A traditional computer operation takes a set of directions (the algorithm) to make a calculation or to trawl a data table to find the piece of information a user is looking for. It isn’t a question of what is likely. It’s a question of what or where the answer is. Machine learning is different.
Patterns, not answers
Machine learning isn’t trying to discover a specific point, like the proverbial needle in a haystack. It is instead trying to figure out the likelihood of something. Its algorithm uses statistical analysis and probability to make informed guesses about what the output should be. That output may have a range of ‘correct’ answers. It does this by ingesting a set of training data, which it uses to seek patterns. The more training data it has, the more precise the patterns it can detect.
The training data might be anything—documents, photographs, maps or traditional data tables, for instance. In the case of a business, the training data could be rich in industry-specific vocabulary or imagery to help the AI get up to speed with what the audience expects. Machine learning can then pick information out of complex data and display relationships that would be hard for a human to conceptualise.
Some things are easy to predict: the weather in Arizona in the summer. Some things are difficult to predict: the weather in England in the summer. Machine learning AI can aggregate decades of data on ocean currents, storm cycles, climate change, satellite imagery, precipitation patterns, topological features, volcanic ash and more to predict what’s going to happen this week in your town. And it also helps scientists generate long-term climate models.
Sometimes data scientists want machine learning to go beyond just calculating the odds, as with weather and climate modelling, and ask it to act on the probabilities. In this situation, the machine learning AI must be robust enough to make a decision like a human would, only better.
Just as a traditional computer can calculate and search a database faster than a human can, a machine learning computer should be able to predict the odds and then make a decision faster than a human can—a decision more logical than a human ever could.
In this respect, machine learning is still highly experimental. Trying to get a computer to predict human behaviour sometimes gives more insight into how strange we are than into how the AI works. This trickiness of human behaviour is one of the major challenges in autonomous vehicle technology, a very practical application of machine learning. In spite of that, self-driving is moving ahead (no pun intended) as a logistics solution that claims to be safer and greener than human driving.
Azure Machine Learning is a cloud-based AI studio where you can manipulate a dataset to try to predict what’s going to happen next using Azure’s toolkit. This might be fluctuations in materials prices or the chances of converting a lead. Or it may be as complex as choosing a geophysically sound site for a new project phase.
There are many different possible ways to experiment with the data in Azure Machine Learning. Most of them require expertise to get the most from the operations as well as an absolutely clear vision of what to achieve and how to achieve it.
The only pre-built AI tool in the Azure studio is Cognitive Services, which are a series of interfaces that allow the use of various AI tools when building apps and web platforms in Azure. Our clients are most likely to implement machine learning through Cognitive Search. Cognitive Search uses a combination of AI capabilities to trawl complex data sets to find what you’re after.
Like all machine learning, it gets better at giving precise results the more data it has because it operates on a mix of probabilities. That’s why it’s able to recognise misspelled search queries, for example (calculating the odds). And also why it is able to show unexpected relationships between data points that would have been difficult for a human to see (acting on the odds).
At the end of the day, however, it’s still up the human being to determine if the machine is correct. Is the relationship between the data points meaningful? Is that search on ‘Iffel Twoer’ meant to pull up data on the Eiffel Tower? Or was it meant to bring up summer haircut discounts: the wiffle twofer at the local barbershop? Only the user can tell, and then the machine will learn.
If you’re intrigued by Azure Machine Learning would like to know if it could help your business, let’s have a chat.