Jacob N. Allen

Electronic nose inhibition in a spiking neural network for noise cancellation

An olfaction detection spiking neural network that detects binary odor patterns is analyzed and implemented. This paper presents a new method for inhibiting spiking neural networks by modulating a detection threshold. Interference noise from active odors is measured by a single inhibitory neuron. The inhibition neuron changes the detection threshold to create tolerance for a system with multiple odors present. A digital implementation of the inhibition is simulated. Comparative results prove that threshold modulation reduces false-positive detection error in high noise scenarios where fifteen odors are active simultaneously.

An E-nose haar wavelet preprocessing circuit for spiking neural network classification

A simulation model for polymer film chemical sensors is developed based on a 1 dimensional diffusion equation. Using this model, electronic nose smell prints produced by the 32 sensor array of a Cyranose 320 are simulated to test pattern classification. A Haar wavelet Alter reduces noise and captures information about the diffusion rate of the analyte in each sensor. Inputs are encoded into a binary Hamming pattern and fed into a binary spiking neural network for pattern classification. The preprocessing circuit for the spiking neural network, including the wavelet Alter, is designed using standard cells for an 180 nm process. Real and simulated results from the spiking neural network classification algorithm are favorably compared to Bayes, canonical, and PCA-PNN classifiers.

A Low-Power Haar-Wavelet Preprocessing Approach for a SNN Olfactory System

A low frequency and low power spiking neural network chip is designed to classify polymer film electronic nose patterns. A simulation model for polymer film chemical sensors is developed and compared to other approaches. The final chip design uses wavelet pre-processing for fault tolerance and operates at just 20 kHz.

Spiking networks for biochemical detection

Artificial neural networks have proven to be a useful tool for olfactory pattern recognition; but most silicon-based implementations have been limited in scale due to inherent constraints on chip real estate and synapse routing. This paper presents a new spiking neural network approach to odorant learning and detection based on new learned information about the mammalian olfaction system. The network is designed to accept more than 1000 inputs and detect odors via an unlimited number of outputs. The basic theory is presented, including stochastic learning, and detection based on hamming distance. Simulation shows the network functions well, even in noisy environments where more than 10% of inputs are contaminated by background noise. Digital hardware implementation using VHDL shows that, a representative system with 128 inputs and 8 outputs fits on a single Xilinx Virtex v1000 chip and would occupy just 0.118 mm/sup 2/ using 0.16um CMOS technology.

Agile hardware development with rapid hardware definition language

Rapid HDL is a hardware definition language developed in C# for the Microsoft .Net platform. Rapid HDL unifies digital FPGA hardware development and co-software integration. Hardware is scripted using reusable software objects, communication between hardware and software is automatic, and synthesis is automated using a free tool chain. Rapid HDL represents 18,000 lines of code and is freely available under the MIT license.

A Compact Correlation Filter For On-Chip Learning in a Spiking Neural Network

A Hebbian learning algorithm based on proportion sampling is presented that can be used to implement on-chip learning for a binary spiking neural network. A correlation filter estimates when statistical independence has been obtained between subsequent samples. Simulation shows that the correlation filter reduces falsely learned connections in environments were inputs are randomly activated an average of 83% of the total time. A correlation filter for 255 binary samples is implemented using 21 gates and a surface area of .0008cm2 for a .5¿ fabrication process. Compared to traditional neural networks, the spiking neural network learned an odor in a single epoch resulting in only a 7% error, while classical learning algorithms required multiple epochs and typically resulted in 30% error.

Introducing RapidHDL: A New Library to Design FPGA Hardware in Microsoft. Net and Automatically Generate Verilog Netlists

RapidHDL is introduced as a new object oriented hardware description library for C# developers. RapidHDL seeks to speed up FPGA development by applying best practices used in software engineering to increase productivity. Logic is rapidly defined in component classes using a structure of truth-table definitions, sink nodes, source nodes, and pass-through nodes. Hardware simulation co-runs with C# programs, and a linked-list of clock events simulates propagation delays. A standardized testing framework allows the developer to write test benches, hardware, and software in a single language. Algorithms are presented to automatically transform RapidHDL objects to Verilog netlist that can be synthesized by calling 3rd party tools

Plasticity recurrent spiking neural networks for olfactory pattern recognition

This paper introduces a novel spiking neural network methodology, and applies it to an odorant learning and detection application. Spike-time dependent plasticity can support coding schemes that are based on spatio-temporal spike patterns. Spiking (or pulsed) neural networks (SNNs) are models which explicitly take into account the timing of inputs. The network input and output are usually represented as series of spikes (delta function or more complex shapes). Plasticity SNNs have an advantage of being able to recurrently process information. Spike-time dependent plasticity can enhance signal transmission by selectively strengthening synaptic connections that transmit precisely timed spikes at the expense of those synapses that transmit poorly timed spikes. Complete theory describes the spiking networks digital implementation. Theory is verified using simulation of a biologically plausible odor environment with 1023 odor receptor inputs. Stochastic on-chip learning uses sampling and a correlation filter to learn odors despite noisy environment conditions. Simulation and Verilog chip design are tested on a field programmable gate array. A scalable field programmable gate array implementation with 1024 inputs, 1 output, and expansion capability for unlimited outputs is synthesized into 87,062 gates

Plastic NNs for biochemical detection

Bio-inspired systems including neural networks (NNs) have proven to be a useful tool for biochemical electronic nose. Silicon embedded olfactory pattern recognition systems however, have been limited in scale due to inherent constraints of synapse routing and chip area. This paper presents a new plastic-based NN approach for a pseudo bloodhound nose for odorant learning and detection. The network uses spike driven plastic synapses, and is designed to accept more than 1000 inputs and detect odors via an unlimited number of outputs. The basic theory is presented, including stochastic learning and detection based on hamming distance. Simulation shows the network functions well, even in noisy environments where more than 10% of inputs are contaminated by background noise. Digital hardware implementation using VHDL shows that a representative system with 128 inputs and 8 outputs fits on a single Xilinx Virtex v1000 chip and would occupy just 0.118 mm/sup 2/ using 0.16/spl mu/m CMOS technology.

Sampling Spiking Neural Network electronic nose on a tiny-chip

Chemicals classification using a new Sampling Spiking Neural Network (SSNN) approach is presented in this paper with experimental measurements using the Cyranose 320 sensor array. The network is unique in its minimal yet powerful design which implements on chip learning and parallel monitoring to detect binary odor patterns with high noise environment. The SSNN architecture is further implemented on a 0.5 um CMOS technology tiny-chip designed to work in conjunction with a 256K external SRAM memory. It handles the routing of spike signal among 32,000 synapses and 255 neurons. At the same time, it tracks and records learning statistics. The chip can be used in parallel with other SSNN co processors for very large systems. Experimental measurements of our SSNN E-Nose classifier, compared to other E-nose systems proved superior in capability, size, and correctness.