1 Comment

Image and Text Recognition with TensorFlow Using Convolutional Neural Networks

by Sophia TurolJune 15, 2016
Learn how to enable image recognition using a simple MNIST data set and text analysis with the word2vec embeddings on top of TensorFlow.

Convolutional neural networks (CNNs) solve a variety of tasks related to image/speech recognition, text analysis, etc. These topics were discussed at a recent Dallas TensorFlow meetup with the sessions demonstrating how CNNs can foster deep learning with TensorFlow in the context of image recognition. The examples featured MNIST, a large data set of handwritten digits, and word2vec, a group of models used to generate word embeddings.

Watch the videos below for more detail.

 

Image and text recognition (MNIST and word2vec)

Viswanath Puttagunta of Linaro provided an overview of neural network basics (weights, biases, gating functions, etc.). Then, he spoke about image recognition with a simple MNIST data set for TensorFlow and how it can be implemented with a convolutional neural network. Viswanath also explored text analysis using word2vec as an example. He demonstrated how not only symbols, but whole words and word sequences can be recognized when described as vectors.

 

 

Find Viswanath’s full slides below.

 

Fireside chat

Abhijeet Sangwan of Speetra, Inc. and Viswanath Puttagunta of Linaro answered some questions from the audience, regarding the importance of open source, the accuracy of ML tools, how much time it takes to train a model, etc.

 

 

Join our group to stay informed about the upcoming events!

 

Further reading

 

About the experts

Viswanath Puttagunta is currently breaking down various statistics and neural network frameworks (Spark Core, MLlib, Caffe, TensorFlow, etc.) to fundamental operations that can be optimized for ARM SoCs. His background is in statistics and signal processing, where he has designed and implemented algorithms related to discriminant analysis, convolution neural networks, etc.

Abhijeet Sangwan specializes in speech and language processing technology, where he has researched and developed several algorithms, systems and products. He currently serves as CTO at Speetra, where he has developed several product lines that leverage automatic speech recognition, keyword spotting, natural language processing, and speech analytics technology.