THIS PAGE INCORPORATES CASCADING STYLE SHEET! TO VIEW PAGE CORRECTLY PLEASE UPDATE TO A CSS COMPLIANT BROWSER
A synchronous firing model of LGN
UR > CS > Research > Vision > Publications > A synchronous firing model of LGN
pdf 128 KB
ps.gz 48 KB
ps 592 KB
@Article{,
  author = 	 {Dana H. Ballard and Zuohua Zhang},
  title = 	 {A synchronous firing model of LGN},
  journal = 	 {Computational Neuroscience},
  year = 	 {2000},
  introduction = {
Recently there has been a huge increase in experiments that imply that neurons communicate with syn-
chronous volleys. Evidence from slice recordings shows that neurons fire with millisecond reliability when
stimulated with realistic simulated input [MS95]. In the retina, the efficacy of synchronous codes has
been noted as a way of retinal encoding [Mei96]. Meister shows how retinal ganglion synchronous codes
trade of communication rate with increased signal delity. Reid, Alonso and Usrey have observed pre-
cise timing in cortico-thalamic connections[RA95, AmUR96, URR98]. Additional evidence that volleys
might be generated comes from experiments by [SJGW94]. In anesthetized cats, neuron outputs in the
LGN were correlated in the presence of the cortex, but uncorrelated (even though their firing rates were
undisturbed), when the cortex was removed. In a tour de force experiment, Casteo-Branco et al[CBNS98]
observed correlations between the retina LGN and cortical areas 17 and 18 in the anesthetized cat.
To interpret this data in a general way requires a theoretical framework. Recently progress has been
made with predictive coding models[OF97, Fie87, RB96, RB98]. The brain has to have a way of assuring
the usefulness of its representations; one check as to usefulness is whether or not the representations can
predict the current input. In such models the the synapses of cells are driven by the statistics of natural
scenes. Adjustment of synaptic strength is based on the ability of a set of neurons to predict its input.
Such models have been able to predict the distribution of cell properties such as receptive field size with
impressive accuracy. The critical feature of the predictive coding formulation is that the structure of
receptive fields can be predicted by assuming that the cortical memory costs metabolically in terms of
spikes and synapses. This allows the general interpretation that the cortex trades of the accuracy of its
computations with the amount of circuitry used to carry them out.
Predictive models have had a signicant drawback in that they relied on the classical rate code
interpretation of neural signaling. Our research extended predictive coding so that it could use volleys
of synchronous spikes instead of a rate code[BRZ99]. The studies reported herein extend this work by
showing that, in a detailed LGN model, the features of the extended model account for features observed
in experimental data, but provide a dierent interpretation: We think that temporal features of receptive
fields reflect the convergence of predictive coding rather than signaling motion features.}
}