Summary form only given. The possibility of proposing models that have temporal characteristics with a view toward electronic circuit implementation is investigated. Similar to the Hodgkin-Huxley membrane model, a mod...
详细信息
Summary form only given. The possibility of proposing models that have temporal characteristics with a view toward electronic circuit implementation is investigated. Similar to the Hodgkin-Huxley membrane model, a model using Josephson junction circuits can be developed. The advantage of the Josephson junction model lies in its high speed and low power dissipation properties. In addition, such a model would have rich temporal dynamical behavior and is suitable for large scale implementation.< >
A prototype two-layer feedforward artificialneural network (FANN) is implemented using standard CMOS VLSI technology. A simple tunable analog scalar/vector multiplier is designed and used to implement FANNs with lear...
详细信息
A prototype two-layer feedforward artificialneural network (FANN) is implemented using standard CMOS VLSI technology. A simple tunable analog scalar/vector multiplier is designed and used to implement FANNs with learning. A modified learning rule is used as a circuit-implementable learning rule for FANNs. Two sequential learning circuits are designed and extensively simulated using the PSPICE circuits simulator. A modular design is proposed for a large-scale implementation of FANNs with learning. A 4*1 module is designed using the MAGIC VLSI editor and has been fabricated via MOSIS on Tinychips. The module chips can be connected vertically and horizontally to realize a large-scale FANNs with optionally using on-chip learning circuit or off-chip learning capability.< >
Custom analog CMOS VLSI components have been designed to allow the construction of neuralnetworks with arbitrary architectures. A wide-range, all-enhancement-mode, MOS analog four-quadrant multiplier has been employe...
详细信息
Custom analog CMOS VLSI components have been designed to allow the construction of neuralnetworks with arbitrary architectures. A wide-range, all-enhancement-mode, MOS analog four-quadrant multiplier has been employed to implement the scalar vector product of the vector of neuron outputs and the vector of the corresponding weights. A nonlinear MOSFET floating element is used to model nonlinear conductive elements; this extends the modeling of the synaptic weights to nonlinear elements. A neuron is realized by a simple CMOS operational amplifier which is compatible with the I/O of the analog multiplier. The neural system can be expanded modularly to large dimensions. SPICE simulations demonstrate the functionality of the all-MOS feedback neural nets using a two-neuron prototype as an example.< >
Models for feedback artificialneural nets (ANNs) which are shown to have qualitatively the same dynamic properties as gradient continuous-time feedback neural nets are presented. These models are based on biological ...
详细信息
Models for feedback artificialneural nets (ANNs) which are shown to have qualitatively the same dynamic properties as gradient continuous-time feedback neural nets are presented. These models are based on biological neural nets where neurons have dendrodendritic connections i.e. where connections among neurons occur via dendrites only. These models have the maximum number of connections equal to n (n+1)/2, where n is the number of neurons. The synaptic weights are naturally symmetric. One model uses nonlinear weights are naturally symmetric. One model uses nonlinear floating MOSFET transistors for its dendritic connection, where its conductance is controlled via the gate voltage. This last model lends itself naturally to analog all-MOS VLSI implementation.< >
An all-MOS circuit realization for a feedforward artificialneural network is described. An all-MOS realization of a modified learning rule is introduced. In addition to analytical verification the modified learning r...
详细信息
An all-MOS circuit realization for a feedforward artificialneural network is described. An all-MOS realization of a modified learning rule is introduced. In addition to analytical verification the modified learning rule is shown, via computer code as well as SPICE simulations, to successfully store into the network any given analog values (within the permissible range). An all-MOS architecture for a prototype two-layer artificialneural network is specifically tested via SPICE simulations. The results demonstrate the learning capability of the all-MOS circuit realization and establish a VLSI modular architecture for composing a large-scale neural network system.< >
暂无评论