Cortical dynamic tuning: The role of intrinsic connections, as revealed by modeling

Current Opinion in Neurobiology

Volumn 6, Issue 6, 1996, Pages 765-772

  Lukashin, Alexander V ^a  

^a VETERANS AFFAIRS MEDICAL CENTER   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

BRAIN CORTEX; FEEDBACK SYSTEM; INFORMATION PROCESSING; MODEL; PRIORITY JOURNAL; REVIEW;

EID: 0030482297     PISSN: 09594388     EISSN: None     Source Type: Journal    
DOI: 10.1016/S0959-4388(96)80026-2     Document Type: Article

References (44)

1. 1. Abbott LF: Decoding neuronal firing and modeling neural networks. Q Rev Biophys 1994, 27:291-331.
2. 2. Lansner A, Ekeberg O: Neuronal network models of motor generation and control. Curr Opin Neurobiol 1994, 4:903-908.
3. 3. Houk JC, Wise SP: Distributed modular architectures linking basal ganglia, cerebellum, and cerebral cortex: their role in planning and controlling action. Cereb Cortex 1995, 5:95-110.
4. 4. Arbib MA (Ed): The Handbook of Brain Theory and Neural Networks. Cambridge, Massachusetts: MIT Press; 1995.
5. 5. Marder E, Abbott LF: Theory in motion. Curr Opin Neurobiol 1995, 5:832-840.
6. 6. Ferster D, Chung S, Wheat H: Orientation selectivity of thalamic input to simple cells of cat visual cortex. Nature 1996, 380:249-252.
7. 7. Vidyasagar TR, Pei X, Volgushev M: Multiple mechanisms underlying the orientation selectivity of visual cortical neurones. Trends Neurosci 1996, 19:272-277.
8. 8. Douglas RJ, Martin KAC, Whitteridge D: A canonical microcircuit for neocortex. Neural Computation 1989, 1:480-488.
9. 9. Douglas RJ, Martin KAC: A functional microcircuit for cat visual cortex. J Physiol (Lond) 1991, 440:735-769.
10. 10. Douglas RJ, Koch C, Mahowald M, Martin KAC, Suarez HH: Recurrent excitation in neocortical circuits. Science 1995, 269:981-985. The feedback excitation resulting from local cortico-cortical interactions is analyzed as a possible mechanism used by populations of cortical neurons to amplify their weak feedforward inputs and to stabilize patterns of activity. The model, which is based on anatomical details and cellular properties, demonstrates that realistic recurrent cortical circuits can achieve 'proportional behavior'-that is, a stable neuronal discharge can be kept proportional to stimulus strength. The results are in agreement with intracellular data. An interesting qualitative interpretation of the equilibrium state in the recurrent network dynamics has been done in electrical engineering terms.
11. 11. Ben-Yishai R, Bar-Or RL, Sompolinsky H: Theory of orientation tuning in visual cortex. Proc Natl Acad Sci USA 1995, 92:3844-3848. An excellent theoretical treatment of mechanisms by which the cortical feedback circuitry may generate orientation selectivity. A neural network that models a cortical hypercolumn (i.e. an aggregate of columns encompassing all orientation preferences) is considered. The basic element of the model is the orientation column, and the strength of the interaction between columns depends on the difference between their preferred orientations, so that the interactions are maximal for similar-orientation columns. The use of mean field approximation makes it possible to describe the behavior of the model using only a few phenomenological parameters. The key point of the study is that for a suitable range of parameters, the mean field approximation yields a solution that represents the generation of orientation tuning in response to cortical interactions alone, that is, in the absence of strong anisotropy in the external input. This solution (termed the 'marginal phase') is characterized by a continuum of states, each of which represents inhomogeneous, orientation-tuned activity of the whole ensemble (in this regime, different states have different location of the peak activity, and the homogeneous state is unstable). In the marginal phase, there are no barriers between inhomogeneous states; therefore, perturbations can move the system from one state to another. However, even a weakly oriented stimulus stabilizes the system in the state in which the location of the peak activity corresponds to the stimulus orientation.
12. ••] {i.e. that the intra-cortical interactions could be a predominant mechanism for orientation selectivity), the model used in this study incorporates more realistic properties of individual neurons and of experimentally reported visual cortical architecture. The results of simulations of this model are consistent with intracellular, extracellular, and pharmacological studies of orientation selectivity. The authors emphasize that the sharp orientation tuning generated by intra-network interactions is an emergent property of this model, as the leading cause of this effect is the recurrent circuitry, which ensures a strong, orientation-selective input to all neurons.
13. •). Secondly, an appropriate range of parameters has been found for a model of a visual cortical hypercolumn, for which the network exhibits a strongly irregular neuronal dynamics that is correlated across the network (the behavior termed 'synchronous chaos'). A possible relation between these results and the observed neuronal stochasticity in cortex is discussed extensively.
14. 14. Suarez H, Koch C, Douglas R: Modeling direction selectivity of simple cells in striate visual cortex within the framework of the canonical microcircuit. J Neurosci 1995, 15:6700-6719. A version of the canonical microcircuit scheme [9] is implemented in a realistic network of excitatory and inhibitory neurons. Both the architecture of the network and the properties of the neurons are derived from experimental data. The functioning of the network is based on massive excitatory cortico-cortical feedback that significantly amplifies a relatively weak geniculate input, thus generating a strong direction selectivity at all contrasts for which there is a response.
