Learning about artificial intelligence
When you are on your smartphone browsing Facebook or reading the news, do you ever stop to think about the technology you hold in your hand? A typical smartphone is a million times more powerful than the computers NASA used for the moon missions in the 1960’s, it’s connected to the world wide web that contains at least 4 billion pages of information, has a virtual assistant to help you with general tasks, and when you aren’t busy taking selfies, you can use it as a phone! Computer technology has advanced incredibly fast in the last 60 years, and with it something called “artificial intelligence”.
Artificial intelligence is defined as the ability to perform tasks normally requiring human intelligence. We used to think this meant doing calculus or playing chess, but computer technology has come so far since the first computers that what is considered artificial intelligence has also taken steps forward. What you’re thinking of when we use the word AI today are most likely the sub-genres of AI called “Machine Learning” and “Deep Learning”. What makes these unique is the machines ability to learn from experience.
Natural reasoning and learning from experience has always been a huge part of AI, but only recently has it become a reality. This is mainly due to three factors:
1: Available data
Thanks to the the world wide web, and the Internet of Things (how everything is connected), we now have access to an incredible amount of data that can be used for learning purposes.
2: Increased processing power
Having loads of data doesn’t help if it takes a computer months to process it all. Processing power has become extremely powerful in a very short amount of time thanks to Moore’s law.
3: Lower costs
Due to the rapid advancement of technology, access to the data and processing power is relatively cheap. You don’t need NASA supercomputers to develop and run AI programs.
Differences between Machine- and Deep Learning
Even though both these AI technologies aim to build a program that can learn and improve as it is used , machine learning is a far more basic than deep learning. Let’s say you have a thousand pictures, and you want a program that can recognize all images containing a car. A machine learning developer would create some classifications that recognize the most basic parts of a car (wheels, shape, doors, etc.) and run the pictures through the program. The first round of results will probably not be very accurate, so you help it by selecting a few images that were not recognized (false negatives), and some images that were mistakenly recognized to contain cars (false positives). The program will add to the classification list, and the accuracy will improve.
If you wanted to solve the same issue using deep learning, it would have been a whole different story. First you’d probably need a lot more than a thousand pictures. Deep learning programs are heavily reliant on big data, which is why only the largest tech companies have incorporated it into their solutions (Google, Apple, Amazon, etc.). Second, the developer would begin meticulously adding layers of detailed classifications, which is impossible to explain in a simple paragraph, but if you’re really interested you should check out this video of Martin Görner from Google that explains it really well. You then run the program on some very powerful computers, and if the problem is solvable, the data good enough, and the developer skilled enough, you’ll end up with something that is more accurate than a machine learning program. How much more accurate? Where machine learning can reach an accuracy of around 90%, deep learning are almost 100% accurate. 9–10% doesn’t sound like much, but if you can’t afford mistakes, it’s a mountain. Moreover, since a deep learning solution uses so many classifications to accurately identify the car, it could eventually have enough data to distinguish different car models!
Do I need AI?
A lot of business owners have been asking that question lately, and the answer is usually “Probably not right now”.
As you can see from the Gartner hype cycle graph, machine learning and deep learning are still relatively immature technologies with a lot of hype and expectation built around it. Those implementing it today would be defined as early adopters. As these early adopters learn more about the possibilities and restrictions of these technologies, they will build competence and knowledge, paving the way for future adopters. It is expected that in two to five years machine and deep learning technologies will reach what is known as the “plateau of productivity”, where we know enough about it to productively start using it, and a mass adoption will occur.
BDO Norway’s approach
Working for BDO’s innovation team, I’ve been part of a project group to map how BDO can use AI. Like everything else, the first step is always knowledge. We did a lot of research on the subject matter, and shared the information we found. In the beginning it often felt quite confusing as many of the solutions within AI are intertwined, and we spent time discussing what was and wasn’t AI. To address this challenge, we decided to define AI as solutions that use machine or deep learning. This decision made the project split into three: AI (Machine/Deep learning), Robotic Process Automation, and chatbots.
Next step was to engage colleagues from different service areas within BDO. We used a design thinking approach and held several workshops to map our needs, challenges, and to come up with solutions. During these workshops we always gave participants a short introduction to AI, but asked them to imagine their ideal solutions without taking into consideration the limitations of technology. There are two reasons behind this:
- We can use these ideas to figure out what problems they are solving, and possibly come up with alternative solutions, which might not need AI.
- Experts within the field have a better understanding of what is and isn’t possible. By asking participants to ignore the limites, we don’t risk missing out on potential great ideas.
Once this is done, it‘s back to traditional product and business development strategy. We look at all the solutions and try to figure out what the benefits are, what the cost will be, the difficulty of implementing it, etc. We prioritize based on the analysis, and finally provide a report of our results to upper management.
To sum it all up
AI can be really confusing, but remember that machine and deep learning are the AI technologies that “learn”. Companies that implement machine or deep learning solutions today are early adopters of the technology. It is definitely a disruptive technology, as it has the ability to save employees from an incredible amount of repetitive and time consuming work. However, providers and experts within the field are still very limited, and it is expected that it will take another 2–5 years before it is mature enough for a mass adoption to occur. If you’re a business owner it might be wise to start mapping the potential areas where you can implement machine learning. As an average user, you are already surrounded by learning technologies thanks to companies like Google, Apple, Microsoft and Amazon. There’s a reason why Google translate is better today than it was ten years ago, and it’s the same reason why Amazon knows what you are going to buy before you do. Machines are beginning to learn, and if you’ve watched terminator, this is the exact point in history where Skynet all began.
