Project

	Prostate cANcer graDe Assessment (PANDA) Kaggle \| Supervised Research Exposition Developed a Machine Learning pipeline to deal with gigapixels multi-resolution wholeslide images. Trained two staged Deep Learning model to to classify the severity of prostate cancer from microscopy scans of prostate biopsy samples. Currently acheived Quadritic Kappa Score of 0.907 and ranked 22 out of 930+ international participanting teams on Kaggle Pulic Leaderboard. [ Code ]
	Recommendation Model Neural Collaborative Filtering Swift for TensorFlow \| Open Source Project Implemented Neural Collaborative Filtering architecture in new deep learning framework Swift for TensorFlow to model latent features of users and items and got it merged in swift-model package. Added new dataset MovieLens in Swift for TensorFlow dataset package to directly load and use it. Experimented on MovieLens-100K dataset for the correctness test of Recommendation model by predicting the top K-items user will interact in coming days. [ Report ] [ Code ]
	Detection of Sign of Depression using Social Media Text University of Cambridge \| Internship Developed a Machine Learning pipeline for detection of depression based on messages from social media platform. Employed Deep Learning model BiLSTM with Attention to learn the mental information from sparse space with unbalanced small dataset.
	Instance Segmentation Advance Machine Learning \| Course Project Implemented deep neural network Mask-RCNN to detect and provide segmentation masks to object belonging to 300 different categories. Extended the model Mask-RCNN to Open Image Dataset provided by Google AI for Instance Segmentation Open Image Challenge 2019. [ Blog Post ] [ Code ]
	Looking to Listen Automatic Speech Recognition \| Course Project Implemented Speech Seperation paper by Google Research to isolate a single speech signal from a mixture of sounds. Build an End to End pipeline consisting of Audio Model, Visual Model both consitiing of Dilated CNN and Fusion Model consisting of BiLSTM followed by fully connected layer. [ Report ] [ Code ]
	Competition and Collaboration Foundations of Intelligent and Learning Agents \| Course Project Trained two agents to play in Tennis environment where the agent must bounce the ball between one another while not dropping the ball out of bounds. Used Actor-Critic network for training where Actor determine best action and Critic evaluate quality of action as determined by Actor inorder to incorporate action taken by both agent for better learning. [ Report ] [ Code ]
	Image Classification on CIFAR Advance Machine Learning \| Course Project Implemented forward and backward pass of different layers of Fully Connected Neural Networks from scratch with the flexibility to take variable size input. Trained the implemented Neural Network for object classification on CIFAR dataset and achieved 75-80% accuracy on each class of CIFAR. [ Code ]
	Automation of Gate Security System Collected around 1700+ images of vehicles and manually annotated all the desired class present in it. Trained model YOLOv3 on annotated images to detect vehicle along with its type and locate the License Plate of the detected vehicle for recognition purpose. Delivered a Detection Module of Automation model for automating the gate security system to ease the guard’s job. [ Report ] [ Code ]
	License Plate Detection and Recognition Computer Vision \| Course Project Implemented an EECV 2018 paper in PyTorch using CNN to extract features and fully connected layers in the end to predict bounding box around License Plate. Built a Recognition Module which exploits Region of Interest from CNN layers to extract features map of interest and several classifiers to predict the corresponding license plate number. [ Report ] [ Code ]
	Non-Invasive Glucometer Electronic Design Lab \| Course Project Designed an analog circuit to get the amplified voltage level for corresponding glucose concentration present in the body. Collected blood sugar concentration data using the designed setup and invasive glucometer and trained Regression model on it. Delivered an alternative low-cost solution to traditional invasive glucose testing method for monitoring glucose-related diseases. [ Report ]
	Toonification of Image Image Processing \| Course Project Implemented Bilateral Filtering for smoothing and quantizing the colors and Edge Detection for detecting and boldening the edges in the image. Combined these implementation to get an artistic and comical effect on a wide range of images. Enhanced speed and accuracy of the algorithm using Fast Bilateral Filtering by working in higher dimensional space. [ Code ]
	Image Classifier Deep Learning with PyTorch(MOOC) \| Udacity Built an Image classifier using Neural Networks in PyTorch to recognize 102 different species of flowers present in the dataset. Achieved a test accuracy of 95% by setting appropriate optimizer, loss function and learning rate. [ Code ]
	Transferring & Receiving Of Image Using Gnu Radio Communication Lab \| Course Project Performed all processing of data like modulation, demodulation, encoding, decoding within GRC. Resolved all effects of a channel which distort the data during transmission through the atmosphere. Resolved all effects of a channel which distort the data during transmission through the atmosphere. [ Code ]
	Reaction Game Digital Circuit Lab \| Course Project Programmed the Altera’s MAX V CPLD board to calculate the reaction time of the player in milliseconds and display it on a 16x2 LCD display over a period of eight iterations. Wrote VHDL programs in Quartus to make an led blink at a random time. The player has to press a button as soon as possible while the CPLD computes the reaction time of the player.
