Atul Dhingra

Built a character level language model to generate new dinosaur names with a character level RNN

Debiasing Word Vectors by neutralizing and equalizing bias for non-gender specific words using GloVe word vectors

Emojify: Making text sentences more expressive with emojis, using word embeddings and LSTM

Developed a LSTM based integer to base-k encoder-decoder with tensorflow

Neural Machine Translation with an attention model to convert human readable dates into machine readable dates using LSTM

Generated novel jazz solos using LSTM networks in a style representative of a body of performed work in training set

End to end trigger-word detection model using 1D CNN and LSTM on speech spectogram obtained from synthesized speech data

Gradient checking and Regularization for a Deep Neural Network

Optimization for a Deep Neural Network with Gradient Descent, Momentum and Adam

Developed a CNN based temperature level detection and prediction model from thermometer images

Implemented YOLO model for car detection system for autonomous vehicles using images from dashboard camera

Eclectic-Nets: A suite of most popular CNNs with keras and tensorflow that include leNet, ResNet-50 etc.

A Deep Learning based face identification, verification and happiness classification using FaceNet architecture with triplet loss

Implemented ResNet-50 model from scratch for image classification task on hand sign(SIGNS dataset)

Implemented Neural style transfer algorithm for generating novel artistic images of the Louvre museum based on painting of Monet

Clustered scenes into groups based on SIFT features with FLANN based matcher using OpenCV

We built a dataset of nearly all the Indian electors. Our data includes information on first and last name, gender, polling station (constituency, district, and state), father or husband's name, among other such details. We assembled this data by scraping and parsing the electoral rolls

A deep learning-based face detection, landmark detection, and gender classification pipeline in a unified workflow using dlib and keras(tensorflow backend)

Stock price prediction using LSTM and 1D Convoltional Layer implemented in keras with TF backend on daily closing price of S&P 500 data from Jan 2000 to Aug 2016

Convolutional Neural Networks(CNN) has had a great success in the recent fast, because of the advent of faster GPUs and memory access. CNNs are really powerful as they learn the features from data in layers such that they exhibit the structure of the V-1 features of the human brain. A huge bottleneck in this case is that CNNs are very large and have a very high memory footprint, and hence they cannot be employed on devices with limited storage such as mobile phone, IoT etc. In this work we… Show more

Convolutional Neural Networks(CNN) has had a great success in the recent fast, because of the advent of faster GPUs and memory access. CNNs are really powerful as they learn the features from data in layers such that they exhibit the structure of the V-1 features of the human brain. A huge bottleneck in this case is that CNNs are very large and have a very high memory footprint, and hence they cannot be employed on devices with limited storage such as mobile phone, IoT etc. In this work we study the model complexity versus accuracy trade-off on MNIST dataset, and give a concrete framework of handling such a problem, given the worst case accuracy that a system can tolerate. In our work we reduce the model complexity by 620 times, and memory footprint by 19.5 times compared to base model, while achieving worst case accuracy threshold. Show less

Detect the wide Receiver in a single shot video clip, where the offensive team is to the left

This classifier uses the sub-strings of size 3 to learn probability associated with each substring such that it belongs to a particular ethnic group using the census 2000 data. When a new string is presented, it is broken down into sub-strings of size 3, and prediction is done on each of the sub-string using the model computed before. The output accuracy of the predictor is 85% even in an unbalanced dataset. The model follows the Naive Bayes assumption of conditional independence

Synthesized clip-art and a 3D mesh onto an image for which the homographies have been calculated using the camera caliberation project.

Calibrate the camera of a robot vehicle for 3d and 2d camera calibration.

Lasso is implemented on CIFAR-10 dataset via Co-ordinate Gradient descent algorithm

This work involves Texture Synthesis and Image Inpainting. The idea of Texture Synthesis by Non-Parametric Sampling proposed by Efros and Leung, ICCV99, and region filling by Criminisi et al “Region Filling and Object Removal by Exemplar-Based Image Inpainting”, TIP 2004 has been implemented.

Implemented and evaluated four basic face recognition algorithms: Eigenfaces, Fisherfaces, Support Vector Machine (SVM), and Sparse Representation-based Classification (SRC) on YaleB dataset

Designed an optical character recognition (OCR) system for hand-written characters, in a fixed-resolution setting.

Spam Classification on the Web Spam Dataset using Percepton and Kernel Perceptron with Polynomial, Gaussian, Exponential and Laplacian Kernels.

