Graduation Year

2026

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Electrical Engineering

Major Professor

Sanjukta Bhanja, Ph.D.

Committee Member

Ismail Uysal, Ph.D.

Committee Member

Wilfrido Moreno, Ph.D.

Committee Member

Dayane Reis, Ph.D.

Committee Member

Ravi Panchumarthy, Ph.D.

Keywords

Conscious AI, Hardware Acceleration, Intelligent Systems, Nueromorphic Computing, Energy-Efficient Learning

Abstract

Artificial intelligence systems excel at narrow, well-defined tasks but remain brittle at theboundaries of their training distributions: they cannot quantify uncertainty, adapt continuously to non-stationary data, or operate efficiently on energy-constrained hardware. This dissertation addresses these limitations through four coordinated contributions spanning probabilistic learning algorithms, biologically inspired temporal memory, cross-entity warning propagation via distributed associative memory, and spintronic processing-in-memory.

The first contribution introduces Boosted Bayesian Neural Networks (BBNNs), which extend standard mean-field variational inference by iteratively constructing a mixture posterior through Boosting Variational Inference. On five clinical medical classification datasets, BBNNs achieve superior uncertainty calibration—lower Negative Log-Likelihood and Expected Calibration Error—on four of five datasets, with the most dramatic improvement on the Hepatitis dataset (17× NLL reduction). Predictive accuracy improves on datasets with high inter-class variability (Heart Statlog +4.84 pp, Cancer +1.50 pp, Diabetes +2.32 pp), at the cost of a 6.5× training time overhead acceptable for offline clinical model development.

The second contribution presents AHTM and H-AHTM, two accelerations of Hierarchical Temporal Memory (HTM) built on a Reflex Memory module that offloads repetitive firstorder temporal inferences from the full Sequence Memory. AHTM reduces per-prediction latency from 0.2868 ms to 39.93 μs (7.2× improvement); H-AHTM replaces the software dictionary with a 1-FeFET AFeCAM hardware array, achieving 2.65 ns per prediction— five orders of magnitude over the Sequence Memory baseline—while maintaining anomaly detection accuracy within statistical noise of baseline HTM across 15 financial datasets and 5 NAB benchmark streams.

The third contribution introduces the Distributed HTM (D-HTM) framework for crossentity warning propagation through inference-time precursor retrieval. A shared Spatial Pooler maps all monitored entities to a common SDR space; pre-anomaly activation windows are stored in a Shared Associative Memory (SAM) and retrieved via a K-of-L overlap condition without any cross-entity weight updates. Evaluated on the SMD, SMAP, and MSL multivariate benchmarks, D-HTM achieves warning F1 of 0.848 on SMD and 0.858 on MSL, with mean lead times approaching the full lookback horizon and recall exceeding 0.96 on all three real-world datasets.

The fourth contribution presents SPIMulator, the first cycle-level functional simulator for Domain Wall Memory-based Processing-in-Memory (DWM PIM), and the Near-Zero Multiplication Escape framework for sparse DNN inference. At threshold τ = 3, the Near- Zero Escape bypasses 66.38% of multiply-accumulate operations with only 0.76 pp accuracy loss on ResNet-18/CIFAR-10, reducing total cycles by 43.2% and energy by 43.9%.

Together, these contributions represent complementary advances toward intelligent machine design: calibrated probabilistic reasoning under uncertainty, ultra-low-latency temporal inference, distributed associative memory for cross-entity warning propagation, and energy-efficient in-memory computation for sparse AI workloads.

Download

Share

COinS