Autonomous learning with complex dynamics. (English) Zbl 0840.92006
We have constructed a neural network model of the olfactory cortex which is able to use its complex nonlinear dynamics for a fast information processing in associative memory tasks. Recognition of noisy input patterns can be sped up by oscillatory or near-chaotic dynamics. We have also included neuromodulatory effects in our model in order to investigate the role of controlled neurodynamics for learning and associative memory. By changing a gain parameter the dynamical behavior can shift dramatically and have a significant effect on memory performance. Changing synaptic transmission during learning can enhance performance further. Neuromodulatory gain control can be used in regulating the accuracy and rate of the recognition process, depending on the current demands. In particular, it is demonstrated that neuromodulatory effects can reduce recall/reaction time considerably. Indeed, intrinsic “neuronal” noise can be used in the same way. We show that such regulatory mechanisms can automatically make the system more or less sensitive to external input and determine the speed and accuracy in the recognition of unknown noisy input patterns, which may change with time. This may have implications both for biological and industrial purposes.
We have tried to make our system as “autonomous” as possible, i.e., with as little interference as possible from the programmer. We give the system an initial network connectivity, grossly resembling that of an evolved real system (the olfactory cortex). Regardless if the system is aimed at simulating aspects of an organism or an artificial automat/robot the system self-organizes and gives an output pattern that is dependent on the present connection strengths and the current input pattern. An input pattern is provided by the programmer and the network treats this input in an “autonomous way”, comparing it to previously stored patterns and if a match is found it gives an appropriate response, otherwise it stores the current input pattern as a novel memory. The new memory is associated with a label given by the programmer (in the general case this could be done by another cortical area or subsystem). An essential idea here is that the system should perform an efficient interaction with the environment mainly by giving a fast (and sufficiently accurate) response to any input. Even though this study is based on the structure and dynamical properties of the olfactory cortex, we believe it can be generalized to many types of pattern recognition and associative memory.
We have tried to make our system as “autonomous” as possible, i.e., with as little interference as possible from the programmer. We give the system an initial network connectivity, grossly resembling that of an evolved real system (the olfactory cortex). Regardless if the system is aimed at simulating aspects of an organism or an artificial automat/robot the system self-organizes and gives an output pattern that is dependent on the present connection strengths and the current input pattern. An input pattern is provided by the programmer and the network treats this input in an “autonomous way”, comparing it to previously stored patterns and if a match is found it gives an appropriate response, otherwise it stores the current input pattern as a novel memory. The new memory is associated with a label given by the programmer (in the general case this could be done by another cortical area or subsystem). An essential idea here is that the system should perform an efficient interaction with the environment mainly by giving a fast (and sufficiently accurate) response to any input. Even though this study is based on the structure and dynamical properties of the olfactory cortex, we believe it can be generalized to many types of pattern recognition and associative memory.
MSC:
| 92C20 | Neural biology |
| 92B20 | Neural networks for/in biological studies, artificial life and related topics |
| 68T05 | Learning and adaptive systems in artificial intelligence |
Keywords:
olfactory cortex; nonlinear dynamics; information processing; associative memory; noisy input patterns; neuromodulatory effects; learning; synaptic transmission; recognition; pattern recognitionReferences:
