Computational properties of single cells recorded in the brain reflect the dynamics of the network in which their are imbedded. Several properties of networks are of particular interest: i) Causes of response latencies, ii) Memory properties, iii) Synchronization and iv) Transfer function.
1. Latencies are an important characteristics of processing in the brain. Determining their causes may reduce the number of hypotheses about the computations performed. Latencies are due to axonal and synaptic transfer time, dendritic transfer time and somatic integration time. In the case of very short EPSPs and parameters set for the multiplication mode (Bugmann, 1991, 1992), the soma shows no memory of past inputs. Thus the firing time is determined by the moment of coincidences of input spikes. In that case, the latencies due to random timing of coincidences have been determined analytically and verified by simulations (Bugmann and Taylor, 1996)
2. The results show that latencies are determined depend on the input firing rates in the case of small input firing rates and are mainly determined by the jitter of the input neurons in the case of high input firing rates. In that later case, the jitter in the very first neurons in a multilayer network can determine the interlayer latency in all subsequent layers.(Bugmann and Taylor, 1993, 1994a, 1995a)
3. In the pyramidal network used to analyse latencies, neurons were supposed to exhibit persistent firing in response to input coincidences. This enabled to process asynchronous inputs (Bugmann, 1997). It also required some control of the duration of the sustained firing. This was proposed in form of inhibitory feedback which silenced its input neurons once it had itself started firing. It turned out that this feedback induced synchronization in the network in a way that made much more sense that the classical model based on lateral connections (Bugmann and Taylor, 1994b, 1995a).
4. The resetting of sustained firing by feedback provided also an interesting mechanism for backward masking which forms the basis of a model that fits well experimental data (Bugmann and Taylor, 1994a, 1995a, 1996)
5. How do properties of a detailed neuron model affect persistent firing in a cluster of neurons with recurrent connections ? We have found that the duration of the sustained firing is decreased when the synaptic transmission probability is small and increased when the synapses are located distally on the dendritic tree (long EPSCs). Further, the cluster shows a gradual reduction in activity with distal recurrent connections and a more binary behaviour (firing at a high rate and rapid transition to a silent mode) with proximal synapses (Bugmann, 1996a)
6. The recurrent cluster can be trained to "recognize" a small set of input patterns. The presentation of a learned patterns leads then to a strong response which persists after end of the input. The presentation of a unknown patterns results in a weak response that decays rapidly after end of the input (Bugmann, 1996b). We have trained the "lateral" weights with a classical hebbian rules, similarly to those used in Hopfield networks to learn attractors. Due to the decaying nature of the attractor, caused by probabilistic synapses, the network becomes a filter, transmitting certain input patterns and blocking others. In a sense, we observe the "transfer function" of a Hopfield network.
7. In general it is noted that all these functions depend on the details
of the parameter of individual neurons. Therefore, details of neuron models
cannot be ignored when "average" or "global" abstract network models are
designed to simulate biological systems (Bugmann, 1996a).
1. What is the actual contribution of jitter to visual latencies ?.
2. Temporal computation in the Cortex.
"A model for latencies in the visual system"
Bugmann, G. and Taylor J.G. (1993)
in proc. 3rd Conf. on Artificial Neural Networks (ICANN'93, Amsterdam), Gielen S. and Kappen B. (eds), p.165-168.
of short-term memory in visual information propagation" (ps.ZIP 62K)
Bugmann, G. and Taylor J.G. (1994a)
Extended Abstract Book of Int. Symp. on Dynamics of Neural Processing, Washington, 132- 136
top-down model for neuronal synchronization" (PDF 59K)
Bugmann, G. and Taylor J.G. (1994b)
Research Report NRG-94-02, School of Computing, University of Plymouth, Plymouth PL4 8AA, UK.
"Modelling visual latencies, masking and synchronization"
Bugmann G. and Taylor J.G. (1995a)
Brain Research Association Abstracts, 12, p. 68. bra95_ps.zip (6623)
"Theory of visual information propagation"
Bugmann, G. and Taylor, J.G. (1995b)
Plausible Neural Computation" (PDF 150K) (ps.ZIP
Bugmann G. (1996a)
Biosystems, 40, pp. 11-19.
"Stochastic activation propagation in a multilayer network of recurrentb
clusters of spiking neurons"
Bugmann G. (1996b)
Invited talk at the Workshop on Neural Dynamics and Pattern Recognition (DYNN'96), Toulouse, March 12-13.
"Towards a model of visual backwards masking "
Bugmann, G. and Taylor J.G. (1996)
"A stochastic short-term memory using a pRAM neuron and its potential
Bugmann, G. and Taylor J.G. (1997)
in Beale R. and Plumbley M.D. (eds) "Recent advances in neural networks", to be published by Prentice Hall.
by Synchronisation: A Task Dependence Hypothesis" (ps.ZIP 13K)
Bugmann G. (1997)
Brain and Behaviour Sciences, 20, pp. 685-686.
a Neural Model of Timing" (PS.ZIP 57K) (PDF
Bugmann, G. (1998)
In Biosystems, 48, pp. 11-19. (ISSN 0303-2647)
"Modelling Relative Recency Discrimination Tasks using a Stochastic Working Memory Model" (ps.zip 161.645) (PDF 73,775)
Bugmann G. and Bapi R.S. (2000)
Biosystems 58:1-3, pp. 195-202. (ISSN 0303-2647)
Please comment to email