Semester of Graduation
Summer 2026
Degree Type
Thesis
Degree Name
Masters in Computer Science
Department
Computer Science
Committee Chair/First Advisor
Sahidul Islam
Second Advisor
Kazi Aminul Islam
Third Advisor
Abdullah Al Hafiz Khan
Abstract
Deep neural networks are increasingly required to run on the devices that generate the data. If such a device must perform more than one task, the standard practice is deploying one model per task, which makes memory grow linearly with task count, which is unacceptable when the entire budget is kilobytes. This thesis asks one question in three settings: how much capability can a network acquire without incurring deployment cost?
The first study takes an ImageNet-pretrained ResNet-18, sweeps the branch point across every residual stage and the classification-head depth across one, ten, and twenty layers, and deploys the resulting multi-head model to an NVIDIA Jetson Orin Nano under INT8 post-training static quantization. The second study takes a 2,775-parameter one-dimensional CNN that jointly performs ventricular arrhythmia detection and arrhythmia subtype classification from intracardiac electrograms, and deploys it to an STM32F303K8 microcontroller. The third extends that detector with approximate Bayesian inference, reusing its existing dropout layers to produce a predictive-uncertainty estimate and a clinician-deferral policy.
Read together, the studies show that the memory saved by sharing depends on the branch point only through the parameter mass it captures, while the accuracy cost depends on task relatedness. When tasks are disjoint these pressures conflict and an intermediate branch is Pareto-optimal. When tasks are hierarchically nested they align, the deepest branch is simultaneously cheapest and most accurate, and the auxiliary structure can be discarded at export. Capability introduced during training need not be carried into deployment.
Comments
I have updated all the errors in my thesis according to the checklist. The only thing missing is the signed Defense Form, which I haven't received from my supervisor yet, since he is travelling for a conference. Once I have it I will submit it as soon as possible.