	Face Recognition Image Processing \| Course Project Implemented a Face Recognition system using the PCA algorithm. Divided the dataset of images into training and test set and than applied PCA algorithm on it to build a face recognition system. Verified the accuracy of Recognition model by computing recognition rate on test data set and also by computing recognition rate on image that was not present in dataset [ Code ]
	Stock Market Clustering Self Project \| Guide: Eduonix Tech Imported some past stock data from the Yahoo Finance and then train K-means clustering algorithmto detect similar companies based on stock market movement. Used Jupyter notebook to develop a python application that can find similarity among companies thatmight not ordinarily be discovered [ Code ]
	Satellite Image Receiving STAB IIT Bombay \| Hobby Project Built our own double cross antenna to receive satellite signals. Decoded the image of NOAA satellite by using RTL-SDR dongle and double crossed antenna
	1D Motion Stage Analog Curcuit Lab \| Course Project Designed the circuit to control the bi-directional motion of the DC motor,controlled by the rotation of potentiometer knob Simulated the circuit in NgSPICE to test various values of resistors and capacitors Used opamps as integrator, differentiator and schmitt trigger to create PWM pulse and later feed thatpulse to L293D to run the motor in both direction [ Report ]
	D.C. motor speed control Analog Curcuit Lab \| Course Project Designed a circuit to effectively control the of DC Motor using dip switches without using microcon-troller Developed the circuit using Preset Counter, J-K Flip Flops, Logic Gates and took input through Dip Switch Implemented Pulse Width Modulation (PMW) by using digital integrated circuits only

	Prostate cANcer graDe Assessment (PANDA) Kaggle \| Supervised Research Exposition Developed a Machine Learning pipeline to deal with gigapixels multi-resolution wholeslide images. Trained two staged Deep Learning model to to classify the severity of prostate cancer from microscopy scans of prostate biopsy samples. Currently acheived Quadritic Kappa Score of 0.907 and ranked 22 out of 930+ international participanting teams on Kaggle Pulic Leaderboard. [ Code ]
	Recommendation Model Neural Collaborative Filtering Swift for TensorFlow \| Open Source Project Implemented Neural Collaborative Filtering architecture in new deep learning framework Swift for TensorFlow to model latent features of users and items and got it merged in swift-model package. Added new dataset MovieLens in Swift for TensorFlow dataset package to directly load and use it. Experimented on MovieLens-100K dataset for the correctness test of Recommendation model by predicting the top K-items user will interact in coming days. [ Report ] [ Code ]
	Detection of Sign of Depression using Social Media Text University of Cambridge \| Internship Developed a Machine Learning pipeline for detection of depression based on messages from social media platform. Employed Deep Learning model BiLSTM with Attention to learn the mental information from sparse space with unbalanced small dataset.
