TLDR: Neural Networks are computational models that stacks units called neurons into layers to create predictions after we have trained the network. Each unit performs a linear transformation on the data followed by a non-linear transformation to produce a result. Forward propagation computes the network result and backward propagation computes the changes in weights and biases for the network in subsequent epochs. The non-linear activation function supports the network to assume any mathematical function, this is called the Universal Approximation theorem.

Introduction #

Every day we are relying on machine learning models to make our lives simpler and hassle-free. Take your average day, where you start, by reading news and weather predictions, or later in the day when you are checking your emails, we unconsciously are relying on machine learning models that run behind the scenes. The list goes on and on with applications in image recognition used in our gallery, audio recognition in voice assistants. This list keeps growing every day.

A few popular classical Machine learning models include Logistic Regression, Linear Regression, Decision Trees, Support Vector Machines, and Naive Bayes models. Although these models are nowadays less commonly used in practice, they are still very useful as understanding them helps to understand the more recent models. Ensemble models such as Random forests and XGBoost are some machine learning models that are commonly used these days.

I first started doing topic modeling when I used to play around with the nips dataset. The first time I tried it, I used scikit-learn for this. I used LDA and NMF for this, and I received results that I was happy with. In this way, I think scikit-learn is one of the most appropriate tools available for exploratory data science tasks. But I had bigger plans, of tackling even bigger datasets. Then I got introduced to another python library gensim which is focused on topic modeling. Among many features it provides, it includes transformations such as online LDA, LSA and HDP, and wrappers to other popular libraries like scikit-learn, vowpal wabbit, and Mallet.

The code can be viewed at my Github repository.

Shelter animal outcomes is a knowledge competition hosted by kaggle. My goal was to learn and get familiarized with the different techniques, methods and tools required to solve classification problems. In my adventures I used pandas, matplotlib, seaborn, jupyter and scikit-learn. The goal of this competition is to predict the outcome of a cat or a dog, given it’s age, gender(also whether spayed/neutered or not), breed, color. The outcomes are return-to-owner, adoption, transfer, euthanasia and died. In some cases we are provided with outcome subtypes too.

All my work related to this can be found in this github repository, it houses my jupyter notebooks.