Machine learning is a type of artificial intelligence (AI) that provides the ability of a system to use data and experience to improve its performance. This entails a system learning without being programmed. Simply put, machine learning is built on data, algorithm and insight.1
You might already be familiar with how machine learning impacts your every day—from the Amazon shopping experience that predicts what you are interested in buying, to Netflix’s ability to predict how you will rate a film or TV show to the iPhone’s Siri feature.
Beyond the Everyday Consumer
While machine learning can make for an efficient customer experience and provide companies with insight regarding customer behavior, this technology has many real-world applications that go far beyond retail.
In healthcare, machine learning is being used to speed up medical diagnostics and improve accuracy and accessibility, create personalized and affordable means of delivering mental health care, identify and prevent prescription errors in real-time and help hospitals and insurance carriers provide better care more efficiently.2,3
Machine learning is also being used in the financial industry as in the case of Goldman Sachs’s AppBank project, which gives computers the ability to learn once human-oriented tasks to run a bank’s systems. If successful, the goal here is to reduce operational risk.4 Feedzai, a company that protects financial firms and merchants against fraud, uses AI algorithms that automatically adapt to business patterns and get progressively more intelligent. Their goal is to enhance their customers’ experiences and make commerce safe.5
In an effort to find villages in need and donate to them, nonprofit GiveDirectly used machine learning and satellite imagery in 2013 to identify the proportion of thatch versus metal roofed homes—thatch being indicative of poverty. Though the algorithm ended up not being accurate enough for GiveDirectly to use on a larger scale, this type of application demonstrates the possibilities of improving humanity one model at a time.6
Another promising example involves UK researchers currently working on a technology model for lip-reading that has already shown a higher degree of accuracy than human lip-readers. Once perfected, this technology could be leveraged to aid people with hearing impairment, automate subtitles, or assist with poor audio quality on a video or mobile call.7
Uncertainty and Best Practices
Though machine learning has potential to positively impact life in so many ways, there is a high level of risk involved. One issue is that the “rules of thumb,” patterns machine learning uncovers from historical data, are not always correct—which could be harmful to a business, person or product. Machine learning might also sometimes draw from historical data that displays racial biases or similar issues, so models should be aware of these things.8
Another large issue is that machine learning can accumulate large amounts of technical debt. Best practices for machine learning include:
- Setting up a source code repository
- Operating over immutable data so that a model can be retrained any time in the future
- Monitoring performance to keep an eye out for significant deviation from expected results
- Regularly retraining models on new data
- Automating retraining, publishing and monitoring the pipeline as often as possible9
Despite potential risks and challenges, machine learning continues to flourish and field experts expect it to remain in high demand.10 The full extent of possibilities remains to be seen, but knowing that machine learning can easily outperform traditional methods on accuracy, speed and scale in certain instances11 will change society as we know it.