With the rapid advancement of neural networks, their adoption across diverse domains and service infrastructures has significantly increased. In practical deployments, pre-trained deep neural networks (DNNs) are often fine-tuned to meet task-specific ...
With the rapid advancement of neural networks, their adoption across diverse domains and service infrastructures has significantly increased. In practical deployments, pre-trained deep neural networks (DNNs) are often fine-tuned to meet task-specific objectives and subsequently compressed to enable efficient execution on mobile or edge devices. However, not all users or developers have access to abundant computational resources, and such deployment processes frequently operate under constrained environments where both computational and data resources are limited. Accordingly, there is a growing need for a systematic methodology that enables the efficient adaptation and deployment of neural networks under resource-constrained conditions.
This dissertation investigates methods for neural network deployment in two primary aspects: fine-tuning and model compression. In the fine-tuning phase, the major limitation arises from the scarcity of high-performance GPUs, as modern DNNs demand substantial memory capacity and computational throughput. In the deployment phase, the principal limitation lies in the absence of calibration datasets required for accurate quantization. To address these challenges, this dissertation introduces two frameworks, FALQON and MimiQ, which respectively mitigate the fine-tuning inefficiency and data scarcity issues in resource-limited environments.
For the fine-tuning stage, this dissertation introduces FALQON, a framework that leverages FP8 quantization for LoRA-based fine-tuning while significantly reducing quantization overhead on consumer-level GPUs. Low-bit floating-point (FP) formats such as FP8 provide substantial acceleration and memory savings during model training owing to native hardware support on modern GPUs and NPUs. However, analysis reveals that FP8 quantization yields considerable speedup primarily for large-dimensional matrix multiplications, whereas quantization overhead dominates in low-rank adaptation (LoRA), which involves small-dimensional matrices for parameter-efficient fine-tuning of large language models (LLMs). To address this limitation, FALQON eliminates redundant quantization overhead by directly merging LoRA adapters into an FP8-quantized backbone during fine-tuning. The forward and backward computations for merged adapters are reformulated to further minimize overhead, and a row-wise proxy update mechanism is introduced to efficiently incorporate substantial updates into the quantized backbone. Experimental evaluations demonstrate that FALQON achieves approximately a threefold increase in training speed compared to existing quantized LoRA methods while maintaining comparable accuracy, offering a practical solution for efficient large-scale model fine-tuning.
For the deployment stage, this dissertation presents MimiQ, a data-free quantization framework tailored for vision transformer (ViT) architectures. Data-free quantization (DFQ) aims to construct a lightweight network from its full-precision counterpart without access to the original training data, often by utilizing synthetic samples. Although several DFQ methods have been proposed for ViTs, their performance degrades significantly under low-bit quantization settings. Examination of existing approaches reveals that synthetic samples frequently produce misaligned attention maps, whereas real samples exhibit strong alignment. Based on this observation, MimiQ enhances quantization performance by enforcing alignment of attention maps between synthetic and realistic representations. Specifically, synthetic data are generated by aligning head-wise attention outputs across spatial query patches, and head-wise structural attention distillation is applied to align the attention maps of the quantized model with those of its full-precision teacher. Experimental results demonstrate that MimiQ substantially outperforms prior DFQ methods, establishing a new state of the art for data-free quantization of ViTs.
Collectively, the proposed frameworks establish a unified methodology for efficient neural network adaptation and deployment under limited computational
and data resources. FALQON enables practical fine-tuning of large-scale models by eliminating quantization overhead in low-rank adaptation, while MimiQ
facilitates accurate deployment without access to original datasets through data-free quantization of vision transformers. Together, these contributions
address critical challenges spanning both the training and deployment phases of modern deep learning systems. The presented approaches advance the feasibility
of deploying high-performance neural networks on resource-constrained platforms, contributing to the broader goal of sustainable and accessible artificial
intelligence.