What is Machine Learning?
I am frequently asked by friends what machine learning is. Here is a very simple helicopter view of the world of Machine Learning.
Before I begin, lets motivate Machine Learning.
Why do we do Machine Learning?
Artificial Intelligence (AI). Machine Learning is a field that develops AI. Simply put, AI is the simulation of human intelligence by machines. A self-driving car, a fraud detection system, or cancer prediction are all examples of AI.
These tasks require human intelligence. The benefit of AI is we can improve human intelligence. We can drive with better accuracy. The AI will not need a map, it has already processed it. And, can respond to sudden changes in the road faster than a human. AI can detect fraud within higher volumes of data than a human. And predict a cancer risk before the individual is afflicted with cancer.
The intention of AI is not to replace human intelligence, but to supplement it — augmented intelligence.
How is machine learning used to build an AI?
It is explained in the name. In machine learning we teach machines / computers to learn without explicitly programming it.
There are various domains in which we do this. Here is a popular, but not exhaustive list of sub-disciplines within machine learning:
Natural Language processing: machines learn from text so they are able to interact with human languages. Examples are: machine translation, sentiment analysis, question and answer programmes.
Machine Vision: the machine interacts with and replicates human vision. Machines are taught from images or videos.
Human Computer Interaction: machine interact with human emotion. The machine typically learns from kinetic motion, facial expressions or changes in physiology to learn human emotion.
Reinforcement learning: learns human behaviour to teach machines what action to take.
Graphical models: the machine learns from graphs, things like social networks, social physics. An example is Google’s page rank.
You have audio-processing which is machine interaction with sound, and there’s probably smell, but I haven’t really seen that yet!
Often, many of these sub-domains come together to build an intelligent system. More often than not however an organisation will likely only use a very small application of machine learning to solve a problem. An example, would be a marketing company sending targeted ads to a cohort of consumers.
How do we do it?…
…data! Big data, small data, analytics. It’s all about the data. If it didn’t make sense before, I guess by now it makes sense why everyone’s been talking about big data. Big data now means we can build systems that work with better accuracy and faster.
The Process at work
Below is the typical process for which we can build an AI. Often times we go back and forth between steps to improve the model accuracy score.
- Define a problem.
- Find a data set.
- Data Pre-processing: This step prepares the data to be processed by your model. It tends to require a thorough investigation and preparing of the data. Some methods involve visualising the data, understanding distributions or correlations of variables, feature selection, feature engineering, perhaps dimensionality reduction.
- Train Model. Obtain a performance score.
- Model Selection. Accuracy Evaluation, maximising the score obtained from Step 4.
- Tuning. Requires adjusting the parameters in the model to obtain the highest accuracy results.
- Test. How does the model do with unseen data.