Low Noise CMOS Active Mixers for Direct Conversion

- . When a nonlinear subneuron is integrated with every synapse, the above specified shape will be proportionally preserved for an -input neuron, regardless of the number of inputs. The resulting unified synapse-neuron block presents a highly modular and scalable solution for the design of VLSI multilayer neural networks with different sizes in different applications and has been successfully used in the implementation of programmable neural network classifiers as proof of concept [8]. A main conclusion in [2] is that increasing the number of nodes per layer in a conventional Madaline increases the required weight accuracy given a maximum allowable NSR (or minimum SNR). In this brief, it was shown that a hybrid distributed-neuron architecture is advantageous in terms of maintaining a better SNR as the number of neuron inputs (or nodes per layer) increases. The larger the network becomes, the more apparent the SNR advantage is, compared to a conventional Madaline network. A final note is that a higher SNR in a distributed-neuron architecture can be traded off at a certain level with a lower bit resolution. Depending on network topology, every 5‐10-dB difference in SNR (6 dB in an average sense) is equivalent to one bit difference in weight resolution. The quantization noise improvement discussed in this brief is a new property emerging in a hybrid (analog‐digital) distributed-neuron architecture and may not be directly applied to a fully analog implementation such as [3]. Besides the quantization noise improvement, a hardware implementation of the proposed architecture benefits from an averaging effect of distributed subneurons over silicon die [8], a property that reduces characteristic variations among different neurons of a sizable VLSI neural network and further contributes to a robust ANN architecture. VI. CONCLUSION This brief demonstrated that in distributed-neuron hybrid artificial neural networks the weight quantization error can influence less the output than in the case of lumped-neuron ANNs. The quantization noise was studied and, as a result, a stochastic model was presented for the first time for an Adaline with distributed-neuron implementation. The self-scaling property of a distributed neuron was formulated and applied to transform an existing model for a conventional (lumped-neuron) Adaline to the one presented for the distributed-neuron Adaline. Based on stochastic analysis and simulations, it was shown that the ratio of signal to quantization noise increases considerably for large networks (having larger number of neuron inputs, or nodes per layer) when a programmable neural network implementation is based on a distributed- rather than a lumped-neuron architecture.