Artificial Intelligence and Machine Learning stand to be the most sought after professions of nowadays and both freshers and experienced professionals are jumping to acquire skills in this field to make their profile shine and stand out. Loads of online and on-premise courses are being launched to certify students and professionals in this area of specialization.
All of that is great and while the trend has boomed so drastically, is AI really that new? Is AI/ML really something that got discovered today given the hubbub about this technology in the recent times? Well that's not true. AI has existed for however many years the concepts of Linear Algebra, Calculus, Matrices, Probability and Applied Mathematics and Statistics were discovered and was also used at a good scale since the early 1950s or may be even before that it was officially introduced as a term.
Here in this blog, we will be discussing about what AI and ML is. It's background and why do we really need it? Sounds mundane does it? But don't be so sure yet.
So What is AI?
It is making a machine able to recognize patterns and behavior of a given type of object based on similar object data samples provided to it. This data called the "training set" contains a lot of parameters on which that object is described so as to capture the pattern accurately. This training set data is fed into set of methods described as training algorithms which after training with a sufficient amount of data is able to predict and make decisions about about new set of data termed the "test set".
A formal way of saying that would be "to train machines to think and make decisions like human beings is called Artificial Intelligence and the process of doing so is Machine Learning". The ability of an algorithm to characterize new test data after learning over the training set data is called "generalization".
Data Sanitation
This is the pre-processing stage of data, before the data is fed into a learning algorithm to train it. This involves various techniques like "dimensionality reduction" where say we reduce the dimensions of an object from 2D to 3D OR "Feature Scaling" where each variable is value is divided by a common denominator to make is smaller. This brings all the training set data to the same level and makes the learning algorithm computations much faster, efficient and accurate.
Pre-processing needs to be applied to the test set data as well before we attempt to make determinations or predictions on any new set of data. While scaling the feature or reducing dimensions, we need to exercise care that while tailoring down the variables, we don't end up discarding features or dimensions that are critical to the definition of the object else the quality of the algorithm gets drastically compromised and we end up given the wrong predictions.
Data sanitation techniques will be discussed in more details in the upcoming posts.
Types of Machine Learning problems
Before getting into the details of the types of machine learning problems, let us first define some terminologies which are an important component or any learning algorithm.
Vectors in mathematics are objects which have both a magnitude and direction. A vector space is a linear space of such objects held together by the vector addition and scalar multiplication rule. In machine learning terms, we have the following:-
An input vector x = {x1, x2, x3, ......, Xn}
A target vector t = {t1, t2, t3, ....}
We now come down to our explanation of the various types of machine learning problems. Machine learning problems are of the following types:-
Supervised Machine Learning
In this problem, the input vector x is the list of variables that describe various features of the object and vector t is a set of desired values obtained after evaluations done on vector x. In short vector t and x are connected by the hypothesis:-
t = h(x) {which means t is a function of x as represented in the figure below}
Depending on the type of output the Supervised Machine Learning Problem is further subdivided into 2 parts:-
Regression Problem - The targer vector t in this case is a value among those falling on the trajetory of a continuous curve. Mathematically speaking, in the regression problem, the hypothesis t=h(x) will be such that, t is a continuous function of x.
For example:- In an advertising firm, the analysis done for a sales of a product being affected by a media and ad campaign is an evaluation done on various factor such as:-
The mediums of advertising,
The amount of space bought in newspapers,
The demographics being targetted,
The number of times the commercial will be telecast,
The complaince limitations for the campaign based on countries involed if it is an international campaign and so on.
What is also taken into account are the patterns of sale changes based on historical ad campaigns.
Here the target vector t will be a sales projections with the current campaign. This is a regression problem.
Classification Problem - Here the target vector t is a defined a set of discrete non-continuous values for a set of input vector x. The output vector can be [0,1] or a set of natural numbers N or so on.
For example:- The overall rating calculated for a program based on the absolute average of total ratings is the number of stars given to that program or movie. Say 1,2,3,4 or 5 stars. If we have to predict the rating on the scale of 1 to 5 stars for a program or movie based on the ratings so far, then the output will be another star rating between 1 to 5. This is a classification problem.
Un-Supervised Machine Learning
In this problem, the input vector x is given to us without any target values vector t. Hence our hypothesis takes the form:-
h(x) = ?
hence the name unsupervised learning, as there is no target vector present. This problem is tackled by making groups of similar data objects within the training data set. This technique is called "Clustering". Then we determine the density of objects present in each on the these groups. This process is called "Density Estimation". This also involves visualizing the data points from a high-dimensional plane down to lower dimensions like 3D.
Reinforcement Machine Learning
These are reward maximization problems. This corresponds to games where we know the rules but we do not know all the moves which can get us to the best possible output. Here the algorithm trains itself by taking steps and observing the output. This happens with the process of trial and error. Some moves are good and give great reward and algorithm has to do "Credit Assignment" for them.
Like every game where there is risk and reward, these algorithms work on tradeoffs between "Exploitation" and "Exploration". Exploration is process of exercising new moves to see how rewarding they are and exploitation is where the algorithm runs tried and tested moves which have given good rewards in the past.
The catch of these algorithms is that too much of exploitation or exploration can lead to bad results like any game we play.
In Closing......these algorithms sound like our life. How we think things may or may not happen in our lives. Sometimes we take calculated decisions based on what we have seen or heard in the past. While there are others like going to a new place, or getting married, a leap of faith where we do not what the future beholds. This is what makes AI so life like and so so real.
Even the maths of it is the poem of life! :)
Stay tuned while we bring more in depth concepts or AI in our blog. I will try to keep things as simple as I can but still do let me know in comments what needs more elaboration or what can I improve. If you wish to connect with me directly. Please email me @ priyadarshani.pandey@gmail.com OR connect with me on LinkedIn - Priyadarshani Pandey | LinkedIn