Abstract

Neuromorphic computing is a neuro-inspired computing gaining traction because of its ability to perform complex calculations faster, with greater energy efficiency, and on a smaller footprint compared to traditional Von Neumann architectures. Due to how fundamentally different their architecture is from the Von Neumann architecture, there are currently significant uncertainties over how we program and use neuromorphic chips. When neuromorphic chip implementation is linked with the implementations employing emerging device technologies, additional challenges related to programming devices are introduced.
To address these issues, neuromorphic frameworks with the abstractions and tools to develop applications might prove beneficial. Lava framework is an open-source neuromorphic framework designed by Intel Labs to build applications that fully exploit the principles of neural computation and map them to neuromorphic hardware. The Lava framework includes high-level libraries for deep learning, dynamic neural fields, and constrained optimization for productive algorithm development. It also consists of tools to map those algorithms to different types of hardware architectures. Lava is the only existing neuromorphic framework that comes with specifically designed optimization solvers. Therefore, we have selected the Lava framework for our resource allocation problem implementation.
This thesis aims to conduct a case study on the use of Lava neuromorphic framework for radio resource allocation. First, we design and formulate the problem as an integer linear program (ILP). Then, it is reduced to a quadratic unconstrained binary optimization (QUBO), a format that can be solved using Lava. To evaluate the performance of the proposed Lava-based approach, we compare our results to the ones obtained with classical CPU-based solvers. Solving problems on Lava requires several hyperparameters as inputs. Currently, tuning the hyperparameters is a tedious task. Determining the optimal set of hyperparameters is even more difficult for large problem instances. We believe that improving the Solver-Tuner utility for hyperparameter tuning can help solve large problems.