Guest blogger Dr Adam Makarucha tells us about real-world use cases of artificial intelligence
It has been the subject of sci-fi novels since Mary Shelley created the genre when she penned Frankenstein, but Artificial Intelligence (AI) is no longer stuck in the realms of fiction. Fortunately, at IBM we are not creating monsters – but in the IBM Systems team, we do teach machines to make independent decisions based on a variety of inputs. And this technology has begun a wave of business gains in all sorts of industries.
There is a common perception that machine learning is solely the domain of big business and universities. While they are certainly investing in this technology, they aren’t the only ones. I’ve seen small engineering firms, supermarkets, even farmers making some big savings. Every business leader should be looking at the implications of machine learning in their own organisation.
How do Machines Learn?
Artificial intelligence (AI) is taking off thanks largely to deep learning, a type of machine learning. The way we teach machines using machine learning is in many ways similar to how we teach children. We show them something, or rather many examples of that thing, such as pictures of animals, let’s say a giraffe and an elephant. We ask them to make guesses, “is this picture an elephant or a giraffe”, all the while correcting them if they guess wrong and reinforcing them when they guess right. This way, the child learns all the differences, or what we call features, that can help them identify a giraffe vs an elephant. Well machine learning is effectively the same process, just with potentially millions of examples, and at the end of this “teaching” (or training) process the machine learns to guess “giraffe” or “elephant” according to the features within the images. This is a highly iterative process and we do it over and over until the machine has learnt everything from the images and it starts to get the prediction correct most of the time. The beauty of deep learning is that the more data we provide the more the machine can learn and, the greater the accuracy it can achieve.
Rather than distinguishing between animals, though, we are teaching machines to make predictions such as whether a credit card holder will default in 3 months, what products will be needed where and when, even how to review medical scans and diagnose different diseases. Reaching those outcomes can be costly, but that doesn’t have to be the case.
Deep Learning: The Right Tool
For a data scientist, these types of use cases involve a lot of time in data preparation, including identifying features that can help the machine to classify the information it receives. Typically, the data scientist can work on one project for many months, preparing many “features” for each unique dataset and problem. However, the promise of deep learning is that these features are learnt by the machine, saving the data scientists time and allowing them to work on more projects simultaneously. The downfall is that deep learning requires more data and time to train than other methods, but we can use infrastructure and tools to accelerate and scale these processes.
To say IBM’s deep learning platform accelerates training times would be an understatement; imagine Usain Bolt in your local primary school sports day, and you’ll be close. The outcome of this is that multiple projects can be run, and results are faster – and even better, it actually scales. The more infrastructure you get, the more thesdata scientists can accomplish. The velocity comes from the tight integration of Nvidia GPUs with IBMs POWER9 CPU, which can get drop training times by 120x1. How does that translate into results? I’ve seen training times taken from four weeks to six hours, this allows data scientists to train, iterate and deploy deep learning-based models in weeks rather than months.
How is Deep Learning Used?
The possibilities are almost endless. Deep learning research at Stanford University has been able to outperform radiologists in identifying pneumonia2. While there is a high success rate in treating pneumonia in wealthy countries, 4 million people a year are killed worldwide. That’s because in the third world, there is often little access to expertise, so it goes undiagnosed until it is too late. The work of the Stanford team democratises expertise – doctors could upload x-rays and get fast, accurate results that save lives.
While it may be less dramatic, a grocer in Ecuador reached out to the data science community to work out what items would sell and when. The results reduced prediction error by 10%. If that process were used in Australia, it could result in annual savings of $15 million – on bananas alone!
From checking massive amounts of complex financial documents to checking cell towers using drones, or making inroads into the $900 million of wheat lost per year to crop disease, deep learning can tackle an incredible amount of problems. My friends at Computer Merchants are using AI to analyse incoming customer emails and track customer satisfaction, so you don’t have to be huge, and your gains don’t have to be measured in millions. That’s what makes this technology so exciting – and gives it the potential to change markets.
AI depends on data to get results – but it doesn’t have to be your own data. Publicly available data has given some customers an excellent start. That said, your own data can give you results that nobody else has, and that’s where competitive advantage is made. If you’re not ready to venture into machine learning just yet, now is an excellent time to get your data in order – and that means chatting to an expert about how you store and manage it. Working on your data now will pay off later.
For more information about deep learning and how it might fit your organisation, contact the team at Computer Merchants, or the IBM Systems team.