Projects

Alzheimer's Disease (AD) is a debilitating neurological condition characterized by progressive cognitive decline and neurodegeneration. Early detection is crucial for timely intervention and treatment. This project focuses on leveraging advanced deep learning methods to analyze 3D brain Magnetic Resonance Images (MRIs) for the detection of AD, Healthy Control (HC), and the intermediate stage of Mild Cognitive Impairment (MCI).

The methodology involves adapting a sophisticated deep-learning model previously developed for AD classification. Post hoc explanations are incorporated using Layer-wise Relevance Propagation (LRP), enhancing interpretability. The extension of the model to include MCI aims to address the challenge of identifying patients in this critical intermediate stage.

Key findings from the analysis include the identification of diagnostic hallmarks in hippocampal and ventricular brain regions, aligning with established literature. The AD score distribution demonstrates distinct mean values for HC and AD classes, showcasing the model's discriminatory capabilities.

However, the project reveals challenges in multi-class classification, indicating the difficulty distinguishing MCI from HC and AD. The study also suggests future directions, such as incorporating additional risk factors like age to enhance classification accuracy.

In summary, this project contributes to advancing AD detection using deep learning, providing insights into model performance, interpretability, and potential enhancements for more accurate classification in complex MCI scenarios. The combination of advanced AI techniques and a nuanced understanding of model outputs offers promise for improving clinical decision-making in diagnosing neurodegenerative disorders.

Our research project focuses on the intersection of cryptocurrency markets, with a particular emphasis on Dogecoin, a notable cryptocurrency, and the impact of Elon Musk's tweets on its price trajectory. Cryptocurrencies, especially Dogecoin, have gained significant attention in financial circles, often influenced by high-profile figures like Elon Musk.

To gain a comprehensive understanding of the correlation between Dogecoin's price movements and the sentiments expressed in Elon Musk's tweets related to Dogecoin, we conducted a thorough investigation. Recognizing the limitations of relying solely on a single influencer's data, we expanded our analysis to include tweets from various users, creating a more robust dataset. Leveraging Twitter's API, we collected and meticulously processed over 45,000 tweets, employing advanced sentiment analysis models for in-depth examination.

Our analysis results reveal a discernible connection between Elon Musk's tweets and the price dynamics of Dogecoin. However, interestingly, this correlation is not reflected in the broader spectrum of Dogecoin-related tweets originating from the wider Twitter user community. Our findings shed light on the nuanced relationship between social media sentiments, individual influencers, and the cryptocurrency market, offering valuable insights into the complex dynamics at play in this rapidly evolving financial landscape.

The focus of this internship project is to explore and develop Intent-Based Dialog Systems (IBDS) and address the challenges associated with modular structures in such systems. In practical applications, Dialog Systems (DS) are often designed with a modular approach, combining multiple DS into a unified system known as Modular Dialog Systems (MDS).

The project introduces the concept of Intent-Based Dialog Systems (IBDS), which differ from traditional research-oriented dialog systems by employing rule-based Dialog Managers (DM) in industry applications. The absence of an established term for such systems in the literature leads to the adoption of "Intent-Based Dialog Systems."

The core challenge addressed in the project is the task of module selection (MS) in IBDS. This involves choosing the appropriate sub-DS for each user utterance to generate a unified response. This problem is particularly relevant in industry applications where IBDS, relying on rule-based DM, combines multiple dialog systems with different intents and architectures.

The project contributes significantly to the scientific literature by formally introducing the novel task of MS and providing an evaluation framework for this task. A benchmark dataset is proposed to evaluate models for MS, and three baseline models are presented. The experiments conducted demonstrate the importance of the user utterance text as a more robust feature for MS models compared to confidence values of Intent Detection (ID).

In summary, the internship project makes valuable contributions to advancing the understanding and development of Intent-Based Dialog Systems. It addresses the unique challenges posed by modular structures in practical applications and provides a clear evaluation framework for the task of module selection.

In the wake of the global COVID-19 pandemic, our innovative project takes center stage in the realm of health and safety. Focused on the pivotal role of face masks in protection, we have harnessed the power of advanced deep learning, unveiling a sophisticated face detection model. This model stands as a guardian, capable of real-time scrutiny in both dynamic video streams and static images.

At its core lies the Single Shot Detector (SSD) architecture, a beacon of accuracy and high-speed prowess. Elevating our approach, we seamlessly incorporate transfer learning in neural networks, allowing our model to deliver a decisive verdict on the presence or absence of a face mask, whether in a still image or a live video feed. The culmination of our endeavors reveals exceptional performance metrics, boasting an impressive precision and recall rate of 100% and 99%, respectively. Beyond the realms of technology, our project emerges as a beacon of hope, showcasing the intersection of innovation and societal well-being. As we navigate the ever-evolving landscape of public health, this endeavor serves as a testament to the transformative potential of cutting-edge technologies in the face of real-world challenges.

Shuffling in MapReduce is the process of transferring data from the mappers to reducers1. This phase is necessary for the reducers, otherwise, they would not have any input (or input from every mapper). As shuffling can start even before the map phase has finished, this saves some time and completes the tasks in lesser time1.

Sorting in MapReduce covers the merging and sorting of map outputs. Data from the mapper are grouped by the key, split among reducers, and sorted by the key1. Every reducer obtains all values associated with the same key1. The keys generated by the mapper are automatically sorted by the MapReduce Framework, i.e., before starting of reducer, all intermediate key-value pairs in MapReduce that are generated by mapper get sorted by key and not by value1. Sorting in Hadoop helps reducer to easily distinguish when a new reduce task should start. This saves time for the reducer1.

