Through the buzzword jungle
Artificial Intelligence (AI), Machine Learning (ML), Data Science, Big Data, Deep Learning, … Those are quite some terms! As a matter of fact, with so many terms, it almost feels like being trapped in a jungle of buzzwords when entering the adventure to find the most beneficial data-driven use cases in a company.
A nice rule-of-thumb to get an overview in this complex jungle is to state that for modern AI applications, Supervised Learning (SL) is an integral part in most cases. SL is a subset of AI, in which models — such as deep neural networks — learn by examples. Let's take a look at how this works.
Automated pattern recognition
For starters, let’s think about how we could implement a use case such as a risk-analysis model for an insurance company. Using regular programming techniques, we could implement some domain knowledge, such as that people driving sports cars have a higher risk to get into an accident than people driving family cars.
A different approach would be to collect historic data for risk investigations, and to do some statistical comparisons, e.g. comparing the risk amongst different car types. This is a manual data analysis to search for patterns. SL aims to do such an analysis in an automated manner by applying a model on the historic data, which recognizes the data-inherent pattern structure.
The definitive advantage of SL is that such pattern recognition is automated. However, this advantage comes at the cost of needing labeled records. A label classifies a given data record, e.g. that client “John Doe” is a riskful client. Without such labeled records, no SL algorithm can learn to recognize patterns.
This would not be too much of an obstacle, if the learning models were as data-efficient as us humans are. Sadly, these SL models are not. To correctly recognize and later on apply a pattern with great accuracy, models like deep neural networks often need more than 10,000 labels.
In fact, deep neural networks excel in modern AI applications as they are capable to learn on massive data sets containing millions of labeled records. For instance, if a deep neural net first learns on 100,000 labeled records, it may achieve an accuracy of 95%. If it is trained again at some later point on 500,000 records, it could achieve an accuracy of near 99% — without changing one line of code.
Supervised Learning (SL) algorithms automatically detect patterns in any kind of data — but to do so, these algorithms must first see many labeled samples. In use cases such as intent classification, models like deep neural networks are provided with pairs of sentences (“Please set an alarm for 6 am”) and the desired outputs for given inputs (“set alarm”). Unlike human learners, SL algorithms need lots of examples. Therefore, for SL to excel in the way it is meant to be, a correct labeling strategy is crucial to successful implementation.