What Is Active Learning Machine Learning

What is Active Learning: Active Learning is a special case of Machine learning, where the Model is provided with information on categories where the data is missing / where the model is nether performing. Equally per the definition from Wiki, the definition is "a learning algorithm tin interactively query a user (or some other information source) to label new information points with the desired outputs". Statistically it means, an experimental design, which is designed for a specific statistical criterion. The objective of the process is to find optimal data which meets the statistical criterion divers for the design.

Why practice we need it: Oft we come across problems where in that location is non enough labelled data, like any problem with highly unbalanced information, where you volition have very limited data or any Multi grade classification problem, where you lot might accept information for some classes, simply some classes will take very express data. In these scenarios labelling the data is very fourth dimension consuming and in scenarios like NLP based bug this is even more cumbersome and more frequent.

How does the solution piece of work: In simple terms, as shown in the epitome above you solve the problem and then attempt to refine. Suppose yous are trying to prepare for a test, what would y'all do? Either you read all the necessary topics for the test and so endeavour the test or you endeavor the exam first and identify which topics are you scoring meliorate, and which topics are you scoring poor and read the topics in which you scored poor. All of us have often came across this approach, now which i is ideal approach, may be the kickoff i, only which ane is faster, definitely the later on one. Applying same concept on Machine learning is called Agile learning. Same thing I tried to explain in the image of Michael Corleone dialogue above.

Active Learning Solution in Particular: Below solution explains the Agile Learning in detail. The steps highlighted in workflow diagram are explained below.

No alt text provided for this image

1) When we have a portion of labelled data and unlabelled information, in first step train the model on Labelled information and go the showtime iteration of Model.

2) Assess the model accuracy based on labelled information and identify the classes where the model is underperforming and define the statistical criteria to extract to label the nest set of labelled data

three) Send the statistical criteria to the custom agile learning model (Frequently an SVM Model / Exponential Gradient exploration) which is called Teacher or Oracle, to label the information based on the statistical criteria. Oracle uses custom query strategy and scenarios to identify the highly effective records and then labels them accordinly

iv) Send the labelled data to AI Model and retrain the Model using new incremental labelled data along with the labelled data from previous iterations.

v) Assess the model to define the model accuracy and define if the model met the criteria and if the newly labelled information met the statistical criterion.

half-dozen) If the Model met the criteria and the model see the last adequate criteria, the target model is generated. If the Model or newly labelled information didn't meet the criteria repeat the procedure to place the new labelled records from Oracle.

Few Notes (Interview Questions)

1) How is active learning different from reinforcement learning?

a. Reinforcement learning updates the model based on user inputs from the surroundings, where every bit active learning updates the surroundings past incrementally increasing the labelled data by labelling through ML model

2) What are different approaches how Oracle labels the information

a. Membership Query synthesis: where new labelled data are newly created or synthesized based on query criteria. Most widely used and popular algorithm for this is GANs (Generative Adversarial Learning)

b. Sampling (Pool-based and stream-based) : Two oftentimes used sampling approaches for this are pool-based, where oracle picks records from entire data set, and stream-based where each unlabelled record is candy to select and label the record

3) What is Query Strategy and how is information technology assessed?

a. Query strategy is the approach used to place (query) the right records to exist labelled by Oracle. Most commonly used criteria to identify the records is Entropy sampling query, margin sampling query and least confidence query

References:

https://en.wikipedia.org/wiki/Active_learning_(machine_learning)

https://world wide web.sciencedirect.com/science/commodity/abs/pii/S0031320311003463?via%3Dihub

Article : 2021-01