Temporal information transformed into a spatial code by a neural network with realistic properties

Science

Volumn 267, Issue 5200, 1995, Pages 1028-1030

  Buonomano, Dean V ^a   Merzenich, Michael M ^a  

^a UNIVERSITY OF CALIFORNIA   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

ARTICLE; INFORMATION PROCESSING; INHIBITORY POSTSYNAPTIC POTENTIAL; MODEL; NERVE CELL NETWORK; POSTSYNAPTIC POTENTIAL; PRIORITY JOURNAL; SPATIAL SUMMATION; TIME PERCEPTION;

EID: 0028910018     PISSN: 00368075     EISSN: None     Source Type: Journal    
DOI: 10.1126/science.7863330     Document Type: Article

References (33)

1. J. A. Anderson and E. Rosenfeld, Eds., Neurocomputing (MIT Press, Cambridge, MA, 1988); J. A. Anderson, A. Pellionisz, E. Rosenfeld, Eds., Neurocomputing 2 (MIT Press, Cambridge, MA, 1990); D. E. Rumelhart and J. L. McClelland, Parallel Distributed Processing (MIT Press, Cambridge, MA, 1986).
2. J. A. Anderson and E. Rosenfeld, Eds., Neurocomputing (MIT Press, Cambridge, MA, 1988); J. A. Anderson, A. Pellionisz, E. Rosenfeld, Eds., Neurocomputing 2 (MIT Press, Cambridge, MA, 1990); D. E. Rumelhart and J. L. McClelland, Parallel Distributed Processing (MIT Press, Cambridge, MA, 1986).
3. J. A. Anderson and E. Rosenfeld, Eds., Neurocomputing (MIT Press, Cambridge, MA, 1988); J. A. Anderson, A. Pellionisz, E. Rosenfeld, Eds., Neurocomputing 2 (MIT Press, Cambridge, MA, 1990); D. E. Rumelhart and J. L. McClelland, Parallel Distributed Processing (MIT Press, Cambridge, MA, 1986).
4. R. S. Zucker, Annu. Rev. Neurosci. 12, 13 (1989); E. H. Buhl, K. Halasy, P. Somogyi, Nature 368, 823 (1994); D. A. McCormick, in Synaptic Organization of the Brain, G. M. Shepard, Ed. (Oxford Univ. Press, New York, 1990), p. 33.
5. R. S. Zucker, Annu. Rev. Neurosci. 12, 13 (1989); E. H. Buhl, K. Halasy, P. Somogyi, Nature 368, 823 (1994); D. A. McCormick, in Synaptic Organization of the Brain, G. M. Shepard, Ed. (Oxford Univ. Press, New York, 1990), p. 33.
6. R. S. Zucker, Annu. Rev. Neurosci. 12, 13 (1989); E. H. Buhl, K. Halasy, P. Somogyi, Nature 368, 823 (1994); D. A. McCormick, in Synaptic Organization of the Brain, G. M. Shepard, Ed. (Oxford Univ. Press, New York, 1990), p. 33.
7. Models that have focused on temporal tasks often have taken the unrealistic approach of collapsing time into a spatial dimension at the input level.
8. J. Gibbon and L. Allan, Ann. N.Y. Acad. Sci. 432 (1984); P. Tallal, A. M. Galaburda, R. R. Llinas, C. von Euler, ibid. 682 (1993).
9. J. Gibbon and L. Allan, Ann. N.Y. Acad. Sci. 432 (1984); P. Tallal, A. M. Galaburda, R. R. Llinas, C. von Euler, ibid. 682 (1993).
10. B fast and slow IPSP components as a single current with a fast onset and a slow decay. The threshold for all Ex and Inh units was -40 and -50 mV (SD = 4), respectively. PPF was simulated with an α function in which the time since the last spike modulated the strength of the synaptic conductance. The α function exhibited a peak conductance increase of 70% at 100 ms, which produced a maximal EPSP facilitation of approximately 60%.
11. B fast and slow IPSP components as a single current with a fast onset and a slow decay. The threshold for all Ex and Inh units was -40 and -50 mV (SD = 4), respectively. PPF was simulated with an α function in which the time since the last spike modulated the strength of the synaptic conductance. The α function exhibited a peak conductance increase of 70% at 100 ms, which produced a maximal EPSP facilitation of approximately 60%.
12. T. Nathan and J. D. C. Lambert. J. Neurophysiol. 66, 1704 (1991).
13. D. A. McCormick, ibid. 62, 1018 (1989); R. J. Douglas and K. A. C. Martin, J. Physiol. (London) 440, 735 (1990).
14. D. A. McCormick, ibid. 62, 1018 (1989); R. J. Douglas and K. A. C. Martin, J. Physiol. (London) 440, 735 (1990).
15. R. Creager, T. Dunwiddie, G. Lynch, J. Physiol. (London) 299, 409 (1980).