A Naive Bayes algorithm is implemented on the Reuters dataset for topic classification("earn" or "acq")

In this thesis, we study two problems based on clustering algorithms. In the first problem, we study the role of visual attributes using an agglomerative clustering algorithm to whittle down the search area where the number of classes is high to improve the performance of clustering. We observe that as we add more attributes, the clustering performance increases overall. In the second problem, we study the role of clustering in aggregating templates in a 1:N open set protocol using multi-shot… Show more

In this thesis, we study two problems based on clustering algorithms. In the first problem, we study the role of visual attributes using an agglomerative clustering algorithm to whittle down the search area where the number of classes is high to improve the performance of clustering. We observe that as we add more attributes, the clustering performance increases overall. In the second problem, we study the role of clustering in aggregating templates in a 1:N open set protocol using multi-shot video as a probe. We observe that by increasing the number of clusters, the performance increases with respect to the baseline and reaches a peak, after which increasing the number of clusters causes the performance to degrade. Experiments are conducted using recently introduced unconstrained IARPA Janus IJB-A, CS2, and CS3 face recognition datasets. Show less

In this work, we consider learning sparse models in large scale setting, where the number of samples and the feature dimension can grow as large as millions or billions. Two immediate issues occur under such challenging scenarios: (i) com- putational cost; (ii) memory overhead. In particular, the memory issue precludes a large volume of prior algorithms that are based on batch optimization technique. To remedy the problem, we propose to learn sparse models such as Lasso in online manner where… Show more

In this work, we consider learning sparse models in large scale setting, where the number of samples and the feature dimension can grow as large as millions or billions. Two immediate issues occur under such challenging scenarios: (i) com- putational cost; (ii) memory overhead. In particular, the memory issue precludes a large volume of prior algorithms that are based on batch optimization technique. To remedy the problem, we propose to learn sparse models such as Lasso in online manner where in each iteration, only one randomly chosen sample is revealed to update a sparse iterate. Thereby, the memory cost is independent of the sample size and gradient evaluation for one sample is efficient. Perhaps amazing, we find that with the same parameter, sparsity promoted by batch methods is not preserved in online fashion. We analyze such interesting phenomenon and illustrate some effective variants including mini-batch methods and a hard thresholding based stochastic gradient algorithm. Extensive experiments are carried out on a public dataset which supports our findings and algorithms.
Show less

Implemented various Route-Finding algorithms as a part of the coursework.

• Created and presented interactive web based algorithm snippets for Minimum-Spanning Tree algorithm
• The project was awarded 3 rd best in approximately 25 projects.

• Investigated the role of repetitive activities(robust plasticity of motor cortex) for neurorehabilitation with
the help of a non-invasive brain computer interface.
• Built a VR rehabilitation game with Unity and C#, and engineered an algorithm to dynamically scale
with and treat upper-arm disabilities with an immersive game design.

Developed a robust speaker verification algorithm invariant to noise and multi-channel input using GFCC,
MFCC and i-vectors.
• The research was published in Springer, Proceedings of the IEEE International Conference on
Signal, Networks, Computing, and Systems 2016

• Explored various applications of Biometrics, including heuristics and authetication based applications as
a part of my undergraduate thesis.
• Studied multi-modal aspect of biometrics and correlated various biometric modalities including hand-veins

• Formulated and published gesture recognition modules for controlling Audio-Books for visually impaired
with OpenCV
• Compared color-based and color-independent tracking of hand based gesture

• Developed and published a novel method for error analysis for eye movement tracking for biometrics.
• Published the research in Springer,Proceedings of Swarm, Evolutionary and Memetic Computing
2013 Part II, LNCS(Springer) 8298, pp. 248-256, 2013

The project was aimed in implementing a Light Detection Circuit using a Light Dependant Resistor, and a Bridge Wave Rectifier, on a Bread Board. so that it could be extrapolated by the organization in fire and explosion applications at the lab, Centre for Fire, Explosion and Environment Safety.

Temperature Measuring and Recording System using PIC 16F877A microcontroller by making use of DS18B20 Temperature Sensor. The system was developed to display and then store temperature readings during the experiment. The project was developed so as the system can be implemented by the organization further for temperature measurements in Explosions And Blasts , using the needed sensors.

Iris Recognition is a rapidly expanding method of biometric authentication that uses pattern recognition techniques on images of irides to uniquely identify an individual. The aim of the project was compiling a robust database of human irides and to compare the results of the acquired images with the previous popular available database such as Chinese Academy of Sciences' Institute of Automation (CASIA). The project became a stepping stone towards my pursuit for field of Computer Vision and its… Show more

Iris Recognition is a rapidly expanding method of biometric authentication that uses pattern recognition techniques on images of irides to uniquely identify an individual. The aim of the project was compiling a robust database of human irides and to compare the results of the acquired images with the previous popular available database such as Chinese Academy of Sciences' Institute of Automation (CASIA). The project became a stepping stone towards my pursuit for field of Computer Vision and its practical applications. This is the first project i have been involved which is the muse for me for going forward in the filed of Computer Vision. Show less