keywords: invoices annotation engine models transfer learning curriculum learning

cnn vs lstm - temporal shift module ngrams - multi conv

keywords: Given a pre-defined list of speakers classify in real time the speaker on stage test data includes adversarially attacked data

Implemented Recurrent Instance Segmentation by Bernardino Romera-Paredes et al. as described here and Fully Convolutional Networks for Semantic Segmentation by Jonathan Long et al. using Keras Framework with TensorFlow backend. The project provided me the opportunity to implement and gauge the following techniques closely:

• Fully Convolutional Networks as opposed to traditional Segmentation networks with CNN ending with Dense layer(s). FCN proved to be independent of the size of input providing ability to use the model in varying picture without adhering to padding. Experimenting and visualizing data manifold with gradient was informative in understanding the role played by activation functions in training and inferenc of a deep model. We experimented on FCN-8 /16 /32 with Relu, Leaky-Relu, Sigmoid, Tanh and dealt with vanishing gradient problem in case of Sigmoid and tanh by experimenting with L1 loss, L2 loss, Batch Normalization, Dropout.
• Attention using Recurrent Neural Network (RNN) - next in line for our experiments.
• Recurrent Neural Network and Attention-based models
• ConvLSTM
• Keras Functional API vs Sequential API (partial towards the former :P)
• Writing your own trainable and non-trainable layers in Keras

keywords: ELL-784 ELL-881 COL-106 - data structure MTL-106 - probability XYZ-ABC - image processing

keywords: style transfer, cells, data augmentation

keywords: target list face classification, with fewer false positives

keywords: virtual assistant few shot learning skewed data

celery, redis, sql, sklearn, pytorch, transformers, transfer-learning, attention, flask, python, TorchScript, C++ inferences, rasa, mutli-lingual - french, websocket (socket.io), siamese,

fast-api, sanic, github workflows