These two phases occur simultaneously and are done by the MapReduce framework1. This process allows for efficient and distributed processing of large datasets across a cluster of computers, making it a fundamental component of the Hadoop ecosystem.

This project aims to implement stream window aggregations using Kafka Streams. It focuses on processing continuous, unbounded streams of data in real-time. The key feature of this project is the use of windowing techniques to segment the stream into discrete chunks, known as windows. These windows can be time-based (tumbling, hopping, or sliding windows) or session-based, depending on the use case.

The project leverages the power of Kafka Streams, a client library for building applications and microservices, where the input and output data are stored in Kafka clusters. It offers sophisticated stream processing capabilities such as filtering, mapping, and aggregating incoming data, as well as joining multiple streams.

By using Gradle as a build tool, the project ensures a reliable and efficient build and testing process, making it easier to manage dependencies and allowing for automated testing and deployment.

Overall, “Stream Window Aggregations using Kafka Streams” is a robust solution for real-time data processing and analytics, providing valuable insights and aiding in decision-making processes.

The Twitter Sentiment Classification project is a suave endeavor designed to harness the power of cutting-edge machine learning techniques for discerning the nuanced sentiment conveyed in Twitter text data. This ambitious initiative seeks to construct a resilient model, finely tuned to decipher the emotional tone encapsulated in tweets, with a keen focus on distinguishing between positive and negative sentiments. The ultimate aim is to craft a highly accurate classifier, poised to revolutionize applications in sentiment analysis, elevate the efficiency of customer feedback processing, and contribute to insightful social media monitoring. By leveraging advanced algorithms, this project aspires to be at the forefront of sentiment classification, unraveling the intricacies of language expressed in the dynamic realm of Twitter.

You can pull it and use it from docker hub. The README.md can be found here.

Initial State.

Test with a NEGATIVE Tweet.

Test with a POSITIVE Tweet.

Previous Next

Initial state: File upload.

Upload a file (.mp3/wav) and select languages.

Results of file upload on their respective sections!

Initial state: Website streaming.

Results of streaming data!

Previous Next

This project is a sophisticated integration of three advanced technologies: Whisper ASR, Facebook’s m2m100 model, and Nvidia’s NeMo. Each of these components plays a crucial role in the pipeline.

OpenAI's Whisper ASR: This is the first step in the pipeline. It takes speech from various sources such as microphones, MP3 files, or live audio streams, and transcribes them into text. This is particularly useful in scenarios like transcribing lectures, meetings, or interviews in real-time.

Facebook’s m2m100 model: The transcribed text is then fed into this machine translation model, which translates the text into a specific language. This is a boon for global communication, breaking down language barriers in international conferences, customer support, or even in social media interactions.

Nvidia’s NeMo: The final step in the pipeline is text-to-speech synthesis. The translated text is converted back into speech, producing audio streams or audio files. This can be used to generate multilingual podcasts, voiceovers for videos, or even for accessibility purposes, aiding those with reading difficulties or visual impairments.

The integration of these three services in a sequential pipeline makes my project a comprehensive solution for multilingual speech processing. It has the potential to revolutionize communication in our increasingly global and digital world, making information more accessible and interactions more inclusive. Whether it’s for business meetings, educational content, customer service, or personal use, your project opens up a world of possibilities. It truly embodies the spirit of innovation and the power of technology to bring people together.

Initial state: Upload an Image File.

Correctly Predicted Bell Pepper Healthy Leaf!

Correctly Predicted the Leaf Disease: Apple Cedar Rust!

Correctly Predicted the Leaf Disease: Tomato Yellow Curl Virus!

Previous Next

PhytoPhinder is an innovative plant disease detection system that harnesses the power of Convolutional Neural Networks (CNNs). It diagnoses diseases in plants based on leaf images with remarkable accuracy. The system is built using PyTorch, a leading deep learning framework, and is deployed as a FastAPI application in a Docker container, ensuring scalability and ease of use. The project utilizes a comprehensive dataset of leaf images from Kaggle, known as the Leaf Disease Detection Dataset.

This dataset provides a diverse range of leaf images, covering various species and disease states, which aids in training a robust and versatile model. The dataset comprise with approximate ~90k of images (train ~70k, test ~20k). The dataset only provide 38 types of disease and healthy images of several plants. EDA has been done in notebooks/phytophinder-the-leaf-detective.ipynb notebook. You can see all the types of plant sample (healty and non-healthy) images. Which means this problem is a multi-class classification problem and we have 38 classes to classify. I also created small_dataset (you can download it from here). I used this dataset in notebooks/notebook_1.ipynb to run the project locally. You can use this dataset to run the project locally. Don't worry it has all equally distributed images from all classes.

I hyper-parameter tune it with differnt CNN models (notebook_1.ipynb, and phytophinder-the-leaf-detective.ipynb) with different inner-layers. Validation-loss defines the perforfence of the best model. Where phytophinder-the-leaf-detective.ipynb is the final version of my code. Later I convert it an anssemble everything in train.py. So you can follow that notebook and that script to evaluate my project. Using the best trained model I make predictions on whether it can detect the classes(disease or healthy) of a plant leaf. Finally, I containerise this application and publish it to docker-hub.