15. 15. Scottun BC, Bradley A, Sclar G, Ohzawa I, Freeman RD: The effects of contrast on visual orientation and spatial frequency discrimination: a comparison of single cells and behavior. J Neurophysiol 1987, 57:773-786.
16. 16. Stemmler M, Usher M, Niebur E: Lateral interactions in primary visual cortex: a model bridging physiology and psychophysics. Science 1995, 269:1877-1880. The long-range lateral connections, which are beyond the scope of classical receptive field models, can explain modulatory influences of surround stimuli on the responses of a cell to stimuli within the classical receptive field (see also D Somers, SB Nelson, M Sur, Soc Neurosci Abstr 1994, 20:1577). Physiological experiments have demonstrated that the responses of cells to stimuli within the classical receptive field can be enhanced by similar stimuli outside the classical receptive field if the central stimulus is weak, whereas the same surround stimuli can suppress the responses of cells to high-contrast, strong central stimuli. The authors argue that this dual effect might be a consequence of specifically organized long-range lateral connections, if their influence depends upon the state of excitation of local cell populations responding to the central stimulus. This hypothesis has been tested by a large-scale model that involved simulations of 10 000 excitatory and 10 000 inhibitory integrate-and-fire cells. It is shown that the desired effect can be achieved if the long-range connections are orientation-specific, and the orientation preferences of postsynaptic cells is drawn from a Gaussian distribution centered around the presynaptic cell's preferred orientation. The simulations demonstrate a detectable enhancement of cell responses to weak central stimuli and a suppression of responses to strong stimuli, by a factor of three.
17. 16. Stemmler M, Usher M, Niebur E: Lateral interactions in primary visual cortex: a model bridging physiology and psychophysics. Science 1995, 269:1877-1880. The long-range lateral connections, which are beyond the scope of classical receptive field models, can explain modulatory influences of surround stimuli on the responses of a cell to stimuli within the classical receptive field (see also D Somers, SB Nelson, M Sur, Soc Neurosci Abstr 1994, 20:1577). Physiological experiments have demonstrated that the responses of cells to stimuli within the classical receptive field can be enhanced by similar stimuli outside the classical receptive field if the central stimulus is weak, whereas the same surround stimuli can suppress the responses of cells to high-contrast, strong central stimuli. The authors argue that this dual effect might be a consequence of specifically organized long-range lateral connections, if their influence depends upon the state of excitation of local cell populations responding to the central stimulus. This hypothesis has been tested by a large-scale model that involved simulations of 10 000 excitatory and 10 000 inhibitory integrate-and-fire cells. It is shown that the desired effect can be achieved if the long-range connections are orientation-specific, and the orientation preferences of postsynaptic cells is drawn from a Gaussian distribution centered around the presynaptic cell's preferred orientation. The simulations demonstrate a detectable enhancement of cell responses to weak central stimuli and a suppression of responses to strong stimuli, by a factor of three.
18. •] is extended to analyze the spatio-temporal patterns of activity in visual cortex. The same principle of angular modulation of interactions is utilized. The addition is that the connectivity matrix is not symmetric with respect to inhibitory and excitatory populations (it is still symmetric, however, with respect to orientation preference of interconnected units), which results in dynamic attractors. Two possible scenarios have been demonstrated: the locking of time-dependent, sharply orientation-tuned activity to the rotating external stimulus, and the generation of spontaneous traveling pulses of activity.
19. 18. Fohlmeister C, Gerstner W, Ritz R, Van Hemmen JL: Spontaneous excitations in the visual cortex: stripes, spirals, rings, and collective bursts. Neural Computation 1995, 7:905-914. A neural network model of the visual cortex is constructed utilizing a phenomenological approach described in [19]. Individual neurons obey stochastic dynamics in discrete time: the instantaneous value of the net synaptic input determines the probability of firing a spike. This approach makes it possible to investigate very large ensembles of neurons over a long period of time without solution of differential equations describing the time dependence of membrane potentials. On the other hand, such important ingredients of realistic neural circuits as refractoriness, noise, axonal delays, and the time courses of excitatory and inhibitory postsynaptic potentials can be explicitly incorporated into the model. The key point of the present study is that the local connectivity is spatially organized as a 'Mexican hat': that is, inside a column with a given direction preference, the cells make excitatory connections with each other, whereas in columns with strongly different direction preferences, the cells inhibit each other. A wide class of dynamic behaviors has been found by varying the parameters determining the spatial modulation of connection strengths. Although there are no external inputs and the dynamics evolve spontaneously, starting from random initial conditions, the patterns of activity propagate in an orderly fashion, suggesting the existence of non-trivial dynamic attractors of the system.
