A couple weeks ago, I traveled to Sikkim with my family. Sikkim is a state in northeast India, known for its glaciers, mountains, tea gardens and diverse flora! It is also home to India’s newest airport (Pakyong Airport) built at 4,500 ft above sea level, whose news sparked our interest to travel there.

The last few months have been a complete whirlwind — wrapping up my undergraduate thesis, final examinations, graduating from college, moving to a new home & starting my first real job at Google! More on all of this in future blog posts but in between all of the adult-ing, I also traveled to eastern Europe with a few friends from university for our grad trip!

Earlier today, I conducted a workshop on the basics of neural networks and implementing them in Keras. You can find my slides below and the supporting implementation code here.

In the past few years, two major patterns in our consumption of online content have come to light — the proliferation of misinformation (or “fake news” as Trump would say) and the entrapment of users in personal filter bubbles. This was especially true during the 2016 US Election and is becoming increasingly prominent in my home country of India, where I constantly see family members being swayed by an article or a claim that panders to their inherent bias. The alarming part is that false (or somewhat false) “news” is no longer limited to dodgy blog sites and virus-ridden webpages but also posts on social media by influential people (read: politicians) with valid credentials and a large following.

As part of the initial studying for my final year project (more on that in a future blog post), I had to learn how to build deep models for text classification. As a newbie to deep learning at the time, I decided to start with a small project to familiarize myself with deep learning concepts, TensorFlow and the field of text classification. After some digging, I decided to build classifiers for the movie review sentence polarity dataset (Pang and Lee, 2004) since this was a relatively smaller dataset, dealt with sentences (as opposed to documents) and was a binary classification task. Additionally, there were multiple existing references that I could use as a starting point, which proved to be extremely useful. The complete source code and detailed results for my project can be found on GitHub.