16. The integrate-and-fire elements of Fig. 1 were incorporated into a circuit similar to a previously suggested neocortical circuit [R. J. Douglas and K. A. C. Martin, Neurol Comput. 1, 480 (1989)]. In particular, we simulated layers 4 and 3. The network was composed of 100 inputs and 150 and 250 elements in layers 4 and 3, respectively. In keeping with known empirical data, 20% of the elements in each layer were inhibitory and 15 to 20% of the connections onto each element type were inhibitory. The Ex units projected forward to both the Ex and Inh units in the next layer. Experimental data suggest that the connection probability between excitatory cells in neocortical layer 3 is 4 to 8% [A. Mason, A. Nicoll, K. Stratford, J. Neurosci. 11, 72 (1991); A. M. Thomson, D. Girdlestone, D. C. West, J. Neurophysiol. 60, 1896 (1988)]. Given that we were simulating relatively few elements, we used connection probabilities of 10 to 20% (between all unit types). Connectivity was uniformly random, and connection strengths followed a Gaussian distribution. For the simulations shown here, connection strengths were adjusted so that approximately 80% of the Ex4, Inh4, and Inh3 units and 25% of the Ex3 units fired in response to the first pulse (in general, 6 to 12 active inputs were necessary to reach the threshold). Performance was not qualitatively affected by connection strengths as long as neither the excitation nor inhibition dominated network dynamics. Delays representing axonal, synaptic, and dendritic conduction delays were incorporated into each connection, with delays of 3 ms and 1 ms for connections onto Ex and Inh units, respectively [H. A. Swadlow, Soc. Neurosci. Abstr. 19, 1704 (1993); C. Welker, M. Armstong-James, H. van Der Loos, R. Kraftsik, Eur. J. Neurosci. 5, 691 (1993)]. Noise was presented in the form of jitter during the pulses and by spontaneous activity in the input channels. In the simulations presented here, the spontaneous activity was set at 0.2 Hz, and the interval tuning curves were tested with spontaneous activity levels from 0 to 10 Hz. Over this range, there was a graceful degradation of the tuning curves, with the longer intervals being affected more.
17. The integrate-and-fire elements of Fig. 1 were incorporated into a circuit similar to a previously suggested neocortical circuit [R. J. Douglas and K. A. C. Martin, Neurol Comput. 1, 480 (1989)]. In particular, we simulated layers 4 and 3. The network was composed of 100 inputs and 150 and 250 elements in layers 4 and 3, respectively. In keeping with known empirical data, 20% of the elements in each layer were inhibitory and 15 to 20% of the connections onto each element type were inhibitory. The Ex units projected forward to both the Ex and Inh units in the next layer. Experimental data suggest that the connection probability between excitatory cells in neocortical layer 3 is 4 to 8% [A. Mason, A. Nicoll, K. Stratford, J. Neurosci. 11, 72 (1991); A. M. Thomson, D. Girdlestone, D. C. West, J. Neurophysiol. 60, 1896 (1988)]. Given that we were simulating relatively few elements, we used connection probabilities of 10 to 20% (between all unit types). Connectivity was uniformly random, and connection strengths followed a Gaussian distribution. For the simulations shown here, connection strengths were adjusted so that approximately 80% of the Ex4, Inh4, and Inh3 units and 25% of the Ex3 units fired in response to the first pulse (in general, 6 to 12 active inputs were necessary to reach the threshold). Performance was not qualitatively affected by connection strengths as long as neither the excitation nor inhibition dominated network dynamics. Delays representing axonal, synaptic, and dendritic conduction delays were incorporated into each connection, with delays of 3 ms and 1 ms for connections onto Ex and Inh units, respectively [H. A. Swadlow, Soc. Neurosci. Abstr. 19, 1704 (1993); C. Welker, M. Armstong-James, H. van Der Loos, R. Kraftsik, Eur. J. Neurosci. 5, 691 (1993)]. Noise was presented in the form of jitter during the pulses and by spontaneous activity in the input channels. In the simulations presented here, the spontaneous activity was set at 0.2 Hz, and the interval tuning curves were tested with spontaneous activity levels from 0 to 10 Hz. Over this range, there was a graceful degradation of the tuning curves, with the longer intervals being affected more.
