Memristive In-Memory Computing for Resource-Constrained Visual-Temporal System
Date Issued
August 2026
Author(s)
Tushar, Sree Nirmillo Biswash
The University of Tennessee
Advisor(s)
Garrett S. Rose
Catherine D. Schuman
Additional Advisor(s)
Ahmedullah Aziz
Anahita Khojandi
Abstract
Resource-constrained computer vision systems must operate under strict power and memory limits, making traditional von Neumann architectures inefficient due to heavy data movement between sensors, memory, and compute units. This dissertation presents a hardware--algorithm co-design framework for visual temporal tasks that leverages memristive in-memory and in-sensor computing with sleep inspired algorithm to enable low-power, real-time visual processing at the edge.
At the device and circuit levels, this work analyzes memristive synapses, including current-controlled synapses, to improve read-current distinguishability, reduce power consumption, and enhance reliability under practical hardware non-idealities. Hardware-aware evaluation metrics, including Same Current Level (SCL), are introduced to quantify output-current distinguishability in memristive dot-product engines (MDPEs). To further improve efficiency, circuit- and architecture-level optimizations are developed, including refractory-period optimization, row pruning, dynamic-resolution DAC programming, and current-mode analog Winner-Take-All (WTA) circuits. These techniques reduce peripheral overhead, spike activity, and energy consumption while preserving classification performance.
The proposed framework is then extended to visual processing applications. A mixed-signal memristive in-memory architecture is developed for template matching using hardware-aware approximations of Pearson Correlation Coefficient (PCC) computation, reducing energy and delay compared with conventional processing. In addition, an in-sensor motion-recognition system is demonstrated, where analog signals from a light-sensing surface are mapped directly to a memristive crossbar and classified using current-domain decision circuitry, reducing data movement and supporting low-power spatio-temporal feature extraction.
Finally, this dissertation investigates adaptive temporal learning using memristive neuromorphic systems. A memristive reservoir computing framework is evaluated under device non-idealities, including input noise, synaptic variability, dot-product-engine variability, and conductance drift. A multi-conductance analog STDP circuit is explored to support adaptive synaptic learning and replay-driven consolidation, while system-level energy estimation is performed using a large-scale RAVENS neuromorphic architecture. Building on these components, a sleep-inspired wake--sleep learning framework is proposed to extract temporal context through replay-based memory consolidation without storing complete histories of past observations. Together, these contributions advance efficient, scalable, and adaptive visual temporal computing for resource-constrained edge applications.
Degree
Doctor of Philosophy
Major
Computer Engineering
Embargo Date
August 1, 2029