Diving into Machine Learning

 Machine learning has revolutionized industries, and will continue to drastically over the next few years as advancements are made. Systems being able to learn from data to make informed decisions allows for huge flexibility and customization.


Image Recommendation System

My goal for creating this application was to simulate algorithms commonly seen in social media platforms to enhance user engagement. This is done by recommending relevant images based on user preferences and interactions. 


Technical details: Collected user interaction data, including likes, comments, and hover times. Preprocessed 5,000+ images using Python and OpenCV for consistent input dimensions, while also normalizing pixel values.


The model architecture utilized was a pretrained ResNet50 for feature extraction of the images. This convolutional neural network (CNN) was important for embedding generation. 


I implemented this custom recommendation model using K-Nearest Neighbors, which I felt was best to calculate the similarity scores between image embeddings. 


Overall the system improved recommendation accuracy by 25%, and a drastic data preparation time reduction of 60%. 


Guava Health Detector

My second project opened up doors for me to train my own customized model. After lots of reading of tensorflow and keras documentation, I was able to create a customized model that could detect the health of a guava with high confidence based on a user inserted image. 


Technical details: I normalized all pixel values between the range of [0,1]. For augmentation I created a custom pipeline to introduce variability in the training data, this helped to substantially prevent overfitting. 


The model architecture consisted of a fine tuned MobileNetV2, which is a pre-trained convolution neural network (CNN). This model accomplished two key aspects, efficiency and extraction of high-quality features from images, which was vital considering the size of the dataset, the length of the training, and the importance of confidence in the user inputted images.


Model compiled with Adam optimizer, categorical crossentropy (good for multi-class classification), and accuracy for performance evaluation. 


Image Colorizer

This project was hard. From creating to testing to training. The goal of this project was unique, I wanted to be able to convert a grayscale image to color, and vice versa. I managed to find a perfect kaggle dataset for training. The approach I took was planned carefully and thoroughly ahead of time.


Technical details:

Generative Adversarial Networks (GANs): Useful for realistic outputs since they involve a generator (which was extremely handy), and a discriminator to evaluate the realism. 


U-Net for the generator: Skipped connections, and allowed for the low level details from the input to be transferred to output (preserves textures and edges).


A simple way to prevent overfitting through augmentation allowed for me to not require additional data. 


Comments

Post a Comment

Popular posts from this blog

Software Engineering Best Practices

Following software engineering best practices allows a developer to be efficient while also ensuring quality and replicable work. 1. Clean code Clean code is simply explained as easy to read and understand which does a couple things.  It allows for reduced time for current and future debugging, editing, and collaboration.  Collaboration is an important aspect, my first real introduction to team collaboration was in my last year of university  where a team of 9 of us worked on a project for a client. This allowed for us to practice applying real world  principles such as design, implementation, testing, and thorough IEEE documentation standards.  https://github.com/mnothman/HD-Requester Critical aspects that should be focused on include: Consistency : Using consistent naming and coding styles. Name styles include- CamelCase, PascalCase, snake_case,  while coding styles/conventions include PEP 8 for Python. Avoiding Hardcoding : Hardcoded values should be rep...
Read more

Netflix and On Demand Streaming

Something that I wanted to explore is how companies, specifically Netflix, deal with big data and have such quick on demand streaming.  Content Delivery Network (CDN)  Netflix created their own proprietary CDN named Open Connect, which is a Content Delivery Network. This CDN is a network of servers that are spread across the world to deliver movies quickly while also remaining efficient to scale well. A CDN is vital to reduce latency and improve performance by simply serving content from a server that is close to a user.  Open Connect stores Netflix’s content, where they are cached on multiple of these spread apart servers around the world. A simple way this works is: A user requests content, which is then delivered from the server that is nearest to the user's location (instead of being from a central server which is common practice but not optimal for streaming services). Edge Servers Netflix uses Open Connect Appliances (OCAs) which are its own servers that are deploye...
Read more

DevOps and CI/CD

DevOps and CI/CD pipelines are crucial for modern software development. DevOps are practices, tools, and standard philosophies that are integrated in software development to standardize development and IT operations in one. CI/CD pipelines are continuous integration and continuous delivery pipelines that are created to speed up the development lifecycle. CI/CD allows for developers to frequently merge code in shared repositories with each other and complete automated testing. The impact of such is that features and changes are pushed quickly while also maintaining a standard of quality through non time consuming testing.  -Continuous integration (CI) is the aspect that ensures that developers code changes are frequently integrated to create less major merge conflicts of codes, and allows for bugs to get caught early on.   -Continuous Delivery/Deployment prepares code changes for deployment through testing and deploying changes to production when ready.  The importanc...
Read more