CIO

Intelligent machines part 1: Big data, machine learning and the future

Why AI is the next wave of disruptive technology

Futurist Ray Kurzweil predicted in 1990 that a computer would beat a human world champion chess player by 1998. In 1997, that actually happened with IBM’s Deep Blue. Since then, artificial intelligence (AI) has continued to advance rapidly, making now a good time to brush up on what is considered the next wave of highly disruptive technology.

AI consists of many sub disciplines such as natural language processing, computer vision, knowledge representation and reasoning. Machine learning executes AI in that algorithms – which are fed with big data – enable computers or machines to pick up on patterns, predict future outcomes and train themselves on how to best respond in certain situations.

The technology is making its way into a broad range of industries from marketing with behavioural targeting, to healthcare with accurate and early detection of complex diseases, to infrastructure with smarter urban planning.

In part 1 of this series, CIO looks at how some of the big players are using AI, including one of the most talked about facets of machine learning – deep learning or artificial neural networks that are made up of many hidden layers between input and output.

IBM, Baidu, Google, Facebook, Apple, Microsoft and others have invested big in AI. Local players that have also long been on the scene include DSTO, NICTA and CSIRO.

“The key to intelligence is learning,” says Alex Zelinsky, chief defence scientist at the Defence Science and Technology Organisation (DSTO). “Once we master machine learning, then you can start to have artificial intelligence. We are intelligent because we can learn; you can learn lessons from doing things and remember those lessons.”

“Very often with a computer program we are just writing a sequence of instructions for the computer to follow in order to accomplish a task,” adds Adam Coates, ‎director of Baidu Silicon Valley AI Lab.

“The idea behind machine learning is that there are some decisions … where it’s very hard to write down the instructions, so we would like the machine to learn to make those decisions based on looking at a bunch of examples. Deep learning is a technology that has really become popular in the last few years, which is a much more powerful version of machine learning,” he says.

Facebook’s director of AI research, Yann LeCun, says that deep learning is becoming pervasive in ways that people don’t yet really realise.

“Whenever you use voice recognition on your smartphone or whenever you upload a picture on Facebook and it recognises your friends, there’s deep learning in it.”

At Facebook, LeCun’s job is to find smarter ways to match content with users’ interests. Sounds simple, but it’s actually a difficult task because it involves training machines to read and understand all kinds of unstructured data such as text, images and videos to serve up relevant content at the right time and to a diverse bunch of users.

“Doing a really good job at this requires understanding content and understanding people. Ultimately that’s an AI problem because understanding people requires intelligent machines if you want to do a really good job.

“With current machine learning techniques it’s very difficult to have a machine read a text, for example, and then remember what the text is about, the events that happened in the text, and then answer questions.”

To achieve this, the machine needs a short term memory or its own sort of hippocampus that we humans have in our brains, LeCun says.

Read: 5 tools and techniques for text analytics

In April this year, Facebook revealed at its F8 developer conference that its memory network can read a short term version of Lord of the Rings in 15 sentences and have it answer questions on ‘where is Frodo?’ and ‘where is the ring now?’

“It’s basically a neural net with piece of memory on the side,” says LeCun.

“If you have a machine that holds dialogue with a person, that machine has to keep a trace of all the things that were talked about or what the topic of discussion is about, figure out what the person knows and doesn’t know, and then do something for her and him. You need to keep track of all of those things and so you need a working short term memory for that.”

Facebook’s memory network answers questions by figuring out where the spoken topic appears in the text and regurgitates it. One step above that is it can find relations between objects and know geometry.

“If I step out of the room and then turn right and you ask the question ‘where should I go to meet Rebecca?’ it has to remember where you are and where did I go to and do some geometry reasoning,” LeCun says.

Speech recognition is another big area where machine learning can be applied. At Baidu, the Chinese-based search giant, the aim is to have mobile phone software accurately transcribe words in languages such as English or Mandarin and understand the request.

Today, the technology is not at a point where it’s more convenient than typing on a small keyboard, Coates says.

“Something we think is a big failing to current speech systems is that they don’t work well in noisy environments. If your phone is sitting on a table a little bit away from you in a room that has poor reverberation or you try to talk to it in a crowded café, especially if it’s not a newer cell phone that has many microphones, it really doesn’t work quite as well.

“We are trying to make the system much, much more accurate so that when you speak to your phone you can do so casually like you and I are speaking together – the phone can understand what you’ve said and give you a really good transcription.

“And if we want a lot of these new and emerging applications like Internet of Things devices to work, we really feel that speech recognition systems have to handle noisy environments much better.”

With its Deep Speech system, which first came out in December 2014, Baidu trained it on more than 100,000 hours of voice recordings, first getting people to read to the machine and then adding synthesized noise over the top to make it sound like they are talking in a noisy room, cafeteria, or car.

“We feed all of those synthetic examples to the deep neural network so that it can learn on its own, so even when I hear this person speaking in different kinds of environments they are always saying the same thing and the neural network will learn how to filter out the noise by itself.”

The other part to this equation, Coates says, is getting the software to understand complex requests.

“For instance, if I ask you to book airline tickets and I give you a very complicated set of criteria and I tell you about this using natural language, it’s quite challenging to get a system to understand this in enough detail so it can go out and do what you want it to do.”