18. The integrate-and-fire elements of Fig. 1 were incorporated into a circuit similar to a previously suggested neocortical circuit [R. J. Douglas and K. A. C. Martin, Neurol Comput. 1, 480 (1989)]. In particular, we simulated layers 4 and 3. The network was composed of 100 inputs and 150 and 250 elements in layers 4 and 3, respectively. In keeping with known empirical data, 20% of the elements in each layer were inhibitory and 15 to 20% of the connections onto each element type were inhibitory. The Ex units projected forward to both the Ex and Inh units in the next layer. Experimental data suggest that the connection probability between excitatory cells in neocortical layer 3 is 4 to 8% [A. Mason, A. Nicoll, K. Stratford, J. Neurosci. 11, 72 (1991); A. M. Thomson, D. Girdlestone, D. C. West, J. Neurophysiol. 60, 1896 (1988)]. Given that we were simulating relatively few elements, we used connection probabilities of 10 to 20% (between all unit types). Connectivity was uniformly random, and connection strengths followed a Gaussian distribution. For the simulations shown here, connection strengths were adjusted so that approximately 80% of the Ex4, Inh4, and Inh3 units and 25% of the Ex3 units fired in response to the first pulse (in general, 6 to 12 active inputs were necessary to reach the threshold). Performance was not qualitatively affected by connection strengths as long as neither the excitation nor inhibition dominated network dynamics. Delays representing axonal, synaptic, and dendritic conduction delays were incorporated into each connection, with delays of 3 ms and 1 ms for connections onto Ex and Inh units, respectively [H. A. Swadlow, Soc. Neurosci. Abstr. 19, 1704 (1993); C. Welker, M. Armstong-James, H. van Der Loos, R. Kraftsik, Eur. J. Neurosci. 5, 691 (1993)]. Noise was presented in the form of jitter during the pulses and by spontaneous activity in the input channels. In the simulations presented here, the spontaneous activity was set at 0.2 Hz, and the interval tuning curves were tested with spontaneous activity levels from 0 to 10 Hz. Over this range, there was a graceful degradation of the tuning curves, with the longer intervals being affected more.
19. The integrate-and-fire elements of Fig. 1 were incorporated into a circuit similar to a previously suggested neocortical circuit [R. J. Douglas and K. A. C. Martin, Neurol Comput. 1, 480 (1989)]. In particular, we simulated layers 4 and 3. The network was composed of 100 inputs and 150 and 250 elements in layers 4 and 3, respectively. In keeping with known empirical data, 20% of the elements in each layer were inhibitory and 15 to 20% of the connections onto each element type were inhibitory. The Ex units projected forward to both the Ex and Inh units in the next layer. Experimental data suggest that the connection probability between excitatory cells in neocortical layer 3 is 4 to 8% [A. Mason, A. Nicoll, K. Stratford, J. Neurosci. 11, 72 (1991); A. M. Thomson, D. Girdlestone, D. C. West, J. Neurophysiol. 60, 1896 (1988)]. Given that we were simulating relatively few elements, we used connection probabilities of 10 to 20% (between all unit types). Connectivity was uniformly random, and connection strengths followed a Gaussian distribution. For the simulations shown here, connection strengths were adjusted so that approximately 80% of the Ex4, Inh4, and Inh3 units and 25% of the Ex3 units fired in response to the first pulse (in general, 6 to 12 active inputs were necessary to reach the threshold). Performance was not qualitatively affected by connection strengths as long as neither the excitation nor inhibition dominated network dynamics. Delays representing axonal, synaptic, and dendritic conduction delays were incorporated into each connection, with delays of 3 ms and 1 ms for connections onto Ex and Inh units, respectively [H. A. Swadlow, Soc. Neurosci. Abstr. 19, 1704 (1993); C. Welker, M. Armstong-James, H. van Der Loos, R. Kraftsik, Eur. J. Neurosci. 5, 691 (1993)]. Noise was presented in the form of jitter during the pulses and by spontaneous activity in the input channels. In the simulations presented here, the spontaneous activity was set at 0.2 Hz, and the interval tuning curves were tested with spontaneous activity levels from 0 to 10 Hz. Over this range, there was a graceful degradation of the tuning curves, with the longer intervals being affected more.