20. 19. Gerstner W, Ritz R, Van Hemmen JL: A biologically motivated and analytically soluble model of collective oscillations in the cortex. I. Theory of weak locking. Biol Cybern 1993, 68:363-374.
21. 20. Wolpert DM, Ghahramani Z, Jordan MI: An internal model for sensorimotor integration. Science 1995, 269:1880-1882.
22. 21. Muller RU, Ranck JB Jr, Taube JS: Head direction cells: properties and functional significance. Curr Opin Neurobiol 1996, 6:196-206.
23. 22. Muller M, Wehner R: Path integration in desert ants, Cataglyphis fortis. Proc Natl Acad Sci USA 1988, 85:5287-5290.
24. 23. Gallistel CR: The Organization of Learning. Cambridge, Massachusetts: MIT Press; 1990.
25. 24. Touretzky DS, Redish AD, Wan HS: Neural representation of space using sinusoidal arrays. Neural Computation 1993, 5:869-884.
26. 25. Skaggs WE, Knierim JJ, Kudrimoti HS, McNaughton BL: A model of the neural basis of the rat's sense of direction. In Advances in Neural Information Processing Systems, vol 7. Edited by Tesauro G, Touretzky D, Leen T. Cambridge, Massachusetts: MIT Press; 1995:173-180.
27. 26. Hartmann G, Wehner R: The ant's path integration system: a neural architecture. Biol Cybern 1995, 73:483-497.
28. 27. Wittmann T, Schwegler H: Path integration - a network model. Biol Cybern 1995, 73:569-575. A model that demonstrates how path integration can be performed by a neural system is presented. An approach is adopted that utilizes a sinusoidal array architecture suggested in [24] for efficient vector operations in neural networks. The model consists of two sub-networks. One sub-network stores the home vector via feedback connections, whereas the other sub-network serves to respond to external signals that code changes induced by self-motion. Because there are connections between the sub-networks, the change in direction is added to the home vector computed so far, thus resulting in the path integration. However, the stability of the activity profile with respect to noise-a crucial problem for rotation-invariant systems-has not been analyzed adequately.
29. ••], in spite of some differences in the shapes of tuning curves and the distributions of connection strengths.) The stable pattern of activity, which is sustained by mutual connections between cells, is referred to as the internal representation of the head direction. A specific mechanism ensuring the transition between the states is proposed: a fast modulation of synaptic connections that shifts the equilibrium towards another state. The synaptic modulation could be provided, for example, by the input from the vestibular system. The author shows that the activity profile may be shifted without destroying its shape, if certain constraints are imposed on the modulatory mechanism.
30. 29. Georgopoulos AP, Taira M, Lukashin AV: Cognitive neurophysiology of the motor cortex. Science 1993, 260:47-52.
31. 30. Georgopoulos AP: Current issues in directional motor control. Trends Neurosci 1995, 18:506-510. A comprehensive review of recent experimental and theoretical studies on the cortical mechanisms subserving complex visually guided and memorized motor actions.
32. 31. Schwartz AB: Direct cortical representation of drawing. Science 1994, 265:540-542.
33. 32. Keller A: Intrinsic synaptic organization of the motor cortex. Cereb Cortex 1993, 3:430-441.
34. 33. Thomson AM, Deuchars J: Temporal and spatial properties of local circuits in neocortex. Trends Neurosci 1994, 17:119-126.
35. 34. Wickens J, Hyland B, Anson G: Cortical cell assemblies: a possible mechanism for motor programs. J Motor Behav 1994, 26:66-82.
36. 35. Abbott LF, Blum KI: Functional significance of long-term potentiation for sequence learning and prediction. Cereb Cortex 1996, 6:406-416. A novel mechanism for learning and generating trajectories of movement is proposed and analyzed. A network of location-tuned neurons with broad and overlapping tuning curves is trained by means of a standard model of longterm potentiation. The paper emphasizes the importance of proprioceptive inputs in movement generation. It is shown that because of experience-induced changes in synaptic connections between coding neurons, the network can continuously generate from current sensory input the next target location based on the training experience. Once this new current position is reached, new proprioceptive input induces the next step along a learned trajectory and so on. The model is stable with respect to spatial perturbations during the movement. If some perturbation occurs, the network always shifts the actual trajectory back toward the learned path. Another noteworthy feature of the model is that the spatial and temporal characteristics of memorized movement are separated. The internal connections store information about the shape of trajectory and not the rate at which it is generated, whereas the retrieval rate is determined by the motor response and subsequent proprioception. In this scheme, the task can be performed at any desired rate, if an additional mechanism is involved that controls the timing of proprioception.
