‘Artificial Intelligence’ Services? – What is AI really?
When you think of Artificial Intelligence (AI), you might think of sentient robots (iRobot) and all-knowing computers (Hitchhiker’s Guide to the Galaxy, Skynet). Although these Hollywood depictions of AI are cool, this is not a totally accurate representation of AI in reality. Let’s discuss!
I'm so confused, what is AI?
Artificial general intelligence (AGI), sometimes known as StrongAI, is the ability of a machine to possess the same cognitive and intellectual capacities as humans. In other words, the actions of a true AGI should be indistinguishable to those of a human and is the representation of AI that we are used to seeing in popular culture. This is the premise of the Turing test you may have heard of – which was designed to test whether or not a computer is capable of thinking like a human being.
However, in reality, when we think of AI, for example self-driving cars or applications that create memes from text descriptions this is considered narrow AI, not AGI. Narrow AI is goal-oriented and designed to accomplish specific tasks – such as facial recognition, language translation, disease diagnostics or playing a game of chess. While narrow AI has the ability to perform far better in both accuracy and speed than a human can on a particular task, it is a long way from AGI and will often struggle outside of the domain on which it was trained. When we encounter AI applications in our daily lives these are narrow AI applications. It is what powers Alexa, Siri, your Netflix recommendations and social network filters and is what we typically mean when we refer to AI.
How does AI work?
In order to properly explain how AI works, we need to take a step back to the fundamentals that underpin it. It’s time to introduce everyone’s favourite scientific fields – computer science and statistics. More specifically, introduce a sub-field of the two called machine learning.
Machine learning marks a shift in approach to programming. For a long time, a deterministic approach where predefined rules are used to find patterns within data and solve problems has dominated the field. Nowadays, machine learning enables a probabilistic approach where a variety of statistical models are leveraged to train an application. This application learns from historical data to improve continuously (automatically fed by new data) without being explicitly (re)programmed along the way.
Before building an AI application, you would traditionally train a machine learning model to perform a particular task using an appropriate statistical model and historical dataset. This task could be anything from extracting fields from invoices to converting audio to text. Once trained, these tasks are often referred to as AI tasks as they are designed to replicate human cognitive functions such as speech or language understanding.
Think about a child learning to read; you begin with the letters and show them several instances of the letter ‘a’ in various situations to assist them with understanding what the letter ‘a’ looks like. You include words that begin with or contain the letter so that they can learn how that letter sounds. Once they can recognise letters and individual words, you might begin giving them phrases or providing further context to help them comprehend the meaning behind the words. Ultimately they will be able to read and comprehend language without having learned every word in the dictionary.
Training a machine learning model follows a similar process, although hopefully much faster. By training an algorithm on thousands of data points, it can begin to learn inherent patterns within the data which can be used to perform specific tasks such as recognising animals in an image. In the above example, once an animal detection model has been trained it can also adapt to previously unseen but relevant images. However, keep in mind that today’s AI is powerful but focused. For example, if you wanted instead to build an application that can analyse medical imagery, you would need to teach a new algorithm using appropriate medical data points, instead of using the animal detection model above.
Finally, once we have an accurate machine learning model that can replicate an AI task, we can combine this with other machine learning models in addition to data and software engineering to create an AI application.
What makes AI so powerful?
Today, AI is everywhere and is transforming the world. It is being integrated with and used in a range of industries. AI has been proven to be incredibly successful at identifying fraudulent activities in banking, diagnosing illnesses in healthcare, offering personalised suggestions in retail, detecting defective items in manufacturing, and various other applications.
So what makes AI so powerful, and why is it becoming increasingly common in everyday applications? There have been three major breakthroughs that have accelerated the adoption of AI since its conception back in the 1950s (fun fact: the first electric programmable computer was invented in 1943). In short, the three breakthroughs are big data, cheap computation, and better algorithms.