Abuse prevention is another area where machine learning comes in handy. Robin Anil – an ex Googler who left the company this year to work on statup Tock with other former Google staff – spent a lot of time at the search giant picking up on offensive edits users made to Maps.

“You’ve probably seen Map Maker that came with the news recently that some people drew something bad on maps between Android and iOS. Those kinds of problems I dealt with.”

Machine learning and ‘trust modelling’ was practical in helping verify which user edits were true and which were false, Anil says.

“The only way we can figure that out is through the power of big data; trying to figure out if a lot of people agree that is the truth and the system figures out that is the truth. So it tries to figure out agreements between people.”

Next page: How DSTO, NICTA and IBM are using AI

Page Break

At DSTO, Zelinsky says machine learning algorithms run unmanned aerial vehicles, unmanned underwater vehicles, unmanned surface vehicles, unmanned ground vehicles, and so on.

“One of the big applications of unmanned systems is what we call intelligence, surveillance and reconnaissance (ISR). They can be done by unmanned systems or satellites. They give you images of the environment or the planet and they have terabytes of data, so big data,” he says.

Instead of having people manually look through tonnes of images for presence of individuals, causalities, or of particular infrastructure, machine learning automates some of that to filter down what could be of interest to the analyst.

“It very quickly labels things to recognise this is water, this is sand, this is forest, these are buildings, etc. And it can filter out most of the uninteresting things, so you’ve got 8 terabytes of data or 100 terabytes of data and it is only letting you look at 4 or 5 per cent of that data where it looks interesting,” he says.

Machine learning also helps with mapping out the state of the environment and infrastructure post natural disaster, where it constructs a map in real time and compares it with a stored map to pin point where the damages have occurred, Zelinsky says.

Toby Walsh, AI researcher at National ICT Australia, is looking at how machine learning can apply to the organisation’s bionic eye project.

“There’s the hardware aspect – to physically connect to someone’s eye ball and can improve their eye sight when they get macular degeneration. But then, like most things, it turns into a software problem.

"You can take all the work that has been done on computer vision algorithms to try and actually improve the quality of the image you are trying to project on the back of the eye ball,” he says.

Computer vision helps with the bionic eye’s capability in spotting moving obstacles and then magnifying them so the user doesn’t run into them. Facial recognition can also be built into the bionic eye so that users are able to identify people they interact with.

“They implant electrodes on the back of the eye. The brain is so wonderfully plastic as well so it’s going adapt to the signal you put on it,” Walsh says.

Route optimisation is another machine learning project Walsh is working on at NICTA. “This is the classic travelling salesman problem. There’s approximation algorithms, local search methods that take a solution and improve it and look at ways of tweaking the solution and improve it again.”

A recent project NICTA helped work on that has turned into a startup is Foodbank Local, where it finds the most optimal route for food charities to pick up and deliver food.

The app factors in a number of variables to suggest the most efficient route, and gives users turn by turn directions of from start to finish in their journey of picking up and delivering food from supermarkets in their local area.

“If we can improve the efficiency which charities work, if we can feed more with less,” Walsh points out.

Walsh adds he is also looking at how game theory can be applied to optimisation, especially as these problems involve many stakeholders with sometimes competing interests.

“We are dealing with the fact that there is not just one person playing here; there are multiple players coming together and they may behave selfishly. So how do we design mechanisms that even if they are going to behave in selfish ways that out of that we get some optimal or good behaviours.”

One of the biggest advances in AI is IBM’s Watson supercomputer, made up of hundreds of machine learning algorithms. Jeff Welser, VP and lab director of IBM Research – Almaden, says he is trying to teach Watson when to ask questions back to the user before making its next best guess.

When going through the process of discovery, Watson sometimes comes back with three or four possible answers that are all about the same level of confidence. Having Watson figure out what information it might need to be able to differentiate its possible answers by asking the user questions is the next step, Welser says.

“It’s really more about how does the system understand what are reasonable answers for it to ask back. So Q&A’s the first step. And then there’s giving you more assistance on doing something back and forth.”

Read: Machine learning used to predict clinical response to anti-cancer drugs

Today Watson is mostly being trained to read through millions of documents for drug discovery, which means it has to understand the ins and outs of disciplines such as chemistry, biology, toxicology and compare different studies on these.

“Our drug discovery process today is very time consuming and costly. It takes hundreds of millions of dollars to make one drug. And our failure rate is over 90 per cent still, partly because a lot of today’s diseases are very non-trivial ... like cancer and multiple sclerosis, which are not very well understood,” says IBM researcher, Ying Chen.

“And the diseases themselves change. Once you make something, the disease adapts itself. So this process makes the discovery extremely difficult,” she adds.

When it comes to Watson coming back to the user with additional questions, it needs to be tuned to specific domains.

“We have some technology that is already doing reasoning and inference based on the questions that are being asked to Watson discovery advisor. But it’s a work in progress because what we’ve realised is when we apply to one particular domain there may be domain specific rules and knowledge that needs to be incorporated as Watson comes back with additional questions,” Chen says.

“We started building the system several years ago, but rapidly we realised you’ve got to give it intelligence to understand a specific area you are looking at. We need to be working hand in hand with domain experts, who really understand this field at a deep level. They are the ones who can explain what pattern is or isn’t interesting to them,” Welser adds.

Follow Rebecca Merrett on Twitter: @Rebecca_Merrett