Embedded Systems and Machine Learning

Course | Master of Science in Electrical Engineering
Semester: Fall 2024
Instructor: Dr. Sudeep Pasricha

Course Overview

Machine learning is rapidly transforming embedded systems—from smartphones and IoT devices to autonomous vehicles. This course focused on designing efficient and optimized deep learning models for embedded platforms. We explored cutting-edge research and implementation techniques that bridge the gap between ML theory and real-world, resource-constrained environments.

Key areas of focus:

• Deep learning fundamentals and architectures (DNNs, CNNs, RNNs)

• Software optimization for low-power and memory-limited platforms

• Hardware acceleration using specialized architectures

• Hardware-software co-design techniques

• Emerging paradigms: processing-in-memory, photonics, memristors

• Hands-on projects using Python, simulation tools

• Deep Learning Architectures

  • Fundamentals of Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs)

  • Training pipelines, forward/backpropagation

  • Techniques to mitigate overfitting: dropout, regularization

• Software Optimization Techniques

  • Model compression: quantization, pruning, and weight sharing

  • Memory-efficient inference with reduced precision arithmetic

  • Latency, power, and memory profiling tools and methods

• Hardware Acceleration for ML

  • Efficient deployment using GPUs, TPUs, FPGAs, and custom ASICs

  • Hardware-aware model tuning and co-design

  • Real-time ML inference and throughput optimization

• Sequence and Time-Series Processing

  • Recurrent models: RNNs, LSTMs, GRUs

  • Applications in speech recognition, sensor analytics, and control systems

  • Optimization of sequential models for embedded platforms

• Unsupervised Learning Methods

  • Clustering techniques (e.g., K-means, DBSCAN)

  • Dimensionality reduction (e.g., PCA, t-SNE)

  • Pattern discovery and feature learning without labels

• Anomaly Detection and Embedded Security

  • Real-time outlier detection for edge devices

  • Machine learning for intrusion and fault detection

  • Building trust and robustness into ML-powered systems

• Generative Modeling

  • Autoencoders and Variational Autoencoders (VAEs)

  • Generative Adversarial Networks (GANs) for synthetic data

  • Applications in data augmentation and model compression

• Advanced Embedded ML Techniques

  • Lightweight architectures: MobileNet, SqueezeNet, TinyML

  • Neural Architecture Search (NAS) and knowledge distillation

  • Balancing trade-offs: accuracy, speed, and hardware efficiency

• Reinforcement Learning

  • Policy-based and value-based methods (e.g., Q-learning)

  • Real-world applications in embedded control and automation

  • Designing efficient RL agents under constrained resources

• Emerging Computing Paradigms for ML

  • Processing-in-memory and near-memory computing

  • Memristor-based learning architectures

  • Photonic computing and neuromorphic systems

  • Future directions in machine learning hardware

Technical Scope

Project : Real-Time Recyclable Waste Classifier

The goal of this project was to develop a real-time machine learning system capable of classifying recyclable and non-recyclable waste using image data. By leveraging transfer learning and lightweight object detection models like SSD MobileNet V2, the system was designed for deployment on resource-constrained embedded platforms such as the Raspberry Pi. The broader objective was to automate waste sorting processes in environments like dining facilities, reducing manual labor and improving recycling efficiency.

Final Reflection

This course significantly strengthened my ability to design and implement resource-efficient machine learning solutions by bridging the gap between high-level ML theory and low-level hardware-aware optimization. Through hands-on labs, technical readings, and the final project, I developed a deeper understanding of how to adapt and deploy ML models within the strict performance and energy constraints of embedded systems. I learned to think critically about the full stack—from algorithm selection and data handling to model compression techniques and hardware acceleration strategies. Importantly, the course also introduced me to cutting-edge research in emerging compute paradigms such as processing-in-memory, neuromorphic computing, and photonics, which broadened my perspective on where embedded ML is headed.