Firstly, as we spoke about in the above, AI applications often require large volumes of training data to accurately perform. In the mid to late 2000s’ there was a growing trend within organisations to generate and store more and more data that could potentially be valuable to the business, ranging from customer interactions to network or systems data. As the volume and quality of data generated and stored by organisations increased over the past two decades, so too did the volume and quality of AI applications trained on these bigger and better datasets.
Secondly, the algorithms used in machine learning have dramatically improved in recent years, thanks to ongoing research efforts from universities, research organisations and industry. Machine learning algorithms range from shallow statistical methods such as linear regressions, random forests or decision trees through to deep learning methods such as convolutional neural networks (CNNs), long short term memory networks (LSTMs) or generative adversarial networks (GANs). When you hear the phrase “deep learning” it is referring to these larger, more complex algorithms. These advances in the underlying algorithms (and their implementation) used within AI applications enable more accurate models to be trained without necessarily increasing the volume of training data or computational power required.
What makes AI so powerful?
AI is undeniably becoming a more powerful force in the business world. It is making our existing technology smarter and unlocking the potential value of all the data that enterprises gather. You may start questioning “How exactly can a business access AI?”. Or you may start getting excited thinking “let’s hire the smartest Machine Learning engineers (clearly we can’t go wrong, there is ‘Machine learning’ in the name right?), and build a smart model so that it can solve a business problem in no time. Unfortunately, that’s not the answer. Sit tight and read on!
“You can’t expect to get anything useful by asking wizards to sprinkle machine learning magic on your business without some effort from you first.”
Cassie Kozyrkov – Google Chief Decision Scientist
‘Artificial Intelligence’ should be and is already accessible to everyone and every business. The term Artificial Intelligence has been kicked around a lot over the years and is often abused and misused. Applied Intelligence is a relatively new concept that suggests we should focus on bringing AI out of the lab and into the real world, rather than focussing on how to construct a super AGI system.
When looking to implement AI within an organisation, it is important to remember that a machine learning model is only as good as the data it is trained on, an accurate machine learning model only creates value if it makes it into production and deploying a machine learning solution to production is only worthwhile if it provides tangible business value.
While the barrier for entry into AI is continuing to lower due to advancement in no/low code approaches, anyone who has been involved in a machine learning project knows that having a successful AI project involves a whole lot more than just teaching a machine to learn.
Finding appropriate AI use cases within a team or organisation is hard. Getting high quality data for machine learning models to be trained on is hard. Deploying and maintaining accurate machine learning models in production is even harder. While machine learning is clearly a critical component of an AI project, without an appropriate focus on the entire solution, a project is unlikely to get off the ground, or stay off it.
That is why at Eliiza we consider the ‘A’ in AI to be not just ‘Artificial,’ but also ‘Applied,’ in Applied Intelligence. Our focus stretches far beyond just training the latest and greatest machine learning models and onto the wider end-to-end lifecycle of an AI solution, from insights, ideation to production, with an overarching ethical framework and governance. We love AI as much as the next engineer but what we love more is helping our clients plan, build and maintain value creating AI solutions, no matter what stage of their AI journey they are in.
At Eliiza, we understand the end-to-end lifecycle of an AI solution is broad and finding a jack of all trades that can do it all is not feasible. We know that drafting AI strategies require different skills then developing data pipelines which in turn require different skills then training machine learning models. With this challenge in mind we have created five different practice areas across Data Analytics & Strategy, Data Engineering, Data Science, Machine Learning Engineering and Machine Learning Products. These specialty practice areas are designed to provide highly targeted advice and expertise to clients who are stuck or need support with a particular step in their AI journey, or can be combined together to provide end-to-end AI solutions.
So no matter where you are in your AI journey, Eliiza has you covered and can provide you with the tools, people and expertise to accelerate your machine learning journey from ideation to production and beyond!
Detecting PII with McGrathNicol
The current process to identify PII or other sensitive data within large volumes of unstructured data can be a time consuming and human resource-intensive process if “traditional” methods are used.
Automating claims processing at nib
We created a image processing pipeline to find and extract, analyse the extracted text and identify information required for the claims process