What is Machine Learning?
Machine Learning (ML) allows computers to learn and improve from experience, without being explicitly programmed for every single task. Machine learning algorithms build a model to make accurate predictions and decisions. The data used to build such a model is called training data. Today, ML is being used in every industry. ML is behind from bots to social media feeds, from driverless vehicles to medical diagnosis.
Subcategories of Machine Learning
There are three subcategories of machine learning:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
Supervised Learning
Supervised learning is where training is done on a labeled dataset. The machine tries to create a mapping between inputs and labels. The outputs can be a limited set of values (classification) i.e. classes or a range of numerical values (regression). For example, classifying emails into spam and not spam.
Unsupervised Learning
Unsupervised learning is where training is done on an unlabeled dataset. The machine tries to identify patterns, trends, or clusters in the data. For example, identifying types of consumers based on purchases they made.
Reinforcement Learning
Reinforcement learning is where a reward system is used to train the machine. The machine tries to learn the best action based on trial and error. For example, learning to play games or driverless cars.
Limitations of Machine Learning
Many times machine learning where it fails to work as expected. The top limitations of machine learning are:
- Bias
- Overfitting
- Explainability
Bias
Bias happens when the model is too simple. For example, if a prediction has to be made based on age and gender, and the model is just using gender to predict, model will not be accounting for influence of age.
Biases can be present in the human labeled data too. For example, if a person is biased against a certain kind of people. This biasness would reflect in the labelling too.
Overfitting
Overfitting happens when the model is too complex and too close to the training data. It captures all the noises in the training data, and fails to capture the patterns. For example, if we have limited data, model can fit with the training data well but would not perform good on new data.
Explainability
Business needs to understand how machine learning models are making decision. As models can be affected by the biases, there is a need for business to validate how certain decisions are being made. For example, let's say a model is trained using only brown horse images and only black cat images to identify horses and cats. If this model has correlated color with the animal, it will give 95% accuracy if it is tried upon 95 brown horses and 5 black horses. But the decision would be based on the color and not on how the animals look. If a explanation of such a decision is present, it will be possible for business to improve.