Contemporary Data Science Application Development

Opening Session

The entire machine learning model development process has significantly changed in last one or two years. A data scientist has many options available for ML development. This has posed many challenges to him. In this lecture, I will present the challenges faced by a data scientist in the modern age.

Due to recent technology changes, the whole approach in model building has changed. In this lecture, you will learn more about MLaaS - Machine Learning as a Service. You will understand the benefits of using it and also its limitations.

You learned the MLaaS approach in my last lecture. When you cannot overcome the limitations of this approach, the next one is considering the use of AutoML. In this lecture, you will understand AutoML and its pros and cons.

When both MLaaS and AutoML options have failed for you, you have left with no choice other than using a traditional way of model development. Still, the major decision comes to the use of appropriate technology - whether to use GOFAI or ANN? So, what is the deciding factor? This lecture tells you which technology to use that is the most appropriate for your current project.   

You are here because your dataset is small and you have decided to use GOFAI approach for model building. This is a well-researched area and thus provides a well-defined workflow for model development. In this lecture, I will describe this workflow to you. If you have tried your hands earlier on Machine Learning, this workflow may not be something new to you. What comes important here is your approach in each step. Especially, the dimensionality reduction and selecting appropriate machine learning algorithm are the two most important steps. I will provide you the definitive guidelines in completing these steps, focusing from the angle of a data scientist. I will tell you precisely what you need to learn to master the GOFAI approach.

Creating ML models using ANN technology is lot more easier considered to classical approach, that you saw in the previous lecture. In this lecture, I will show you what all you need to do, or rather what all you need to learn to take the ANN approach. You will understand the various activation, optimisation functions and layer types. I will describe all the steps required in this approach so that you can create your own ANN-based ML models and deploy them on your production servers.

A trivial ANN is no good for advanced applications in computer vision and NLP domains. We need DNNs. Though it is fairly simple to train a DNN, the resource requirements in terms of hardware and the training time are very high and beyond the reach of a common person. Many behemoths in this industry have trained such large networks and made them publicly available. The question is, which one is the most suited for you. So, in this chapter, after introducing you to the various pre-trained models, I will also guide you on their comparative study. Towards the end, I will also tell you why customised DNNs are still required.

In this lecture, I will summarise all your learnings so far. I will provide you a cheat sheet that gives you a consolidated workflow summary of all the development paths that you have studied so far. This will make your life easier while handling a new data science project.