Kunal Grover

Implemented an RNN using TensorFlow. The dataset was comprised of many cocktail recipes. We wanted to generate a recipe based on user input.

Try it out: http://tensorflow-frontend.s3-website-us-east-1.amazonaws.com/

How does it work?
We used RNN because we were trying to train what comes next given part of a word. We know that the context of the next character depends on previous characters. Trained data for an RNN uses X number of characters as training data with a prediction value as the next character. Now, we know that RNNs have gradient problems, and we also require a longer context (e.g., what was the word ten words before this), so we used long short-term memory (LSTM) to solve these problems.

Steps:
Character tokenization into integers.
Implement a simple RNN with TensorFlow and LSTM.
For training, feed a large amount, e.g., 60 characters to predict the next character.
Next, add characters sequentially to retrieve the complete sentence.
Result: A simple Flask-based app on the back end working with a ReactJ front end as the user interface.

Problem:
The received set of text cards displayed to the user are required to create clusters or topics for the content so that the user receives suggestions based on similar content.

Solution:
Data cleaning with stop word removal.
Keyword extraction using term frequency-inverse document frequency (tf-idf). We selected top X words that would be used to create clusters with Word2Vec pre-trained models from Wikipedia data (fastText).
A select subset of keywords helped to get our keywords and cluster vectors. Cluster sizes could be determined by using the Elbow method. We could also do hierarchical clustering to help visualize the clustering, and thereby select good clusters.
Each article can be assigned a cluster based on the mode of clusters of keywords in that article.

Alternative approach:
Doc2Vec, didn’t try it out, but fairly similar, it combines document Word2Vecs into a single vector for comparison.

A classifier developed to detect and analyze the adultering of milk sample tests using Computer Vision. The classifier involved detecting areas where the samples reacted with the chemical and found the result of adulteration based on the test strip analysis.