Learning Experience \(E\)¶
Learning Objectives¶
| Prediction | Estimation of unseen data |
| Modelling/ Characterization/ Inference | How do inputs affect output Obtain the Sample CEF/Conditional Distribution which closely matches the Population CEF/Conditional Distribution |
| 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 Paradigms¶
| Method | Meaning | Application | |
|---|---|---|---|
| Supervised | Uses labelled data, to derive a mapping between input examples and target variable. | \(D_\text{train} = X, y\) | |
| Unsupervised | Learning from unlabelled data | \(D_\text{train} = X\) | |
| 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. | \(D_\text{train} = X, \text{Feedback}\) | 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 | |
| Online | Streaming |
Training Method¶
| Advantage | ||
|---|---|---|
| Batch | \(\hat f: X \to y\) | Better model |
| Streaming/ Online/ Passive-Aggressive | \(\hat f_b: X_{i \le b} \to y_{i \le b}\) where \(b= \text{Mini-batch}\) | - Adaptive to new data points - Computationally-cheap |
| Hybrid | - Batch training start of day - Online training intra-day |
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 |