Tappe Maestro, Rafael (2023) Function generation from a Sum of Oscillator Signals an Evaluation of Algorithms. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2023_TappeMaestroR.pdf Download (12MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (130kB) |
Abstract
Energy demand in data-intensive applications is rising. The human brain's energy efficiency far exceeds that of today's digital computers, prompting exploration into neuromorphic computing using analog devices. In this work, an ensemble of oscillators is designed and run in simulation with the goal of time-series approximation. Oscillator-neurons are formed by a vanadium dioxide memristor in series with a RC-circuit. Multiple gradient-free optimization algorithms are explored to perturb oscillator parameters. When a target's frequency band lies within the oscillators' frequency band, fit can be achieved as quantified by the Root Mean Squared Error (RMSE). The approximation of chirp signals is difficult, none of the real-world targets can be fit. The system benefits from broad phase and frequency diversity. Loss increases as dynamic range is increased for most algorithms. Surprisingly, increasing the number of oscillators tends to increase loss. The Las Vegas and random walk algorithms perform best, surpassing Simulated Annealing and Differential Evolution. We achieve our best RMSE of 0.02 with Las Vegas. Serving as a benchmark, linear regression outperforms this result by 13 orders of magnitude. While vanadium-dioxide oscillators are powerful, their frequency response makes them unsuitable for arbitrary function generation. Lastly, gradient-free algorithms might benefit from negative oscillator gains.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Schomaker, L.R.B. and Cipollini, D. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 10 Aug 2023 08:58 |
| Last Modified: | 10 Aug 2023 08:58 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/31138 |
Actions (login required)
 |
View Item |