available for a biological system of interest.

References

Achacoso T.B. and Yamaoto W.S. (1990) Artificial ethology and computational neuroethology: A scientific discipline and its subset by sharpening and extending the definition of artificial intelligence. Perspectives in Biology and Medicine, 33:379-389

Altman J.S. and Kien J. (1989) New models for motor control. Neural Computation, 1:173-183.

Arbib M.A. (1987) Levels of modeling of visually-guided behavior. Behavioral and Brain Science, 10:407-465

Arbib M.A. (2003) Rana computatrix to human language: Towards a computational neuroethology of language evolution. Philosophical Transactions of the Royal Society of London A - Mathematical, Physical and Engineering Sciences, 361:2345-2379

Beer R.D. (1990) Intelligence as adaptive behavior: An experiment in computational neuroethology. Academic Press. This manuscript first appeared as a Ph.D. dissertation in 1989.

Beer R.D. (1995) A dynamical systems perspective on agent-environment interaction. Artificial Intelligence, 72:173-215

Beer R.D. and Gallagher J.G. (1992) Evolving dynamical neural networks for adaptive behavior. Adaptive Behavior 1:91-122

Beer R.D., Ritzmann R.E. and McKenna T. (1993) Biological Neural Networks in Invertebrate Neuroethology and Robotics. Academic Press

Braitenberg V. (1984) Vehicles: Experiments in Synthetic Psychology. MIT Press

Chiel H.J. and Beer R.D. (1997) The brain has a body: Adaptive behavior emerges from interactions of nervous system, body and environment. Trends in Neurosciences 20:553-557

Cliff D. (1991a) Computational neuroethology: A provisional manifesto. In J-A. Meyer and S.W. Wilson (Eds), From Animals to Animats 3: Proceedings of the First International Conference on the Simulation of Adaptive Behavior (pp. 29-39). This manuscript first appeared as a technical report in 1990.

Cliff D. (1991b) The computational hoverfly: A study in computational neuroethology. In J-A. Meyer and S.W. Wilson (Eds), From Animals to Animats 3: Proceedings of the First International Conference on the Simulation of Adaptive Behavior (pp. 87-96). MIT Press

Cliff D. (1992) Neural networks for visual tracking in an artificial fly. In F.J. Varela and P. Bourgine (Eds), Toward a Practice of Autonomous Systems: Proceedings of the First European Conference on Artificial Life (pp. 78-87). MIT Press

Cliff D. (2002) Neuroethology, computational. In M.A. Arbib (Ed), The Handbook of Brain Theory and Neural Networks, 2nd edition (pp. 737-741). MIT Press

Cliff D., Husbands P. and Harvey I. (1993) Explorations in evolutionary robotics. Adaptive Behavior, 2:73-110

Collett T.S. and Land M.F. (1975) Visual control of flight behaviour in the hoverfly Syritta pipiens L. J. Comparative Phsyiology 99:1-66

Ekeberg Ö. (1993) A combined neural and mechanical model of fish swimming. Biological Cybernetics, 69:363-374

Ekeberg Ö., Blümel M. and Büschges A. (2004) Dynamic simulation of insect walking. Arthropod Structure and Development, 33:287-300

Ekeberg Ö. and Pearson K.G. (2005) Computer simulation of stepping in the hind legs of the cat: An examination of mechanisms regulating the stance-to-swing transition. J. Neurophysiology, 94:4256-4268

Ijspeert A.J. (2001) A connectionist central pattern generator for the aquatic and terrestrial gaits of a simulated salamander. Biological Cybernetics, 84:331-348

Novakovic V.A., Sutton G.P., Neustadter D.M., Beer R.D. and Chiel H.J. 2006. Mechanical reconfiguration mediates swallowing and rejection in Aplysia californica. J. Comp. Physiol. – A. 192: 857 – 870.

Pearson K.G. (1976) The control of walking. Scientific American, 235:72-86

Pearson K.G. Ekeberg Ö. and Büschges A. (2006) Assessing sensory function in locomotor systems using neuro-mechanical simulations. Trends in Neurosciences, 29:625-631

Snyder J.B., Nelson M.E., Burdick J.W. and MacIver M.A. (2007) Omnidirectional sensory and motor volumes in electric fish. PLoS Biol 5(11): e301.

Sutton G.P., Mangan E.V., Neustadter D.M., Beer R.D., Crago P.E. and Chiel H.J. (2004) Neural control exploits changing mechanical advantage and context dependence to generate different feeding responses in Aplysia. Biol. Cybern. 91: 333-345.

Walter W.G. (1963) The Living Brain. W.W. Norton. This is a reprint of a book that was originally published in 1953.

Webb B. (2001) Can robots make good models of biological behavior? Behavioral and Brain Sciences, 24:1033-1050

Weiss K.R., Chiel H.J., Koch U. and Kupfermann I. (1986) Activity of an identified histaminergic neuron, and its possible role in arousal of feeding behavior in semi-intact Aplysia. J. Neuroscience, 6:2403-2415