(written by Lizzie Cope, revised by Stefan Hesse)
After being a family member at TLR in Malaga for roughly 6 months, I had the opportunity to present at our monthly TGIF event at the beginning of the year. My topic of choice: Artificial Intelligence (AI)! My objective for this talk was to break down the complex topic of AI for people without any technical or mathematical background.
What is AI and what can it do?
According to Merriam Webster, AI is “the capability of a machine to imitate intelligent human behaviour”.
In other words, this means that AI systems typically demonstrate behaviours associated with human intelligence, such as planning, learning, reasoning, problem-solving and, to a lesser extent, social intelligence and creativity. For example: Amazon’s Alexa or Apple’s Siri on your phone uses AI to recognise what you say in order to provide you with a virtual assistant.
In general, AI can be broken down into two sub-categories: narrow AI and general AI. First, we’ll take a quick look at narrow AI, which is what we see all around us nowadays. Intelligent systems that have been “taught” how to do one specific thing without being specifically programmed how to do it (for example in speech and language recognition devices, or self-driving cars). These systems have been trained to e.g. recognise cars in an image by showing it lots of examples of cars in images.
Then there’s general AI, which is the type of adaptable intellect found in humans. It’s a flexible form of intelligence capable of learning how to carry out different tasks, ranging from cutting hair or to reason about a wide variety of topics based on accumulated experience (e.g. the things you see in the movies such as Terminator or i Robot featuring Will Smith). This sort of AI doesn’t actually exist today and experts are divided over how soon it will become a reality or if it is achievable at all. So let’s not panic about robot world domination, just yet at least!
Key AI terminology
1. Machine Learning: A computer system is fed large amounts of data, which it then uses to learn how to carry out a specific task, such as understanding speech or captioning a photograph.
2. Data Science: The science of extracting knowledge and insights from data (recording, storing and analysing data).
3. Deep Learning (Neural Networks): One specific method within machine learning and mostly responsible for the great breakthroughs we have seen in recent years in AI. This method is vaguely based on the human brain.
But what can AI NOT do?
As impressive as AI is becoming, almost all of the progress that has been made within AI is made in the narrow AI category. The methods used in this area have certain requirements, which poses a problem in order to make further progress in the area. Below is an excerpt of typical problems a machine learning engineer faces:
Problem #1: Data Labelling. Most current AI models are trained through so-called “supervised learning”, which requires a lot of examples categorised by a human. Therefore in order to achieve any meaningful results a large amount of human labor is necessary to annotate data.
Problem #2: Obtaining large data sets. Directly connected to the problem above is the size of data that for example, deep learning requires. Depending on the task no dataset may exist or might be too expensive to create. This point also outlines the biggest difference between human intelligence and artificial intelligence, since a human generally can abstract the concept of e.g. a chair by seeing one picture. Our current AI technology can not.
Problem #3: Explainability. The problem of explainability has been a long-standing issue within AI and got, even more, pressing through EU-regulations demanding that certain decisions made by algorithms have to be explained to the person concerned by it.
The more complex the algorithm gets, the harder it is to explain in a profound way of why the AI derived at this decision. This is a serious problem because without knowing what is going on, it is also hard to improve the algorithms.
Problem #4: Transfer Learning. For humans, it is very easy to transfer their knowledge from one domain to another. We humans can make abstractions and find similar patterns within different domains, which is not the case for artificial intelligence. There is some success within image recognition, but transferring knowledge across domains has not been shown to work.
Problem #5: Human Bias. Computers are great because they make emotionless, rational decisions assuming we program them that way. In AI we do not program them, but let them learn by our experience which includes all the biases that humans have and we are transferring them onto the AI.
How can you work with AI?
In order to use AI nowadays, you just have to take your phone out of your pocket, but this is just the consumer end of things. If you want to make an AI work on your own problem then you can do that too. Though the options differ if you are a programmer or not.
The most important part, if you are a programmer or not, is to define your objective, identify what data you have, and then try to iterate on the problem. Do not make the mistake to simply gather a lot of data. A lot of data might be required to solve your problem, but it also has to be the right data.
If you are a programmer you could start by doing one of the thousands of online courses about AI. These courses do not necessarily require a strong mathematical background. Even though I would always advocate picking a class that puts a strong emphasis on math since AI without math brings you only so far.
For none-programmers, there are also options. Going strong on deep neural networks is probably not the route to go without any programming knowledge, but there are tools like “WEKA”, which provide a graphical user interface, where almost anybody could throw some data in and see some results. For more sophisticated projects it would be necessary to consult with an expert in the field. I would personally recommend coursera as a platform, this isn’t for the faint-hearted however as this course is specifically about deep-learning!
To wrap up, I’d love to share the slides with you, which I used during my presentation at TLR:AI-4-Everyone
Stefan Hesse, a TLR family member since March 2019, studied computer science with a specialisation in data science. He is passionate about artificial intelligence and studied various machine learning algorithms for more than five years and applies them on a daily basis. Check out the company I work for logarithmo to find out further information about what we do!