20. The integrate-and-fire elements of Fig. 1 were incorporated into a circuit similar to a previously suggested neocortical circuit [R. J. Douglas and K. A. C. Martin, Neurol Comput. 1, 480 (1989)]. In particular, we simulated layers 4 and 3. The network was composed of 100 inputs and 150 and 250 elements in layers 4 and 3, respectively. In keeping with known empirical data, 20% of the elements in each layer were inhibitory and 15 to 20% of the connections onto each element type were inhibitory. The Ex units projected forward to both the Ex and Inh units in the next layer. Experimental data suggest that the connection probability between excitatory cells in neocortical layer 3 is 4 to 8% [A. Mason, A. Nicoll, K. Stratford, J. Neurosci. 11, 72 (1991); A. M. Thomson, D. Girdlestone, D. C. West, J. Neurophysiol. 60, 1896 (1988)]. Given that we were simulating relatively few elements, we used connection probabilities of 10 to 20% (between all unit types). Connectivity was uniformly random, and connection strengths followed a Gaussian distribution. For the simulations shown here, connection strengths were adjusted so that approximately 80% of the Ex4, Inh4, and Inh3 units and 25% of the Ex3 units fired in response to the first pulse (in general, 6 to 12 active inputs were necessary to reach the threshold). Performance was not qualitatively affected by connection strengths as long as neither the excitation nor inhibition dominated network dynamics. Delays representing axonal, synaptic, and dendritic conduction delays were incorporated into each connection, with delays of 3 ms and 1 ms for connections onto Ex and Inh units, respectively [H. A. Swadlow, Soc. Neurosci. Abstr. 19, 1704 (1993); C. Welker, M. Armstong-James, H. van Der Loos, R. Kraftsik, Eur. J. Neurosci. 5, 691 (1993)]. Noise was presented in the form of jitter during the pulses and by spontaneous activity in the input channels. In the simulations presented here, the spontaneous activity was set at 0.2 Hz, and the interval tuning curves were tested with spontaneous activity levels from 0 to 10 Hz. Over this range, there was a graceful degradation of the tuning curves, with the longer intervals being affected more.
21. The learning rule for the Ex3 to output unit connections established that at the end of each stimulus, recently active connections to the appropriate output unit were strengthened, and active connections to other output units were weakened. Connections were also decreased if an inappropriate output unit fired during a stimulus.
22. G. Ogden and D. Massaro, Psychol. Rev. 85, 172 (1978); C. Wood, J. Acoust. Soc. Am. 60, 1381 (1976).
23. G. Ogden and D. Massaro, Psychol. Rev. 85, 172 (1978); C. Wood, J. Acoust. Soc. Am. 60, 1381 (1976).
24. Phonemes were synthesized with SynSen (for Macintosh). We then converted the spectrograms of the phonemes into spike trains by assigning each input unit of the network to a given frequency of the spectrogram. For this task, the connections from the inputs to the layer 4 units followed a Gaussian function in order to establish some spatial topography.
25. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
26. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
27. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
28. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
29. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
30. V. Braitenberg, Brain. Res. 25, 334 (1967); J. W. Moore, J. E. Desmond, N. E. Berthier, Biol. Cybern. 62, 17 (1989); S. Grossberg and N. A. Schmajuk, Neural Networks 2, 79 (1989); D. W. Tank and J. J. Hopfield, Proc. Natl. Acad. Sci. U.S.A. 84, 1896 (1987); W. E. Sullivan, J. Neurophysiol. 48, 1033 (1982). The nervous system does use a spectrum of time delays for binaural sound localization; however, these delays are less than 1 ms [C. E. Carr, Annu Rev. Neurosci. 16, 223 (1993)].
31. P. S. Antón, G. Lynch, R. Granger, Biol. Cybern. 65, 407 (1991).
32. D. Margoliash, J. Neurosci. 3, 133 (1983).
33. This work was supported by National Institute of Mental Health fellowship F32 MH10431 and NIH grant NS-10414. We thank A. Doupe, S. Lisberger, H. Mahncke, M. Mauk, K. Miller, and J. Raymond for helpful comments and discussions.