Alyssa Apsel (Electrical Engineering) - Swarthmore College
SPATIO-TEMPORAL VELOCITY DETECTION IN TWO DIMENSIONS WITH A NEURAL NETWORK AND A SILICON RETINA
Advisors: Ralph Etienne-Cummings, Prof. Paul Mueller, Prof. J. Van der Spiegel
ABSTRACT: A short range algorithm for velocity detection in 2-dimension is implemented in analog hardware via a neural network. Data is received from a silicon retina chip and transferred as contrast invariant (edge detected) images of viewed obj ects to the neural network. Serial outputs from the retina chip are interfaced at clock rates of 25K frames per second through a board of shift registers to the parallel inputs of the neural network. The input lines are summed across x and y directions in order to examine each velocity component individually, and processed through 18 spatio-temporal feature extracting filters in both x and y directions designed to detect individual peak velocities. The temporal filters are designed to prevent signal delay . Data is sampled from the neurons over a range of velocities, and the output of each filter was plotted in tuning curves reflecting the maximum tuning velocity. Theoretical power spectrums based on the synaptic values and time constants used to implement the spatio-temporal filters are plotted over spatial and temporal frequency. The peaks of these plots are used to predict the theoretical velocities to which each filter is tuned. Theoretical and experimental values are then compared to reveal a close co rrelation. Tests of the 2-dimensional filter design to 45¡ motion also indicate effective performance in 2-dimensions.
Everton Gibson (Electrical Engineering) - Temple University
MICROFABRICATION OF A METAL-INSULATOR-METAL SENSOR FOR THE DETECTION OF HYDROGEN
Advisor: Dr. Jorge Santiago-Aviles
ABSTRACT: Research in the area of sensor technologies has led to many new discoveries about gas sensors. A hydrogen sensor has many significant applications in industrial facilities and the medical field. This sensor has the structure of a metal- insulator-metal capacitor. Its composed of palladium, silicon nitride and platinum. Platinum and palladium are catalytic metals in the dehydrogenation and dissociation of hydrocarbons and other hydrogen compounds. When exposed to hydrogen, palladium adsor ption of hydrogen atoms occurs. This reaction forms a dipole layer at the interface of the palladium metal and the silicon nitride insulation. The effect of this dipole layer is the formation of a measurable potential difference and a modification of the palladium work function.
Jennifer Healy-McKinney (Physics) - Widener University
MICROFABRICATION OF SILICON LAMP FOR METABOLISM MONITOR
Advisor: Dr. Jay Zemel
ABSTRACT: A silicon lamp consisting of a polysilicon beam surrounded by a vacuum enclosed in a polysilicon shell was built for the metabolism monitor. This device can be used as an infrared light source when a voltage is applied across the polysi licon beam. Part of measuring the metabolism of an individual is measuring the carbon dioxide concentration in the person’s breathe with an infrared sensor. The silicon lamp is necessary for the function of the IR sensor because of its ability to give off infrared light. The design of the silicon lamp consists of an etched region one micron deep into the silicon substrate. The silicon tub is 500 microns squared. A sacrificial layer one micron thick of oxide and phososilicate glass is placed above and belo w the polysilicon beam. A polysilicon beam is obtained that is 800 microns long, 100 microns wide, and 1 micron thick. A thin layer of silicon nitride is placed on top of the beam as well as between the beam filament and the substrate for insulation. A fi nal layer of polysilicon is deposited and acts as a shell to enclose the area. To create a vacuum around the polysilicon beam, the sacrificial layer is etched away with p-etch and hydrofluoric acid by entering through the sides of the substrate. These hol es are then closed by oxidation and the deposition of silicon nitride.
Peter Jacobs - Swarthmore College
THE NITRIC OXIDE SENSOR
Advisors: Prof. J. J. Santiago-Avils and M. Zeno
ABSTRACT: This research project centered around the design and fabrication of a microsensor capable of sensing micromolar concentrations of nitric oxide (NO) in solution. The support and conduction lines of the sensor were built using planar and photolithography technology in the Microfabrication Lab. The active part of the sensor is nickel (II) tetrakis (3-methoxy-4-hydroxyphenyl) porphyrin (Ni(II)TMHPP) which had to be deposited electrochemically in a sodium hydroxide solution using cyclic volt ammetry (CV). The sensor was tested in a solution of pH 7.4 phosphate buffer using differential pulse voltammetry (DPV). The results show that the porphyrin was deposited correctly onto the gold electrode and that the sensor did indeed show sensitivity to NO in the micromolar range.
Sang Yoon Lee (BE) - University of Pennsylvania
PATTERN RECOGNITION USING SOFTWARE SIMULATION OF THE NEURAL NETWORK
Advisor: Christopher Donham (EE)
ABSTRACT: This paper describes a software simulation of neural networks that was created to accomplish pattern recognition tasks. With flexibility in the number of layers and “neurons” – simple computing units – within each layer, the simulator u ses error back propagation to find the optimal final configurations of the neural net. The ultimate goal of the simulator would be the ability to incorporate the final neural net settings into the analog neural computer. As recognition tasks, two differen t types of patterns were used – digital and analog. The XOR and Parity problems were used for digital analysis, while vowel recognition was the goal behind analog analysis. The simulator was able to be trained for both types of inputs, and without certain physical limitations of the analog neural computer programmed into its learning process, the recognition rate was found to reach 100%. But as a complete simulator of the hardware net, recognition was achieved for 98% of the vowel training data. When the trained net (simulator) was tested with vowel articulations that were not used for its training, 61% of them were recognized.
Paul Longo - University of Pennsylvania
REAL TIME VELOCITY-BASED TARGET TRACKING WITH SILICON RETINA AND PC
Advisors: Ralph Etienne-Cummings, Dr. Jan Van der Spiegel and Dr. Paul Mueller
ABSTRACT: A velocity-based target tracking system using a pre-existing smart silicon retina chip was designed and constructed. A novel aspect of the approach taken lies in the inherent simplicity of the device, a feature due in large part to the on-chip edge detection, log compression, and Difference-of-Gaussian approximation abilities of the silicon retina. A DC motor driven retina-lens mount is linked to a PC and controlled via software written in C. The software was developed to generate contr ol signals for the retina chip, calculate target velocity based on chip output, and send a proportional voltage to the motor control circuit to drive the DC motor. The algorithm displays a linear output for targets moving with an on-chip velocity of up to 1.7 cm/s, and the system performs well in physical tracking experiments.
Laura Sivitz (Physics) - Bryn Mawr College
SYNTHESIZING FERROLELECTRIC CERAMICS FROM POWDER PRECURSORS
Advisor: Prof. Jorge Santiago-Aviles
ABSTRACT: The transducer for the ultrasonic mammographer being developed at Penn will be fabricated from the ferroelectric ceramic PZT. Preliminary research on the processing of ferroelectric ceramics has demonstrated that such ceramics can be re adily produced. Slip casting or spin casting are preferable to the processes described for fabricating a ferroelectric pellet because casting favors the growth of parallel crystal planes.
Zachary Walton (Physics) - Harvard University
AUTOMATIC REAL-TIME PHONEME RECOGNITION
Advisor: Chris Donham
Abstract: An analog neural network was implemented in hardware for use as a general programmable neural computer. To test the design of the hardware, it was programmed to extract features of the speech waveform for the purpose of phoneme identifi cation. Circuits were designed to detect the following features of the speech waveform: energy onset, energy offset, formant transitions, signal durations, energy amplitude, noise, and voicing. Using a digital computer, these circuits were programmed into the analog neural net. Continuous speech was played at the programmed neural net and the output was monitored by a digital computer. The output was analyzed, and the detection circuits were modified to more accurately extract the desired features. The ou tputs of the feature detection circuits were then combined to enable the neural net to distinguish individual phonemes.