MATHEMATICS REQUIRED FOR ML / DL /AI ?

WHAT MATHEMATICAL CONCEPTS YOU NEED TO HAVE BEFORE TAKING UP COURSES ON ML /DL /AI

THERE ARE PREDEFINED LIBRARIES IN PYTHON WHICH MAKES CREATING MODELS SUPER EASY FOR ANYONE .FOR A BETTER UNDERSTANDING OF WHATS ACTUALLY HAPPENING BEHIND THE SCENES YOU NEED TO HAVE COMMAND OVER THE FOLLOWING MATHEMATICAL DOMAINS.

SINCE “DATA SCIENCE” IN ITSELF IS A VAST DISCIPLINE .LETS DISCUSS THE MOST COMMON DOMAINS AND THEIR CORRESPONDING REQUIREMENTS ONE BY ONE

SUPERVISED AND UNSUPERVISED MACHINE LEARNING

FOR UNDERSTANDING THE SUPERVISED LEARNING ALGORITHMS YOU NEED TO HAVE A GRASP ON BASIC EQUATIONS OF LINES AND SHAPE OF POLYNOMIAL CURVES UPTO 5-6 DEGREES . SUPERVISED LEARNING PROBLEMS CAN BE REGRESSION OR CLASSIFICATION PROBLEMS . HERE THE WORD “SUOERVISED ” SHOWS THAT YOU HAVE A LABEL CORRESPONDING TO EVERY ENTRY IN THE DATA SET . HENCE EVERY TIME YOUR MODEL MAKES A PREDICTION YOU HAVE SOMETHING WHICH TELLS YOU HOW WELL YOUR MODEL HAS PERFORMED AND HOW FAR YOU ARE FROM THE DESIRED PREDICTION . A FIRM KNOWLEDGE OF STATISTICAL TERMS LIKE VARIANCE, MEAN , BIAS , RMS ,GAUSSIAN DISTRIBUTIONS ARE REQUIRED. FOLLOWING ARE THE MOST COMMON ALGORITHMS USED :

1. LINEAR REGRESSION : FOR CREATING A MODEL THAT PREDICTS CONTINUOUS VALUES
2. LOGISTIC REGRESSION

UNSUPERVISED LEARNING AS THE NAME SUGGESTS , DOES NOT HAVE A LABELED DATA SET . IT IS JUST A COLLECTION OF FEATURES . MOSTLY UNSUPERVISED LEARNING CONSISTS OF CLASSIFICATION PROBLEMS .

. FOR UNDERSTANDING UNSUPERVISED LEARNING ALGORITHMS YOU MUST KNOW ABOUT TREES , PROBABILITIES , LOG FUNCTIONS , ENTROPY AND SOME BASIC COORDINATE GEOMETRY AND MATHEMATICS. STATISTICAL KNOWLEDGE IS REQUIRED TOO. KNOWING ABOUT ENSEMBLES AND AVERAGE MODELS IN MATHEMATICAL MODELLING WILL HELP WINNING DATA SCIENCE COMPETITIONS. ENSEMBLING HELPS MAKE BETTER PREDICTIONS .

REINFORCED MACHINE LEARNING

IN REINFORCED LEARNING YOU NEED TO HAVE STRONG MATRIX VISUALISATION SKILLS , A DYNAMIC THINKING APPROACH , AND GOOD AMOUNT OF REASON . THE REINFORCEMENT LEARNING METHOD SOUNDS VERY STRAIGHT FORWARD BUT THE ALGORITHMS USED ARE DIFFICULT FOR FEW PEOPLE TO UNDERSTAND . REINFORCEMENT LEARNING IS A RATHER VAST DOMAIN THAT CONTAINS VARIOUS ALGORITHMS .

DEEP NEURAL NETWORKS

NEURAL NETWORK , AS THEY GET COMPLEX , DEAL WITH HUGE MATRICES ,WITH DIMENSIONALITIES IN THOUSANDS AND MILLIONS. SO ONE NEEDS TO HAVE STRONG MATRIX VISUALISATION SKILLS . WHEN UPDATING THE WEIGHT PARAMETERS OF THE MODEL (TRAINING OF THE MODEL) NEURAL NETWORKS WORK ON TWO BASIC OPERATIONS ,FEED FORWARD AND BACKPROPOGATION .

THESE TWO PROCESSES ARE COMPLETELY DEPENDENT ON GRADIENT CALCULATION WHICH FALSE UNDER VECTOR CALCULUS . FOR UNDERSTANDING THESE CONCEPTS YOU NEED TO HAVE KNOWLEDGE IN DOMAINS LIKE CALCULUS .

MOSTLY DERIVATIVES ,PARTIAL DERIVATIVES AND CHAIN RULE.YOU NEED TO BE THOROUGH WITH PROBABILITIES , STATISTICAL TERMINOLOGY AND CONCEPTS ALONG WITH CONCEPT OF ENTROPY . AGAIN I CANNOT EMPHASIZE ENOUGH THE IMPORTANCE OF MATRICES AND VECTORS .

VECTOR CALCULUS

A NEURAL NETWORK IS ALL ABOUT GRADIENTS !!! UPDATING WEIGHTS OF THE NEURAL NETWORK LAYERS ARE ALL BASED ON GRADIENTS . GRADIENTS REFER TO THE VECTOR REPRESENTING THE MAX SLOPE AT A PARTICULAR SURFACE AND ITS DIRECTION . VISUALISING GARDIENTS IS AN IMPORTANT ASPECT OF UNDERSTANDING NEURAL NETWORKS . GRADIENT DESCENT IS THE STARTING POINT OF NUMEROUS ALGORITHMS THAT TRY TO OPTIMISIZE THE LEARNING PROCESS . BUT WHAT DO WE OPTIMISIZE?

ONLY MAKING CORRECT PREDICTIONS IS NOT THE FACTOR ONE NEEDS TO CONSIDER . ONE IMPORTANT FACTOR IS HOW MUCH COMPUTATION ONE NEEDS TO DO IN ORDER TO DO SO . OPTIMISING THIS IS ALSO IMPORTANT . VARIOUS ALGORITHMS LIKE MOMENTUM ,ADAM , ADAGRAD CATER TO THESE NEEDS .