Learning Experience \(E\)¶
Goal of Experience¶
-
Learning/Inference
-
Obtain the Sample CEF which closely matches the Population CEF
or
-
Obtain the Sample Conditional Distribution which closely matches the Population Conditional Distribution
-
Prediction/Decision
-
Making prediction of \(y\) based on the learning
Learning Types¶
Prediction | Estimation of unseen data |
Modelling | Characterization How do inputs affect output |
Optimization | What input values produce desired outputs (both mean and variance) |
Control | How to adjust controlled inputs to maximize control of outputs |
Simulation | |
Causal inference |
Use ML models for developing structural models, and then let the structural models to make the predictions, not the ML models
- Why: Black swans can be predicted by theory, even if they cannot be predicted by ML
- How: Use a non-parametric ML to identify important variables and then develop a parametric structural form model.
Learning Methods¶
Method | Meaning | Application |
---|---|---|
Supervised | Uses labelled data, to derive a mapping between input examples and target variable. | |
Unsupervised | Learning from unlabelled data | |
Semi-Supervised | \(\exists\) labelled data and large amount of unlabelled data. Label the unlabelled data using the labelled data. For example, love is labelled as emotion, but lovely isn’t Cotraining, Semi-Supervised SVM | |
Lazy/Instance-Based | Store the training examples instead of training explicit description of the target function. Output of the learning algorithm for a new instance not only depends on it, but also on its neighbors. The best algorithm is KNN (K-Nearest Neighbor) Algorithm. Useful for recommender system. | |
Active AL | Learning system is allowed to choose the data from which it learns. There exists a human annotator. Useful for gene expression/cancer classification | |
Multiple Instance | Weakly supervised learning where training instances are arranged in sets. Each set has a label, but the instances don’t | |
Transfer | Reuse a pre-trained model as the starting point for a model on a new related task | |
Reinforcement Learning RL | Learning in realtime, from experience of interacting in the environment, without any fixed input dataset. It is similar to a kid learning from experience. Best algorithm is Q-Learning algorithm. | Game playing |
Bayesian Learning | Conditional-probabilistic learning tool Each observed training expmle can incrementally inc/dec the estimated probability that a hypothesis is correct. Useful when there is chance of false positive. For eg: Covid +ve | |
Deep DL | Multi-Layered ANNs | Computer Vision |
Federated FL | Distributed | Privacy |
Types of Learners¶
They are not adapted by the ML algo itself, but we can use nested learning, where other algorithms optimize the hyperparameter for the ML algo.
Eager Learner | Lazy Learner | |
---|---|---|
Training | Learns relationship between class label & attributes | Stores training records |
Evaluation | Perform computations to classify evaluation record | |
Training Speed | Slow | Fast |
Evaluation Speed | Fast | Slow |
Example | - Decision Tree - Rule-Based Classifier | - Nearest-neighbor classifier |