Introduction and artificial intelligence
Machines that can learn and adapt marks a massive step forward for computing, but who's teaching them? Computer systems only get smarter when they get data – from where and from whom they get that data is critical. Machine learning researchers have so far focused on how they can build better, faster and more precise algorithms, but what do data scientists and computer programmers know about the world? What machines really need is to be trained by experts in every field of human endeavour.
What is machine learning?
Machine learning is essentially all about getting computers to take the initiative without input from humans. They do this by statistical learning after identifying patterns in data.
While the basic idea of machine learning is very simple, its execution is complicated. "Thanks to machine learning, your email inbox is mostly free of spam and other unwanted email, and your smartphone can constantly improve its understanding of what your personal needs are based on what you say and do," writes Microsoft, also citing the real-time voice translation found in the upcoming Skype Translator and Cortana, and even Bing.
"In essence, machine learning runs through various possible actions and predicts which action will be most successful based on the collected information," says Dr Kevin Curran, IEEE Technical Expert. "Of course, the computer can only solve problems it is programmed to solve as it does not have any generalised analytical ability."
Artificial intelligence
Everyday examples of machine learning include search engines ranking web pages, your smartphone producing a map of where you've been using geo-tag data from the photos you've taken, automated spam filters, spell correction in word processing software, face recognition in cameras, speech recognition by all virtual personal assistants, and all kinds of recommendations while shopping, browsing the web or streaming music or movies. Its next destination is the connected car.
"Machine learning has become a hot area within computer science in recent years," says Curran, who name-checks autopilot and the magnificent gyroscope ability of Segways as places where machine learning algorithms are running. "In fact, many large computing giants are spending on big salaries to attract those with machine learning experience."
When ML goes wrong
But machine learning – a form of artificial intelligence – does not always work. Take Uber, which has an algorithm that responds to high demand by raising the price. That makes perfect business sense under normal conditions, but is quadrupling the price of a taxi during a siege really the kind of thing a human-run business – aware of the public relations implications – would ever do?
Uber's algorithm, and hundreds like it, needs some human-like morality in the form of a specific model that relates to real-world scenarios, not just the basic number-crunching of a data scientist. Ditto Siri's habit of relying only on keywords, thus replying to calls for help with alcohol and gambling with details of nearby off-licences and casinos. A business knowingly doing that would be called psychotic.
Machine teaching
The next phase
In a video posted in July, Microsoft Research described its efforts to help people without a machine learning background to teach their systems to learn from experience. "No one has really built a machine learning tool for the layman," said Patrice Simard, Distinguished Engineer and a Deputy Managing Director at Microsoft Research, who is leading a new machine teaching research project. His aim is to make the process of training a machine 'easy, fast and universally accessible'.
"Surprisingly, machine teaching is neglected by ML practitioners both in the academic and the industrial communities," says Simard, who calls machine teaching a 'paradigm shift'.
What is machine teaching?
Machine learning's in-tray is bulging. So good is machine learning, goes the argument, that there are not enough people with machine learning expertise to work on such projects. Machine teaching is a specialised subset of machine learning that attempts to find an optimal training set given a machine learning algorithm and a target model. Say what? "Machine learning is where software algorithms 'learn' to typically 'classify' data in an 'intelligent' form," says Curran.
What is Azure Machine Learning?
Stuffed with tried-and-tested algorithms, Microsoft's Azure ML is designed exclusively for data scientists to use, largely to save time on repetitive tasks. The goal is to produce a machine learning model quickly, and to host it in the cloud for easy accessibility. Niche industry will benefit. For example, with over 400 possible errors, ThyssenKrupp's 1.1 million lifts/elevators are tricky to diagnose and fix, so they feed data into an intelligent information loop created using Azure ML that displays real-time performance on an online dashboard, and tells its technicians how to fix them. At its best, it's predictive, and even pre-emptive.
Machine learning for the masses?
It's time for data scientists and machine learning experts to get out of the way. Microsoft Research wants to widen the field of people who can create, teach and maintain computer models. "Normal non-techies could potentially be 'teaching' machines in the near future, but the reality is more likely that they will be using bespoke highly tailored niche computer applications which are tailored to certain vertical markets that help these people 'teach' machines," says Curran. That still means capturing the as yet untapped 'analogue' expertise of millions of high-levels experts, academics, and others.
How important is Microsoft's involvement?
At present, machine learning is in the hands of Google. That's hardly surprising since language, speech, translation, and visual processing – an evolving search engine's bread and butter – all rely on it. Many of the leading AI and machine learning researchers work for Google. "This effort by Microsoft is to be welcomed," says Curran. "Entry into the field of machine learning is not trivial, and tools to make it easier for non-experts to adopt machine learning should be welcomed – but the research is only in its infancy."
- How Microsoft's machine learning algorithms will make for smarter apps
from www.techradar.com
