文摘
This thesis explores some of the more unusual architectures and devices being considered today as the basis for information processing, emphasizing architectures that are highly parallel and devices that are extremely small compared to current standards.;The first part of this thesis theoretically and numerically analyzes analog electronic neural networks in which competition within neuron clusters leads to pattern classification and feature extraction abilities. Global stability theorems, derived using a Liapunov approach, provide general guidelines for network design and operation. The theorems state that with continuous-time updating, competitive networks converge only to fixed points, while with discrete-time, parallel updating, they converge to either fixed points or period-two limit cycles. A stability criterion guarantees that discrete-time networks converge only to fixed points when a quantity related to the neuron gain, or transfer function slope, is sufficiently small.;A set of analytical phase diagrams for competitive associative memories is derived using a combination of statistical mechanics and nonlinear dynamics. The diagrams classify attractor types as a function of pattern storage fraction and neuron gain. Numerical tests agree well with the diagrams.;Analog annealing, a technique for improving network performance by reducing neuron gain, is shown to improve performance in an analog associative memory by dramatically reducing the number of fixed points. The number of fixed points decreases exponentially with network size with a scaling exponent that decreases with neuron gain. Numerical data based on fixed-point counts in small networks support the results.;The second part of this thesis discusses low-temperature tunneling measurements at zero magnetic field through double and triple quantum dots with adjustable inter-dot coupling, fabricated in a GaAs/AlGaAs heterostructure. The devices have capacitances so small that the charging energy of adding an electron is much greater than the thermal energy at dilution refrigerator temperatures. The measurements, which explore how changing inter-dot coupling affects device conductance, are important for quantum dots used as "artificial atoms" or as "single-electron transistors" in larger arrays.;For single quantum dots, single-electron charging leads to dramatic conductance peaks. For arrays of two and three quantum dots, the conductance peaks each split into two (double dot) or three (triple dot) peaks as the inter-dot coupling increases. The splitting closely tracks the measured tunnel conductance and experimentally determines the interaction energy. Coupled double and triple dots with different gate capacitance show quasiperiodic beating. Monte Carlo simulations of a classical capacitive charging model qualitatively reproduce the observed structure, even though the underlying splitting mechanism is most likely quantum mechanical.