| [1] | Freeman, Brain Res. Rev. 10 pp 147– (1985) |
| [2] | Baird, Physica 22D pp 150– (1986) |
| [3] | Li, Biol. Cybern. 61 pp 379– (1989) |
| [4] | Li, Biol. Cybern. 62 pp 349– (1990) |
| [5] | Hopfield, Proc. Natl. Acad. Sci. USA 88 pp 6462– (1991) |
| [6] | Érdi, Biol. Cybern. 69 pp 57– (1993) |
| [7] | Granger, J. Cog. Neurosci 1 pp 61– (1989) |
| [8] | Liljenström, Int. J. Neural Systems 2 pp 1– (1991) |
| [9] | Wilson, J. Neurophys. 67 pp 981– (1990) |
| [10] | Bressler, Brain Res. 409 pp 294– (1987) |
| [11] | Ambros-Ingerson, Science 247 pp 1344– (1990) |
| [12] | Haberly, Trends in Neurosci. 12 pp 258– (1989) |
| [13] | Skarda, Brain and Behav. Sci. 10 pp 161– (1987) |
| [14] | The Synaptic Organization of the Brain, Oxford Univ. Press, New York, 1979. |
| [15] | Haberly, Chem. Sens. 10 pp 219– (1985) |
| [16] | Mass Action in the Nervous System, Academic Press, New York, 1975. |
| [17] | Freeman, Elect. Clin. Neurophys. 44 pp 586– (1978) |
| [18] | Bressler, Electroenceph. Clin. Neurophysiol. 50 pp 19– (1980) |
| [19] | ”Functional organization of primate visual cortex,” In Cerebral Cortex, Vol. 3, Visual Cortex, and (Eds.), Plenum Press, New York, 1985. |
| [20] | Biedenbach, Exp. Neurol. 16 pp 464– (1966) |
| [21] | Bland, Prog. Neurobiol. 26 pp 1– (1986) |
| [22] | Rowntree, Brain Res. 362 pp 98– (1986) |
| [23] | and , ”The cholinergic hypothesis of memory: A review of the animal experiments,” In Psychopharmacology of the Ageing Nervous System, and , (Eds.), Plenum Press, New York, 1989, pp. 237–324. |
| [24] | Larson, Science 232 pp 985– (1986) |
| [25] | Freeman, Biol. Cybern. 33 pp 237– (1979) |
| [26] | and , ”Acetylcholine and cortical oscillatory dynamics,” In Computation and Neural Systems, and (Eds.), Kluwer, 1993, pp. 523–530. · doi:10.1007/978-1-4615-3254-5_79 |
| [27] | Hasselmo, J. Neurophysiol. 67 pp 1222– (1992) |
| [28] | Kennedy, IEEE Trans. Circuits and Sys. 36 pp 1133– (1989) · doi:10.1109/31.192428 |
| [29] | Madison, J. Neurosci. 7 pp 733– (1987) |
| [30] | Hasselmo, J. Neurophysiol. 67 pp 1230– (1992) |
| [31] | and , ”Some comments on the weaving of contemporaneous minds,” Proceedings of the Conference on Measuring Chaos in the Human Brain, ( and (Eds.), World Scientific, 1991, pp. 138–155. |
| [32] | and , ”Period adding hierarchical protein modes and electroencephalographically defined states of consciousness,” Proceedings of the 1st Experimental Chaos Conference, S. Vohra et al. (Eds.), World Scientific, 1992, pp 175–194. |
| [33] | Mandell, J. Stat. Phys. 70 pp 355– (1993) |
| [34] | Electrophysiology of small cultured hippocampal neurons, Doctoral Thesis, Stockholm, 1991. |
| [35] | The Organization of Behavior, Wiley, New York, 1949. |
| [36] | Hasselmo, Neural Computation 5 pp 32– (1993) |
| [37] | and , ”The role of noise in cortical associative memory,” manuscript in preparation. |
| [38] | Crick, Seminars Neurosci. 2 pp 263– (1990a) |
| [39] | Eckhorn, Biol. Cybern. 60 pp 121– (1988) |
| [40] | Gray, Proc. Natl. Acad. Sci. USA 86 pp 1698– (1989) |
| [41] | Gray, Nature 338 pp 334– (1989) |
| [42] | , and , ”A model of cortical associative memory based on Hebbian cell assemblies,” In Computation and Neural Systems, and (Eds.), Kluwer, 1993, in press. |
| [43] | ”Response properties of a population of neurons,” Int. J. Neural Systems, 1993, in press. |
| [44] | Llinas, Proc. Natl. Acad. Sci. USA 88 pp 897– (1991) |
| [45] | Traub, J. Physiol. 451 pp 653– (1992) · doi:10.1113/jphysiol.1992.sp019184 |
| [46] | Yao, Neural Networks 3 pp 153– (1990) |
| [47] | Rapp, Biologist 40 pp 89– (1993) |
| [48] | Grassberg, Phys. Rev. Lett. 50 pp 346– (1983) |
| [49] | Gencay, Physica 59D pp 142– (1992) |
| [50] | Kapitaniak, Chaos Solitons and Fractals 1 pp 67– (1991) |
| [51] | Brindley, Physica 51D pp 28– (1991) |
This reference list is based on information provided by the publisher or from digital mathematics libraries. Its items are heuristically matched to zbMATH identifiers and may contain data conversion errors. In some cases that data have been complemented/enhanced by data from zbMATH Open. This attempts to reflect the references listed in the original paper as accurately as possible without claiming completeness or a perfect matching.
