University of Illinois Chicago
Browse

Efficient Neural Network Inference and Training Using Early Exit Strategies

Download (6.23 MB)
thesis
posted on 2024-12-01, 00:00 authored by Alperen Gormez
This work aims to reduce the inference and training costs of deep learning models by utilizing early exit networks. In particular, we introduce four algorithms: 1. E2CM, a simple and lightweight early exit algorithm that reduces the inference cost. In a separate line of work, we also show how early exit networks can be combined with model pruning. 2. CBT, an algorithm to further decrease the inference cost of early exit semantic segmentation networks. 3. EEPrune, a novel dataset pruning algorithm that uses early exit networks to reduce training cost. 4. Class-aware EE LLM, a novel weight initialization algorithm for early exit large language models to accelerate pre-training.

History

Advisor

Erdem Koyuncu

Department

Electrical and Computer Engineering

Degree Grantor

University of Illinois Chicago

Degree Level

  • Doctoral

Degree name

PhD, Doctor of Philosophy

Committee Member

Natasha Devroye Abolfazl Asudeh Mesrob Ohannessian Besma Smida

Thesis type

application/pdf

Language

  • en

Usage metrics

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC