Machine learning, broadly defined, involves computer programs and algorithms that automatically improve their own performance at specific tasks through experience. The primary objective of these tasks, whether they involve classification or relating various inputs to expected
outcomes, is out of sample performance. The algorithm is first trained on a sample of inputs for which the target output responses are known. Training enables the algorithm to learn highly complex and intricate relationships in high-dimensional data, rather than pre-imposing assumptions on how inputs and
outcomes are related. Algorithm performance is then assessed on a different sample to determine how it would perform in the real world, with new data.
Enabling algorithms to discover relationships in the data has led to significant improvements in performance. The algorithms have the potential to replace costly manual, repetitive tasks and to complement problem-solving and decision-making tasks in the workplace. Moreover, in tandem
with the recent rise in data availability, machine learning is often used with datasets too large to inspect manually. The algorithms can be combined with dramatic improvements in computation speed, e.g., the use of graphical processing units (GPUs), to provide exceptional performance even with large
The applications of machine learning have been varied and numerous; the breadth and number of applications continues to rise rapidly. Recent applications, implemented by Analysis Group and others, have been extremely varied in nature.
- Labor Economics:
- Relating employment performance to online, pre-interview screening tests
- Relating management training of supervisors to survey-based measures of employee motivation
- Intellectual Property:
- Classifying patents and drugs in early stages of development, according to the likelihood of their future success
- Classifying large sets of electronic information, e.g., emails, attachments, or spreadsheets, based on the manual review and classification of a subset
- Identifying relationships between stock returns and various factors to develop stock-picking strategies