Demystifying AI, ML, automation:

Usually, AI/ML is a buzz phrase worldwide for a tech or non-tech user segment of the global population.

“AI” and “ML” are sometimes used interchangeably, and any automation is termed “AI-powered” these days (I would even term it a marketing gimmick). However, every automation is not AI.

Before diving into the types of Machine learning algorithms and how ordinary people are connected to them in their daily lives or how multiple industries use these algorithms in their product features, let’s first clarify Automation from AI and ML from AI.

So, do we need AI everywhere? The answer is no; it depends very much on the specific use case.

So, automation and AI are two different topics. An AI capability may or may not be automated, and automation may not be related to AI.

Let’s take an example of a spam filter application in your Outlook or Google Mailbox or any other filtering application any product company builds.

If you already know what needs to be filtered out or rejected, a programmer can take a traditional logic/coding approach (It’s automation, not AI). In this case, you apply filtering logic based on your preference.

If Products must filter out an unpredictable set of emails, then it must be decided based on user preference. Learning user preference thru email consumption behaviour is necessary for the app to filter intelligently. “AI-powered App” is relevant in such cases.

Deep Learning is, further (neural network based) study within field of ML.

We can say,

• Artificial Intelligence (AI) = Machine Learning (ML) + action based on ML algorithm to generate a prediction

AI is an app / thing that can mimic human behaviour

• Machine Learning (ML) = Classic Machine Learning + Deep Learning (Neural network-based learning)

ML is the decision-making process behind the action an AI app / thing will take.

Machine Learning systems classification.

3 different ways, Machine learning approach is classified.

• The way a ML system is trained.
• The way an ML model learns
• The way the dataset is used for training.

Training method-based ML algorithm classification & usage:

4 ways to train ML models, either its Supervised, Semi-supervised, Reinforcement learning or Unsupervised learning.

Supervised Training:

 The dataset for training AI apps, it needs pre-tagging or labelling, or historical data which serves as learning reference point.

e.g.: classic ML

• Classification: Identification of emails for spam detection, identifying an object etc.
• Regression: Forecasting future sales in a company etc.
• Nural Networks: Identifying comments for expression (NLP) etc.

All the above needs some reference (tagging or historical data) to make decisions on, hence its supervised training method.

There are other such algorithms, which we will not go into in depth (K-Nearest Neighbours (KNN), Logistic, Linear Regression (there are subtypes of Regression), Support Vector Machines (SVM), Decision Tree, Random Forest, Neural Networks (CNN, RNN, etc.)

 Unsupervised Training:

This method of training does not require historical data or pre-labeled information.

Data is required, but none of the features/attributes are labeled with a past value.

This type of algorithm scans the data to find affinity in the provided data points, so a grouping can be done by looking at the data; the model can establish a natural affinity in the data.

Clustering/ Hierarchical Clustering: categorizing / Sub categorizing YouTube comments. (K-Means, DBSCAN, Hierarchical clustering-HCA)

Anomaly & Novelty detection: Identifying outlier or Odd one out. Pick defective products in the production factory floor or finding a new category of commenter group in a post. (Isolation forest, single class SVM)

Association Rule: Product bundling for targeted marketing based on customer purchase data, Coke & Burger, Fish & chips. Even sequential upsell can be achieved (car sale & offering accessory deals)

(statistical methods such as Apriori, Eclat etc.)

Semo Supervised: 

It’s a combination of a supervised and unsupervised approach & it works with labeled & unlabeled datasets. 

It can reduce the cost of labeling the entire training dataset.

For example, in a cloud-based Photo upload service, once you tag/label a photo (each native in the photo), the service can tag the same native in upcoming images, too.

Reinforcement Learning:  

We will not deep-dive into this type of learning, as it needs detailed focus, but if I summarize it, it’s like training a child.

A child needs validation after acting. You let the baby know whether it’s OK to do or not. Baby decides on future action by referring to your reaction to her action.

A similar training method is applied to Robots through scoring.

In context of learning system, below image explains the actors & actions. Policy definition helps select response to actions. It helps with validate the action with positive or negative scoring.

Learning method-based ML algorithm classification & usage

Another way ML systems are classified is in context to what mode the model learns the data.

I will not provide too much detail here (Not necessary for people who are not data scientists)

Online Learning:

Some models can learn incrementally in micro batches without tuning the AI model instance off.

Batch Learning:

If the training data is large or requires the entire dataset for training, the model learns in a non-real-time fashion, which you may call offline learning.

Dataset based ML algorithm classification & usage.

Let’s remember some of our friends in school or university days.

Some learned by scanning every sentence in the book (First benchers in school usually 😊)

Some learned by learning the pattern of questions & deriving the formulas (the Einstein(s) in the class 😊)

The first kind is known as Instance-based learning,

The second kind is known as Model-based learning.

This quick and short article should give you an overall understanding of how AI works, where it’s used, and what is and is not AI.