Project Background:
The lightweight deep learning models for resource-constrained devices address the critical challenge of deploying artificial intelligence (AI) capabilities in environments with limited computational resources. With the increasing prevalence of edge computing, IoT devices, and mobile technologies, there is a growing demand for efficient AI solutions operating on devices with constrained processing power, memory, and energy resources. This project mainly aims to leverage the advancements in lightweight deep learning models characterized by reduced model complexity and optimized architectures to enable the deployment of AI functionalities in scenarios such as mobile devices, IoT sensors, and edge computing nodes. The primary objectives include achieving a balance between model efficiency and task performance, exploring automated model design techniques, and addressing the unique challenges associated with real-time inference, energy efficiency, and adaptability to dynamic conditions.
Problem Statement:
- As the deployment of AI expands to edge computing, IoT devices and mobile platforms prove to be impractical due to their high computational demands.
-
The challenge lies in developing models that are accurate, lightweight, and capable of operating seamlessly on devices with constrained processing power, memory, and energy resources.
-
Address the challenge by investigating and innovating in lightweight deep learning models, seeking solutions that optimize model architectures, reduce memory footprints, and ensure energy-efficient inference.
-
The problem underscores the critical necessity to bridge the gap between sophisticated AI capabilities and the constraints imposed by the computing resources available on devices at the edge, contributing to the advancement of AI deployment in real-world, resource-limited scenarios.
Aim and Objectives:
- To develop and optimize lightweight deep learning models tailored for resource-constrained devices, focusing on enabling efficient and accurate artificial intelligence applications.
-
Develop model architectures and optimization techniques to significantly reduce computational demands and memory footprints on devices with limited resources.
-
Achieve low-latency and real-time inference capabilities in lightweight models deployed on resource-constrained devices.
-
Design models that minimize energy consumption during inference, contributing to prolonged battery life and energy efficiency.
-
Investigate automated methods for designing lightweight model architectures of adapting models to resource constraints and tasks.
-
Develop techniques for privacy-preserving on-device processing for data transmission to external servers and ensuring sensitive information.
-
Ensure lightweight models are scalable across various resource-constrained devices adopted in edge computing environments.
Contributions to Lightweight DL Models for Resource Constrained Devices:
- Development of novel model architectures optimized for resource efficiency, reducing computational demands and memory footprints without compromising task performance.
-
Enhanced adaptability of lightweight models to dynamic conditions, ensuring robust performance across varying scenarios and input variations.
-
Techniques for privacy-preserving on-device processing, minimizing the need for external data transmission and ensuring the security of sensitive information.
-
Solutions to ensure the scalability of lightweight models across a diverse range of resource-constrained devices, promoting widespread adoption in various edge computing environments.
-
Improved generalization capabilities of lightweight models, allowing them to perform effectively across a spectrum of tasks without extensive task-specific fine-tuning.
Deep Learning Algorithms for Resource-Constrained Devices:
- MobileNet
-
SqueezeNet
-
TinyYOLO
-
ShuffleNet
-
ESPNet
-
Binarized Neural Networks (BNNs)
-
Firefly
-
HarDNet
-
MCUNet
-
Posenet
Datasets for Resource Constrained Devices:
- CIFAR-10
-
MNIST
-
Tiny ImageNet
-
Speech Commands
-
Edge Impulse Datasets
-
Intel Image Classification
Performance Metrics for Resource Constrained Devices:
- Accuracy
-
Inference Time
-
Energy Consumption
-
Memory Footprint
-
Precision
-
Recall
-
F1 Score
-
Privacy Metrics
Software Tools and Technologies
Operating System: Ubuntu 18.04 LTS 64bit / Windows 10
Development Tools: Anaconda3, Spyder 5.0, Jupyter Notebook
Language Version: Python 3.9
Python Libraries:
1. Python ML Libraries:
- Scikit-Learn
- Numpy
- Pandas
- Matplotlib
- Seaborn
- Docker
- MLflow
2. Deep Learning Frameworks:
- Keras
- TensorFlow
- PyTorch
