Projects and Challenges

Project guide: Saya Date, Cofounder and CTO, Linecraft AI, Pune 

Elevated operational efficiency by optimizing an existing feature through the implementation of the home state marking algorithm, streamlining the computation of home states for individual assets 

Engineered a machine learning model that forecasts the probability of asset failure within a designated time window for all assets of a specific type, enhancing predictive maintenance strategies

Unearthed valuable insights from data analysis, elucidating the 93% recall rate in failure prediction and enhancing model interpretability, driving actionable intelligence

This contains all the assignments and reports written in IEEE conference formats, as a part of the coursework in EE4708, offered by Prof. Gaurav Raina, Department of Electrical Engineering, IITM

This repository consists of all the assignments, projects, and lab sessions which I worked in the course AI-2 Computer Vision, offered by Prof. Pavlos Protopapas, Univ.AI. In the final project we worked on the problem, "Computer Vision for hurricane disaster management". Some of the notable lab sessions include autoencoders, transfer learning, effect of several optimizers on training etc.

We have used pretrained networks, namely DenseNet, Incetion, Inception-ResNet, VGG, and ResNet for the detection of covid-19, pneumonia from the analysis of chest X-ray images of patients. For each network, we have added two dense layers and trained them on the dataset. The weights of these layers for each network are present in the '.rar' files. Visualisations.ipynb shows the Saliency map and GradCAM for a particular X-ray image of a patient diagnosed with covid-19. Deep patterns.ipynb shows the evolution of complex features learned by the convolutional network which may be key to understand the functioning of the black-box model. Ensemble.ipynb gives us ways to improve the generalisability of the pretrained models by proposing a soft-voting classifier and a stacking ensemble. Feature disentanglement.ipynb gives us a proof that the model extracts excellent features that are generalisable and robust. This implies that the model can perform well even on unseen data, though it is trained on very few data samples!

In the notebooks, a) Class 0 represents 'Covid' b) Class 1 represents 'Normal' c) Class 2 represents 'Pneumonia'

Distinguished ourselves by competing against 700+ teams nationwide in the Electrothon ML challenge, organized by NIT Hamirpur, clinching victory in ”The Best COVID Solutions” category

Link for demo video 

External references:

1. https://github.com/keisen/tf-keras-vis/blob/master/examples/attentions.ipynb

2. https://www.tensorflow.org/tutorials/generative/deepdream

3. https://arxiv.org/pdf/2004.13175.pdf

Tackled a challenging problem statement necessitating the development of a domain-specific QA system, prioritizing efficiency while upholding performance standards comparable to state-of-the-art (SOTA) methods, culminating in the conceptualization of Retrieve Twice, Rank and Answer (R2RA), an end-to-end pipeline tailored for domain-specific question answering

Achieved exceptional results with our system boasting a top-1 retrieval accuracy of 87.1% and a reader F1 score of 81.7%, positioning it as not only accurate but also one of the most efficient solutions in the competition, boasting an overall latency of 487ms

Secured the top position among teams representing 23 IITs across India

The link to the paper is provided

A computer vision and deep learning framework that helps identify individual Spotted Deer(Chital) to help aid wildlife conservation

Showcased the findings of the project and discussed several applications in Engineer’s Conclave, Inter IIT Tech Meet, 2021, IIT Guwahati