Related Experiment Video
Updated: Sep 14, 2025
11:18
Closed-loop Neuro-robotic Experiments to Test Computational Properties of Neuronal Networks
Published on: March 2, 2015
10.4K
NACHOS: Neural Architecture Search for Hardware-Constrained Early-Exit Neural Networks
Summary
This study introduces NACHOS, a novel neural architecture search framework for designing efficient early-exit neural networks (EENNs). NACHOS automates the joint design of EENNs, optimizing accuracy and computational efficiency under hardware constraints.
Area of Science:
- Artificial Intelligence
- Computer Science
Background:
- Early-exit neural networks (EENNs) offer efficiency and effectiveness by allowing predictions at intermediate stages.
- Manual design of EENNs is complex, time-consuming, and requires expert knowledge for optimal configuration.
- Automating EENN design using neural architecture search (NAS) is an active research area.
Purpose of the Study:
- To present NACHOS, the first NAS framework for designing hardware-constrained EENNs.
- To enable the joint optimization of EENN backbone and early-exit classifiers (EECs).
- To satisfy constraints on accuracy and multiply-accumulate (MAC) operations.
Main Methods:
- Developed NACHOS, a NAS framework for joint backbone and EEC design.
- Incorporated hardware constraints (accuracy and MAC operations) into the search process.
- Identified Pareto optimal EENN solutions balancing accuracy and computational cost.
Main Results:
- NACHOS successfully designs EENNs that are competitive with state-of-the-art models.
- The framework provides a set of admissible Pareto optimal solutions.
- Investigated novel regularization techniques for auxiliary classifiers.
Conclusions:
- NACHOS offers a fully automated approach to designing efficient EENNs.
- The framework effectively balances accuracy and computational efficiency for hardware deployment.
- Future work includes further exploration of regularization strategies for EENN optimization.
