Understanding Some Terms of Machine Learning Visually : Visual Challenge ML Scholarship Program for MS Azure 2020
I’ve recently participated in a visual challenge as a part of my task for phase 1 in Machine Learning Scholarship Program for Microsoft Azure 2020 Organised by Udacity and Microsoft. I’ve prepared some visuals/Diagrams to explain some terms/topics used in Machine Learning. I’m sharing them in this blog.
I’ve tried to draw diagrams and other visuals to explain following topics
- Difference between Machine Learning Algorithm and Machine Learning Model
- Difference between supervised and unsupervised learning
- Clustering
- Underfitting, Overfitting , Bestfitting
- Difference between Parametric and non-parametric algorithm
- Different types of Classifiers.
Machine Learning Algorithm VS Machine Learning Model
When Machine Learning Algorithm is trained with data it becomes Machine Learning model.
As you can see I have used equation of linear regression algorithm. it has coefficients B0 and B1. when it is trained with input and output, from training data it figures out the values for B0 and B1, for this case i assumed the value is 5, 6 . thus it becomes a model and now capable of predicting output (y) for unseen input (x).
Supervised VS Unsupervised learning
In this visual i tried to simplify highlight the key difference between supervised and unsupervised approach of machine learning .
In supervised approach data are labelled, in the first diagram you can see the model is trained with labelled data of Class A, Class B and Class C, and for new data the trained model predicts the Class.
In Unsupervised approach there data are not labelled. for second diagram you can see the model is trained with pictures of apple and grape. the model figures out the hidden pattern of the data and make a cluster of data having similar pattern.
CLUSTERING
In short clustering is just grouping the data having similar pattern together.
Clustering is an unsupervised machine learning approach, where models are trained with unlabelled data. when the clustering models are trained they figure out the hidden pattern of data and group the similar data together.
Underfitting VS Overfitting VS Bestfitting
Underfitting models tend to have high bias, Model with high bias make more assumption about target function, for the first character “ noodles “ can be represented as an incorrect assumption that model is making as a result it performs bad in both training and testing
In Overfitting, the models pick up all the noises from the dataset, and fails to generalize for new dataset. for the second character by “ Memoizing word by word ” he may perform good for seen data even get 100/100 but when new questions are given he may fail to generalize.
For Besfitting model, the models learn and perform well both in training and testing. best fitting models have balanced bias and variance error.
Parametric VS Non-parametric Machine Learning Algorithm
In parametric Machine Learning algorithm, the mapping function is assumed and and it has fixed number of inputs. for my diagram i represented the function with Y=B0 + B1X1 + B2X2.
Example: Linear regression Algorithm
Non-parametric Machine Learning algorithms are flexible, mapping function is not fixed, the inputs are flexible.
Example: Deep learning algorithms
Different types of Classifiers
So there are mainly 3 types of classifiers.
For Binary classifiers there are only 2 classes and the output belongs to one class only.
Example: suppose you are training a model to identify the image of dog and cat. there are 2 classes and the output will either be dog or car but not both.
For Multi-class single label classifiers the class number is more then 2, and output belongs to only one class.
Example: suppose you are training a model to identify the image of dog, cat and rabbit. there are 3 classes and the output will belong to only one class
For Multi-class multi label classifiers the class number is more then 2, and output can belong more the one class.
Example: suppose you are training a model do assign tags for articles based on the contents , one articles can have multiple tags
