Data Science Associates
- Many times Health Care Personels (HCPs) are not able to detect some diseases that might any patient have because these are rare-diseases.
- Therefore, here, I worked on rare-disease detection from patient journey.
- To solve this problem, I worked on learning patient-disease representation using BERT. After extracting the representation, Semi supervised GAN and PU learning strategies are used for calssification because of the limited availability of the data.
- Worked on enhancing different SOTA NLP models (i.e., BERT, RoBERTa, and XLNet) for various biomedical downstream tasks such as NER, Document Classication etc.
- Developed ZS NLP python package for document classification. This package supports supervised, unsupervised, and weakly supervised methods. Moreover, it also supports different types of word embeddings such as GLoVe, word2vec, and language models such as BERT, Bio-BERT, RoBERTa, XLNet.
- Furthermore, I worked on speech enhancement for reconstruct the distorted speech due to internet connectivity and backgrond whisper noise.
- Worked on deep learning applications in Voice Conversion, Multimodalities (Whisper, NAM, and Normal) speech conversion for development of silent interfaces, and impaired speech analysis.
- Mainly I have worked on designing different mapping functions for various tasks using deep generative models, which are further enhanced with representation learning methodologies.
- As research assistant at Speech Research Lab, I've contributed into several research works, which lead to 8 research publications.
Research and Development Intern
- Worked on object detection algorithm for driving scenarios (especially Potholes Detection) using deep learning.
- For this task, I also develop road segmentation using transfer learning and implemented a faster-rcnn method for potholes detection. [Resules]
- In addition, helped to recruit new student interns for the computer vision profile.
- Worked on query optimization techniques for medical documents. For this task, I used different IR algorithms(such as Rochhio) to dierent standard ML algorithms(such as SVM, ANN, Random Forest).
- We got a 9.7% increase (in terms pf P@5) in result compared to the baseline system (Elastic Search, an advanced search engine).
- Additionally, developed online platform, which taken disease input and returns the genes according to it's importance.
- Designed the algorithm which can extract top positive and negative sentences by covering all the details about the given product.
- For this task, I used IR techniques (i.e., Cosine similarity, TF-IDF), where each and every sentences were rated according to their importance. And sentence similarity module was designed to filterout repetative and reducdant sentences.
- Also, developed an API in Django rest framework for general users. [GitHub]
Maitreya Patel, Mirali Purohit, Jui Shah, and Hemant A. Patil, "CinC-GAN for Effective F0 prediction for Whisper-to-Normal Speech Conversion," accepted in 28th European Signal Processing Conference (EUSIPCO), Amsterdam, The Netherlands, August 24-28, 2020.
Mirali Purohit, Mihir Parmar, Maitreya Patel, Harshit Malaviya, and Hemant A. Patil, "Weak Speech Supervision: A case study of Dysarthria Severity Classification," accepted in 28th European Signal Processing Conference (EUSIPCO), Amsterdam, The Netherlands, August 24-28, 2020.
Harshit Malaviya, Jui Shah, Maitreya Patel, Jalansh Munshi, and Hemant A. Patil, "MSpeC-Net : Multi-domain Speech Conversion Network," in 45th IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, May 04-08, 2020.
Mirali Purohit, Maitreya Patel, Harshit Malaviya, Ankur Patil, Mihir Parmar, Nirmesh Shah, Savan Doshi, and Hemant A. Patil, "Intelligibility Improvement of Dysarthric Speech using MMSE DiscoGAN," in International Conference on Signal Processing and Communications, Bangalore, India, July 20-23, 2020.
Divyesh Rajpura, Jui Shah, Maitreya Patel, Harshit Malaviya, Kirtana Phatnani, and Hemant A. Patil, "Effectiveness of Transfer Learning on Singing VOice Conversion in the Presence of Background Music," in International Conference on Signal Processing and Communications, Bangalore, India, July 20-23, 2020.
Maitreya Patel, Mihir Parmar, Savan Doshi, Nirmesh Shah, and Hemant A. Patil, "Adaptive Generative Adversarial Network for Voice Conversion," in Asia-Pacic Signal and Information Processing Association Annual Summit and Conference (APSIPA-ASC), Lanzhou, China, Nov. 18-21, 2019.
Anery Patel, Maitreya Patel, Tushar Gadhiya, and Anil Roy, "PolSAR Band-to-Band Image Translation Using Conditional Adversarial Networks," in IEEE Sensors, Montreal, Canada, October 27-30, 2019.
Maitreya Patel, Mihir Parmar, Savan Doshi , Nirmesh Shah, and Hemant A. Patil, "Novel Inception- GAN for Whispered-to-Normal Speech Conversion," in the 10th ISCA Speech Synthesis Workshop (SSW), Interspeech, Vienna, Austria, Sep. 14-16, 2049.
Mihir Parmar, Savan Doshi, Nirmesh J. Shah, Maitreya Patel, and Hemant A. Patil, "Effectiveness of cross-domain architectures for whisper-to-normal speech conversion," in 27th European Signal Processing Conference (EUSIPCO), A Coruna, Spain, September 2-6, 2019.
Few shot learning is one of the most necessary and important research area in deep learning. However, with fast pacing field few shot learning methods aren't tested and analysed for various applications such as Image, Text, and Speech. Currently, I've analysed the Relation-Network for image and text classification tasks. And results are published at my GitHub repository and blog. Moreover, I've analyzed effects of different architectures and transfer learning to design more efficient methods to push the SOTA results in simplest way possible.
Accident detection and Road quality measurement
In this project, we designed the algorithm to detect the accidents in various scenarios such as free fall, a sudden break, etc. with precision on whether the bike is falling in the left or right direction and road roughness measurement using different IoT modules. We used different hardware such as Accelerometer, Gyroscope, Pressure Sensor, GPS, GSM, and Raspberry Pi 3B+. In addition, we designed deep learning based accurate potholes detection using tensorflow.
An online learning platform with advanced techniques to help students learn better such as press a key in between any video and get the related books, notes for that timestamp of the video and a lot more. Also, get the most reasonable images related to your selected sentence for remembering anything easily. Developed all this using different information retrieval algorithms and django and react-native. 3rd rank in tic-tech-toe hackathon.
Any product ratings given on e-commerce websites are simply average of all the rating given to a particular product. However, person to person experience will vary and there can be the case where two person had same experience but they gave different ratings. To overcome this problem we developed the system using Natural Language Processing which analyses each and every review of a given product and based on this it will give the more robust rating.
Deep Visualisation - torchdv [GitHub]
Designed python library for deep visualization of any CNN classier in PyTorch. Implemented Guided Back-propagation with positive and negative salience, with Heat-map, Guided-Grad-CAM, Smooth Grad, Layer and Filter visualization using activation map, Inverted image representation, Deep dream, Class-specic image generation, etc. for researchers.
Farmer Boy [GitHub]
Researched and Developed a CNN based architecture for classification of 39 different plant diseases. This model was able to run on very simple system like Rassberry pie with only 512Mb RAM with an accuracy of 98.6%. Though state-of-the-art accuracy is 99.4% but that model is not able run on low computing system. Additionally, developed Django-rest-framework based web platform to test the trained model on reallife scenarios.