A primer to AI, ML, DL

Throughout my years of studying and applying AI to solve problems in my domain, I have always wanted to provide a concise overview of the field. This overview is not complete, but it does provide a glimpse into many of its concepts, terms, and techniques.

Day 1

What are the terms Artificial Intelligence, Machine Learning or Deep Learning actually cover?

Artificial Intelligence - The ability to perform tasks that are normally associated with human intelligence, such as perception, reasoning, decision-making, and problem-solving.

Machine Learning - Subset of AI that focuses on the usage of statistical techniques and developing algorithms that learn from data with little human assistance and make decisions.

Deep Learning - Subset of ML that uses artificial neural networks (analogous to human neuron system) to learn from data and perform human-like tasks.

So starting with ML - what are the building blocks for Machine Learning?

1. Thorough understanding of data - can the data be sliced and diced by dimensions such as customer types, product types, products sold by customer type etc. There is an integral part of understanding data - SME knowledge; never underestimate what you get
2. Trying to predict the outcome - New product launch, Offers on certain product segments. Would the sales increase in both of these cases?
3. Define Models - Build appropriate models and test for accuracy. This step defines the driver to achieve the business outcome. Product pricing, and discounts offered could be the output of this step.
4. Validation/Takeaways - Did the prediction help in achieving the business outcomes? If so, what went right or wrong.

Before getting into modeling let us look at EDA (Exploratory data analysis)

• Deeper understanding of data, correlation, insights, and application of domain knowledge; these steps help in building the right datasets, designing correct models
• Output of EDA usually encompass identifying key features, characteristics, correlations, missing data/anomalies/outliers and visualization of trends, and patterns with multiple dimensions
• Usually 50-60% of AI project is spent on cleaning, and exploring datasets

Let us now look at the 2 different flavors of ML - Supervised Learning & Unsupervised Learning

Supervised Learning - Primarily works with a labeled dataset to train algorithms to classify or predict trends.

Types of Supervised Learning

• Classification - predicts discrete class labels

1. Logistic regression
2. Decision trees
3. Random forests
4. Support vector machines
5. Neural networks

• Regression - predicts the discrete continuous value

1. Linear regression
2. Ridge regression
3. Support vector regression
4. Random Forest regression

There are a few terminologies that need to be understood before we start looking at the models.

• Bias - Errors in the model due to wrong assumptions. For example, data could be highly skewed, and assuming normal distribution will lead to Bias resulting in missing several key features. This leads to model underfitting.
• Variance - Errors introduced into a model due to the random variation in the data. Training with too little data usually leads to Variance and model overfitting.
• Loss function - A simple measure that shows the deviation between predicted and actual values. They are used in ML models to minimize the deviation.
• Recall - Proportion where the right values were rightly predicted (TP/(TP+FN)) This metric is primarily used where we cannot miss any positive instances like identifying a disease, tumor, etc. and the goal is to minimize False Negatives.
• Precision - Precision indicates how effective the model was in identifying positive samples among all the positives. (TP/(TP+FP)) This metric is primarily used where we need to minimize too many positives. To flag your email as spam (while it is not) might not be desirable, hence the goal is to minimize False Positives.
• Confusion Matrix/Classification Table - Performance measurement technique presenting Actual vs Predicted in a simple table.
• F1 Score - Predicts the accuracy of a model, leverages both Recall and Precision - 2 * (P*R)/(P+R).

Unsupervised Learning - Learns to predict outcomes based on unlabelled data. For example, look at an image and predict it to be a cat or fish.

• K Means
• Hierarchical Clustering

Day 2 - Will focus on the Classification techniques