Now showing 1 - 1 of 1
No Thumbnail Available
Publication

Ising Machines and Spiking Neural Networks: Non von-Neumann Computing using Networks of Coupled Oscillators

2021, Vathakkattil Joseph, George, 0000-0002-6757-0660

Computation has become synonymous with digital computation in the 21st century. The exponential growth in computational demand has been identical in trend to the exponential growth of compute resources available on a chip. However, recently the latter has started showing signs of slowing down as the physical limits are approached of the semiconducting substrates that have continually allowed transistor miniaturization for decades. This has brought into scrutiny the foundational architecture of general purpose computers established at the advent of the digital computer - the von Neumann architecture. Separating data and instructions, the von-Neumann architecture has been a simple yet highly successful model of computation. However, many classes of problems such as 'cognitive tasks' in image recognition, anomaly detection, speech recognition, etc. require very high throughput of data and are sub-optimally served by this architecture. Another class of problems is combinatorial optimization where a the optimal solution to a function with a large discrete configuration space is to be found. The exponential growth in the number of configurations even for moderate real-world problems makes this class among the hardest for traditional computers to solve. This thesis explores the potential of using coupled analog oscillators to construct non-von-Neumann architectures for such tasks where a significant advantage may be gained over traditional models. To solve combinatorial optimization, the concept of 'Ising Machines' has been proposed where dynamics of coupled oscillators is adapted to find the solution of the problem as the lowest energy state of a physical system. The Ising machine paradigm is studied using a general coupled oscillator model including amplitude dynamics for the first time. A control scheme called 'Parametric cycling' is proposed that prevents local minima traps to a significant extent. The performance is demonstrated on Max-Cut problems of fully connected graphs and cubic graphs. Further, drawing on the neuroscientific abstraction of neurons in the brain as coupled oscillators, a neuromorphic architecture is constructed suitable for anomaly detection problems. Spiking Neural Networks are proposed to solve structural health monitoring problems by extracting cepstral coefficients as features. The implementation is tested on a novel hardware platform named Intel Loihi, offering the potential of highly energy efficient operation of neuromorphic algorithms. This thesis showcases the computational potential of oscillators beginning with a single resonator, a piezoelectric energy harvester, which may be used to estimate extreme values of responses and fragility of structures. It progresses to coupled (Stuart-Landau) oscillators and demonstrates combinatorial optimization capability. Finally, coupled integrate-and-fire oscillators or spiking neurons are used to detect damage induced anomalies in the vibration response of a structure. The research presented in this thesis thus establishes coupled oscillator networks as prime contenders for computing primitives of the future in classes of problems where digital computers based on the von-Neumann architecture are fundamentally constrained.