Tryst with Machine Learning - A Short Dossier

It is almost certain by now that Automation and AI will play a major role in most of the things we would do in future. There is plenty of content available to emphasize the importance why AI and Automation shall be adopted and to assert the importance of including this in our toolkits and ways of working.

started to understand this area in slightly more detail – as to ‘why’ machine learning is such a critical technology and ‘how’ does it solve the automation problems and ‘what’ kind of problems it can solve – Thereby giving insights into its implications and applications. Attempting to capture my notes and learning here and connect a few dots which might be helpful to gain the better understanding of ML - more from applied usage standpoint (not focused on programming syntax).

It is Important to understand that AI and Machine Learning are not exactly the same. Although, AI and machine Learning are used interchangeably these days, which is not entirely true. AI is a much broader topic leveraging the strengths of NLP (Natural Language Processing), ML (Machine Learning) and Deep Learning. Many of the practical implementations being seen now-a-days are fundamentally the machine learning capabilities. This will be one of the crucial building-blocks to learn about.

Anatomy of Machine Learning

ML gives the ability for computers to learn on its own without being explicitly programmed. A huge data-base of historical data and set of Algorithms are the central building blocks of ML. Algorithms regress a hypothesis based on data-base and predict the result. The word hypothesis is important here since ML are not programmed in traditional sense of computing.

Tom Mitchell has given a precise definition “A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.

E = Historical (or Training) Data-base

T = Intended output of a machine learning hypothesis.

P = Algorithms for outcome with inbuilt performance precision.

What kind of Problems Machine Learning Solves

Pretty much all which are repetitive in nature and carries a huge amount of data to analyze from. Largely categorized in three buckets:

1. Predicting the right answers in form of continuous output – known as Supervised Regression Learning. E.g. predicting the volume of transactions during a specified period at a branch or ATM.
2. Predicting the right answers in form of discreet (or binary) output – known as Supervised classification learning e.g. classification of emails as spam or not.
3. Clustering the data and Identifying the new patterns and insights based on the clusters – known as unsupervised learning e.g. identifying market segmentation, social or customer behavior insights.

Clue – Think of the interesting problems around you and try to categorize in above three buckets for a solution.

Getting Interested? Here are some Fun facts and Starter links..

1. Try the “cocktail-Party-Algorithm”. Separating a singer’s voice from background music has always been a uniquely human ability. Check here, here and here.
2. Get detailed understanding of ML Algorithms. Although there is no single source to learn it from.
3. Get familiarized with one of the language Octave, MATlab, Python, R etc. Octave/Matlab are easiest to get started with and learn the concepts faster.

Leave your comments....any other source worth getting connected to...will love to hear. Lastly...it's always incomplete without Dilbert. :-)