Energy Efficient Learning on Neuromorphic and Compute-in-Memory Architectures

Date of Award

8-15-2026

Degree Name

Ph.D. in Electrical and Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Tarek M. Taha

Abstract

The growing presence of artificial intelligence (AI) in edge devices, such as smart sensors, autonomous systems, and mobile platforms, has intensified the demand for energy-efficient, secure, and real-time on-device learning. Traditional AI architectures relying on cloud-based training suffer from several critical limitations including high latency, privacy vulnerabilities, and prohibitive energy consumption. This dissertation addresses these challenges by proposing a suite of hardware-software co-designed systems based on memristor-enabled neuromorphic and Compute-In-Memory computing (CIM) architectures that support online, adaptive, and federated learning directly at the edge. A major contribution of this dissertation is the development of an analog, extremely energy efficient anomaly detection and learning system tailored for edge security applications. The architecture utilizes dual autoencoder neural networks, one optimized for inference and the other capable of real-time adaptation, built on a memristive crossbar array. The design eliminates the need for large-scale digital computation and dataset storage by implementing sample-wise learning and analog signal processing. A fully analog module computes similarity metrics for anomaly detection, thereby avoiding traditional floating-point computation. This system enables on-device learning and personalization for intrusion detection, ensuring data never leaves the local platform. Extending these principles, this dissertation presents a memristor-based deep reinforcement learning (Mem-DRL) framework, which demonstrates the feasibility of on-chip training in dynamic environments. The architecture incorporates analog neural processing blocks and exploits the intrinsic parallelism of memristive arrays to execute learning and inference tasks efficiently. By reducing analog-to-digital conversion and supporting direct analog ReLU computation, the system significantly minimizes overhead. The Mem-DRL architecture demonstrates the potential of adaptive, self-learning agents that can operate entirely within edge hardware without dependence on cloud resources. To address privacy-preserving collaborative learning, this dissertation introduces the first known demonstration of a federated learning system implemented using memristor-based inmemory computing. In this system, edge devices independently train local models and communicate only weight updates to a central server, preserving data privacy and reducing network load. The local learning process is entirely hardware-implemented, leveraging analog matrix operations and memristive storage. The system maintains inference accuracy comparable to conventional software implementations while dramatically reducing energy and memory overhead. Its utility is demonstrated through its application in federated intrusion detection, though the architecture is generalizable to a wide array of personalized edge learning applications. The dissertation also presents a comprehensive review of edge AI processors, including dataflow, neuromorphic, and CIM architectures. This survey evaluates these platforms across performance metrics, fabrication processes, and supports AI workloads. The analysis reveals a clear industry shift toward emerging non-von Neumann architectures, particularly for applications requiring real-time, low-power computation. Neuromorphic and CIM-based systems are identified as especially promising for applications that demand continual learning and enhanced privacy. Finally, this research investigates on-chip training using analog CIM architectures for convolutional neural networks (CNN). Multiple configurations of SRAM- and RRAM-based CIM systems are explored, considering weight precision, ADC quantization, multiplexing strategies, and network topologies. Lightweight networks such as LeNet-5 and VGG variants are optimized for training tasks like image classification and cognitive signal (RF modulation) recognition. The analysis reveals critical design trade-offs and demonstrates that efficient training can be achieved with reduced hardware complexity and energy footprint. Furthermore, the integration of transfer learning into CIM systems significantly reduces training cost while preserving classification accuracy, offering a practical pathway for deploying deeper neural networks in constrained edge environments. In conclusion, this dissertation presents a unified framework for secure, adaptive, and energy-efficient AI computing at the edge through the development of novel analog and memristive architecture. By demonstrating online learning, reinforcement learning, and federated learning on hardware-constrained platforms, this dissertation is a significant contribution for developing next-generation intelligent edge devices capable of autonomy and resilience. The proposed systems pave the way for a new class of edge AI solutions applicable to security, healthcare, cognitive communication, and autonomous systems, with future directions including large-scale integration, cross-layer optimization, and hybrid analog-digital co-processing, multimodal communications.

Keywords

Artificial Intelligence, Computer Engineering, Design, Electrical Engineering

Rights Statement

Copyright 2026, author

Link to Full Text

Share

COinS