	Instance Segmentation Advance Machine Learning \| Course Project Implemented deep neural network Mask-RCNN to detect and provide segmentation masks to object belonging to 300 different categories. Extended the model Mask-RCNN to Open Image Dataset provided by Google AI for Instance Segmentation Open Image Challenge 2019. [ Blog Post ] [ Code ]
	Looking to Listen Automatic Speech Recognition \| Course Project Implemented Speech Seperation paper by Google Research to isolate a single speech signal from a mixture of sounds. Build an End to End pipeline consisting of Audio Model, Visual Model both consitiing of Dilated CNN and Fusion Model consisting of BiLSTM followed by fully connected layer. [ Report ] [ Code ]
	Competition and Collaboration Foundations of Intelligent and Learning Agents \| Course Project Trained two agents to play in Tennis environment where the agent must bounce the ball between one another while not dropping the ball out of bounds. Used Actor-Critic network for training where Actor determine best action and Critic evaluate quality of action as determined by Actor inorder to incorporate action taken by both agent for better learning. [ Report ] [ Code ]
	Image Classification on CIFAR Advance Machine Learning \| Course Project Implemented forward and backward pass of different layers of Fully Connected Neural Networks from scratch with the flexibility to take variable size input. Trained the implemented Neural Network for object classification on CIFAR dataset and achieved 75-80% accuracy on each class of CIFAR. [ Code ]
	Automation of Gate Security System Collected around 1700+ images of vehicles and manually annotated all the desired class present in it. Trained model YOLOv3 on annotated images to detect vehicle along with its type and locate the License Plate of the detected vehicle for recognition purpose. Delivered a Detection Module of Automation model for automating the gate security system to ease the guard’s job. [ Report ] [ Code ]
	License Plate Detection and Recognition Computer Vision \| Course Project Implemented an EECV 2018 paper in PyTorch using CNN to extract features and fully connected layers in the end to predict bounding box around License Plate. Built a Recognition Module which exploits Region of Interest from CNN layers to extract features map of interest and several classifiers to predict the corresponding license plate number. [ Report ] [ Code ]
	Non-Invasive Glucometer Electronic Design Lab \| Course Project Designed an analog circuit to get the amplified voltage level for corresponding glucose concentration present in the body. Collected blood sugar concentration data using the designed setup and invasive glucometer and trained Regression model on it. Delivered an alternative low-cost solution to traditional invasive glucose testing method for monitoring glucose-related diseases. [ Report ]
	Toonification of Image Image Processing \| Course Project Implemented Bilateral Filtering for smoothing and quantizing the colors and Edge Detection for detecting and boldening the edges in the image. Combined these implementation to get an artistic and comical effect on a wide range of images. Enhanced speed and accuracy of the algorithm using Fast Bilateral Filtering by working in higher dimensional space. [ Code ]
	Image Classifier Deep Learning with PyTorch(MOOC) \| Udacity Built an Image classifier using Neural Networks in PyTorch to recognize 102 different species of flowers present in the dataset. Achieved a test accuracy of 95% by setting appropriate optimizer, loss function and learning rate. [ Code ]
	Transferring & Receiving Of Image Using Gnu Radio Communication Lab \| Course Project Performed all processing of data like modulation, demodulation, encoding, decoding within GRC. Resolved all effects of a channel which distort the data during transmission through the atmosphere. Resolved all effects of a channel which distort the data during transmission through the atmosphere. [ Code ]
	Reaction Game Digital Circuit Lab \| Course Project Programmed the Altera’s MAX V CPLD board to calculate the reaction time of the player in milliseconds and display it on a 16x2 LCD display over a period of eight iterations. Wrote VHDL programs in Quartus to make an led blink at a random time. The player has to press a button as soon as possible while the CPLD computes the reaction time of the player.
	Face Recognition Image Processing \| Course Project Implemented a Face Recognition system using the PCA algorithm. Divided the dataset of images into training and test set and than applied PCA algorithm on it to build a face recognition system. Verified the accuracy of Recognition model by computing recognition rate on test data set and also by computing recognition rate on image that was not present in dataset [ Code ]
	Stock Market Clustering Self Project \| Guide: Eduonix Tech Imported some past stock data from the Yahoo Finance and then train K-means clustering algorithmto detect similar companies based on stock market movement. Used Jupyter notebook to develop a python application that can find similarity among companies thatmight not ordinarily be discovered [ Code ]
	Satellite Image Receiving STAB IIT Bombay \| Hobby Project Built our own double cross antenna to receive satellite signals. Decoded the image of NOAA satellite by using RTL-SDR dongle and double crossed antenna
	1D Motion Stage Analog Curcuit Lab \| Course Project Designed the circuit to control the bi-directional motion of the DC motor,controlled by the rotation of potentiometer knob Simulated the circuit in NgSPICE to test various values of resistors and capacitors Used opamps as integrator, differentiator and schmitt trigger to create PWM pulse and later feed thatpulse to L293D to run the motor in both direction [ Report ]
	D.C. motor speed control Analog Curcuit Lab \| Course Project Designed a circuit to effectively control the of DC Motor using dip switches without using microcon-troller Developed the circuit using Preset Counter, J-K Flip Flops, Logic Gates and took input through Dip Switch Implemented Pulse Width Modulation (PMW) by using digital integrated circuits only

Sudhir Kumar Suman

Project