37. ••], in which the authors concentrated on the learning and retrieving of single trajectories. In this study, they analyze the case of multiple superimposed paths. In their model, the next predicted location depends primarily on the current location, and at the junction of several trajectories, some confusion can arise. Nevertheless, it is shown that a large number of correlated training paths do not destroy the orientation mechanism suggested. Rather, it results in a cognitive space map that extracts common features of training examples. The authors illustrate how this map can be used to guide navigation in the environment.
38. 37. Lukashin AV, Wilcox GL, Georgopoulos AP: Modeling of directional operations in the motor cortex: a noisy network of spiking neurons is trained to generate a neural-vector trajectory. Neural Networks 1996, 9:397-410. The authors describe a model of interconnected integrate-and-fire neurons that produces a well-learned neural-vector trajectory when supplied with an appropriate external input that acts only to initiate the trajectory. To train the network of noisy spiking units, a modified variant of the simulated annealing procedure has been used. This approach guarantees the stability of learned dynamics. The results of simulations suggest that intrinsic connections in the motor cortex may encode learned movement trajectories without requiring external inputs and that the same internal mechanism can provide tuning properties of individual neurons.
39. ••,19]. The connectivity is organized so that there is a large general-purpose core, within which connections between neurons are symmetric and maximally excitatory/inhibitory for the neurons with similar/opposite preferred directions. This part of the network does not carry any information about specific trajectories. There are also a set of relatively small networks that are highly specific for particular trajectories and that are interconnected with the general-purpose network. The information about the shapes of particular trajectories is stored in synaptic connections within specialized networks and between them and the core network. These interconnections should be adjusted specifically to each of the trajectories to be memorized. The specialized networks form a mutually inhibitory 'winner-take-all' system. When a particular trajectory needs to be performed, a short-lasting external signal must activate the appropriate specialized network, which inhibits all other small networks and activates the core. The activated-specialized part and the core, as one recurrent network, generate the desired neural-vector trajectory as a dynamical attractor of the system. The initiated dynamics is governed for internal feedback exclusively, and there is no need for external signals that specify the dynamics. However, in order to switch the network from one attractor to another, a new external signal must intervene the dynamics for a short period of time (see [39] and Figure 1).
40. 39. Lukashin AV, Wilcox WL, Georgopoulos AP: Overlapping neural networks for multiple motor engrams. Proc Natl Acad Sci USA 1994, 91:8651-8654.
41. 40. Pellizzer G, Sargent P, Georgopoulos AP: Motor cortical activity in a context-recall task. Science 1995, 269:702-705. Using neurophysiological recordings in the brains of behaving monkeys, the authors analyze and compare the dynamics of motor cortical activity for two visuomotor tasks: a mental rotation task, which requires the production of movement at an angle from a stimulus direction, and a memory scanning task, which requires the selection of an appropriate movement direction, depending on the serial position of stimuli in a sequence. They have found a fundamental difference between dynamic behaviors of the neuronal population vector in these two tasks: gradual rotation of the vector in the mental rotation task and abrupt change of the population vector direction in the memory scanning task.
42. 41. Redish AD, Touretzky DS: The reaching task: evidence for vector arithmetic in the motor system? Biol Cybern 1994, 71:307-317.
43. 42. Saunas E, Abbott LF: Transfer of coded information from sensory to motor networks. J Neurosci 1995, 15:6461-6474. In modeling of sensory-guided motor tasks, one of the problems is how to transfer directional information from an array of sensory neurons coding target location to an array of neurons coding direction of upcoming movement. The authors have derived conditions on synaptic connections between the two arrays that assure a properly aligned motor response. Coupled sensory and motor networks are considered similarly as populations of location-tuned neurons. The goal is to find a set of connections such that firing in the sensory array corresponding to a particular target location would generate activity in the motor array that encodes movement toward the target. It appears that all that is required to align two networks is a simple rule that the strength of the synaptic connection between neurons depends only on the magnitude of the difference between their preferred locations. In fact, this result is valid regardless of what type of directional information is transferred and to what particular variable the neurons are tuned.
44. 43. Lukashin AV, Amirikian BR, Georgopoulos AP: Neural computations underlying the exertion of force: a model. Biol Cybern 1996, 74:469-478. A model is developed that simulates possible mechanisms by which motor commands encoded in the population activity of motor cortical cells could be transformed by downstream networks into balanced activation of muscles. The central ingredient of the model is a pattern of connections between the motor cortical population and spinal cord interneurons. The model predicts that connection strengths should be correlated with directional tuning of interconnected units in the way that the strongest connection must be between units with similar directional preferences.