Nikhil Chowdary Paleti

In this comprehensive project, we developed a real-time sales management system with three main components:

• Real-Time Data Streaming: Implemented a system to stream shopping invoices from client to database using MQTT, Kafka, and Spark.
• Data Analysis (EDA): Performed exploratory data analysis on the collected data using Spark, gaining insights such as top customers, total sales, and busiest days.
• Recommendation System: Developed a product recommendation system using Alternating Least Squares (ALS) algorithm to provide personalized suggestions to customers.

Key features:

• Used MobileNetSSD for object detection in the checkout process
• Implemented data streaming architecture with MQTT Broker, Kafka, and Spark Streaming
• Stored data in MongoDB Atlas for efficient retrieval and analysis
• Conducted various analytical queries to extract meaningful insights from the sales data
• Preprocessed data and built an ALS model for the recommendation system

This project demonstrated our ability to work with big data technologies and apply machine learning in a real-world retail scenario.

In this project, we developed an autonomous differential drive robot:

• Built a robot with Arduino Uno, L298N Motor Driver, DC Motors, and Nvidia Jetson Nano.
• Implemented differential drive control for flexible movement and turning.
• Utilized ROS (Robot Operating System) for overall robot control and communication.
• Integrated a deep neural network (ResNet18) optimized with TensorRT for object detection.
• Implemented object avoidance logic using OpenCV and the neural network outputs.

This project combined hardware assembly, low-level motor control, and high-level AI-driven decision making to create a fully autonomous robot capable of navigating and avoiding obstacles.

In this project, we explored the limitations of traditional Principal Component Analysis (PCA) and implemented Robust Principal Component Analysis (RPCA) to overcome these limitations:

• PCA Limitations: Demonstrated how standard PCA is not robust to outliers, affecting its performance in real-world scenarios with noisy data.
• RPCA Implementation: Implemented RPCA by framing it as an Alternating Direction Method (ADM) problem, providing a more robust approach to dimensionality reduction.
• Yale Face Dataset Application: Applied our RPCA implementation to the Yale face dataset, showcasing its effectiveness in image processing tasks.
• Noise Introduction and Denoising: Introduced various types of noise to the dataset, including:
  • Random patches
  • Salt and pepper noise
  • Other artificial distortions
• Comparative Analysis: Demonstrated the superior performance of RPCA in denoising these corrupted images compared to traditional PCA.
• Mathematical Foundation: Derived and implemented the mathematical foundations of RPCA and the Alternating Direction Method.

This project deepened our understanding of robust dimensionality reduction techniques and their practical applications in image processing and denoising. It showcased the power of RPCA in handling real-world data with outliers and corruption, providing valuable insights into advanced data analysis methods.

In this comprehensive project, we explored and implemented a range of reinforcement learning algorithms, from classic to cutting-edge:

• Q-learning: Implemented this foundational value-based algorithm, understanding the concept of off-policy learning and the balance between exploration and exploitation.
• SARSA (State-Action-Reward-State-Action): Developed this on-policy TD control algorithm, comparing its performance with Q-learning in various environments.
• DQN (Deep Q-Network): Extended Q-learning to deep learning, implementing experience replay and fixed Q-targets to stabilize training in complex environments.
• REINFORCE with baseline: Implemented this policy gradient method, incorporating a baseline to reduce variance and improve learning stability.
• PPO (Proximal Policy Optimization): Implemented this state-of-the-art policy optimization algorithm, known for its sample efficiency and reliability.
• DDPG (Deep Deterministic Policy Gradient): Developed this actor-critic algorithm for continuous action spaces, combining insights from DQN and deterministic policy gradients.

For each algorithm, we:

• Derived the mathematical foundations and update rules
• Implemented the algorithm from scratch in Python
• Tested and benchmarked performance on various OpenAI Gym environments
• Analyzed learning curves, policy behavior, and algorithm trade-offs

This project deepened our understanding of RL fundamentals, policy and value-based methods, and advanced techniques in deep reinforcement learning, providing hands-on experience with algorithms at the forefront of AI research.

In this project, we delved deep into the fundamentals of neural networks:

• Implemented a single-layer neural network from scratch, understanding the basic building blocks of neural computation.
• Extended the implementation to a multi-layer neural network, exploring the power of deep architectures.
• Derived and implemented the backpropagation algorithm for both single and multi-layer networks, gaining insights into gradient-based learning.
• Introduced the concept of implicit layers, exploring advanced neural network architectures beyond traditional explicit layers.

This project provided a strong foundation in neural network theory and implementation, paving the way for understanding more complex deep learning models.