Artificial Intelligence: 3 Things It Can Do, And 3 Things It CannotYou have probably heard people mention during meetings or kick-offs: “Can we not use an AI solution for that? Automate it with machine learning?”
Artificial intelligence is all the rage these days and cited as “the technology that is going to bring us the fourth industrial revolution” by the World Economic Forum. It is clear that it is vital for companies to stay on top of this.
But it is easy to get carried away and believe artificial intelligence can do everything. Despite rapid developments and many possibilities, there are still certain things it cannot do yet.
To get a sense of where opportunities for using artificial intelligence in your company lie, here is an overview of a few things artificial intelligence, and more specifically its subset machine learning, can currently do, and a couple of things it cannot.
What artificial intelligence can do1. Detect and categorise objects & peopleThink of:
Categorise email as spam or non-spamDetect whether a screen is broken or notDetect whether there is a car in an image and its positionJudge the probability of a tumour to beRecognise faces and voices from images and voice recordingsAI systems can detect and categorise objects, by learning to produce a certain type of output based on the input it gets. Examples of input are images or sentences. Examples of output include categories, identifying and locating objects, and calculating probabilities.
A good rule of thumb to keep in mind here: Any judgment that you can do in less than a second can probably be done by an AI system as well. But at some activities, like calculating the probability that a tumour is malignant, an AI system might actually perform better.
This is often done via a technique called supervised learning, where the AI system is being taught to produce the right output according to the input it receives, with a training set of examples.
To train an AI system well it is important to have enough data so the system can achieve a high enough level of accuracy. You do not need enormous data sets like Google and Facebook have: a training set of a 1000 examples is seen as the minimum. But take into account that more is needed for complex problems.
Note: Many problems in AI these days are solved via a combination of supervised and unsupervised learning, the technique mentioned next.
2. Find patterns in unstructured data to recognise clusters, trends and objectsThink of:Find new clusters of symptoms that could indicate a certain diseasePredict future housing prices based on past housing prices and market factorsPerform market segmentation based on customer behaviourAI systems can also learn to make sense of unstructured and unlabeled data where you do not know the desired outcome yet. Here the system starts clustering by itself based on patterns in the data it detects.
The resulting clusters can help discover previously undetected patterns and new trends. Examples include new market segments based on customer behaviour and new clusters of symptoms that can indicate the presence of an illness. This is achieved via a technique called unsupervised learning.
A major challenge with applying unsupervised learning techniques is the difficulty of testing the accuracy of the system. As the expected outcome is unknown, you cannot compare the results you get with a sample set of data. Thus it can be hard to test if your AI system came up with an accurate outcome.
That is why, first of all, it’s important to consider how big the chance is there will be natural trends in your data that can be of use to your business. Also think of how real-world scenarios you can use to test the accuracy of the outcomes.
In the case of detecting market segments, you could use A/B tests and compare the results you get from using your usual market segments with the one you get from your AI system. This can give you an indication of the accuracy and of how much value this technique can be in this particular case.
Because of the difficulty of testing their accuracy, unsupervised learning solutions currently show the most potential in unearthing possible new trends in your data, which can then be explored and tested further. But we should be aware of the risk of taking the outcomes at face value.
3. Quickly learn complex behaviours and come up with creative solutions, given a clear set of rulesThink of:Have prosthetic limbs teach themselves how to moveTeach self-driving cars to adapt themselves to driving in different situations (different rules, environments)Given a clear set of rules, an AI system can learn complex behaviours rapidly and find creative ways to achieve an outcome. It can, for example, achieve an extraordinary level of chess-playing skills quicker than any human could. But it can also learn complex behaviours such as how to move like a human limb (for bionic limbs), and fly autonomously (for helicopters).
In the process, AI systems can come up with very creative solutions to achieve the desired outcome. A famous example is AlphaGo, the AI system developed by DeepMind that in 2016 defeated the world champion Go player. Go was for a long time considered too complex for a machine to master, but due to using ingenious strategies, AlphaGo managed to do so.
AlphaGo was developed using a technique called reinforcement learning. Here the AI system is given a clear set of rules and the outcome to strive for and starts running many simulations (in parallel) to test all the different possibilities. In the process, it is being rewarded or punished dependent upon the outcomes of these different simulations.
But despite impressive results there are drawbacks to this method. The system needs clear rules to run its simulations. That is why the technique has been so successful in teaching AI systems to play games. However, this can prove tricky in real-world situations where rules tend to be much less clear-cut.
Moreover, you have to be clear on the desired outcome and what constitutes a good result. Else you run the risk of your AI system coming up with a result that might get the job done, but not in the way you like. Such as a prosthetic leg that ends up jumping in “kangaroo style” rather than walking in human fashion, or a helicopter flying upside down.
What artificial intelligence can currently not do1. Perform under high uncertainty with unclear rules and desired outcomesThink of:Make a final diagnosis for a patientMake ethical judgments on whether someone deserves a penalty or notAI systems are becoming increasingly sophisticated and creative, but even techniques like reinforcement learning need clear rules to work well. In situations where these are ever-changing, or a definite “right” outcome is lacking, AI systems struggle.
An example is when a doctor needs to deliver a final diagnosis to a patient. She has to take many factors into account without necessarily one clear answer, but rather a set of possibilities.
How AI systems can help here is by providing information on the likelihood of a patient having a certain disease, or indicating possible risks for further examination.
AI systems also cannot (yet) create innovative solutions in complex fields lacking clear rules, and tackle for example large-scale societal challenges. But it can help to sift through large and complex sets of data to come up with ideas that can then be tested further, as the AI Economist being developed by Salesforce is doing.
2. Produce empathic and compelling outputThink of:Write empathic customer service emailsCreate compelling research reports and presentationsIn both of these cases, what makes a good result depends upon many contextual factors: who you are talking to, at what time, what has happened before. And the result can be achieved in many different ways: using a combination of words and sentences placed in a certain order, which all influence the meaning of the whole.
As such it is incredibly hard for an AI system to create a result that would sound natural to us, even with huge amounts of data.
AI systems composing music are a good example of this. Producing music in the style of Bach has been possible for many years already, but creating compelling pop music has proven to be a much bigger challenge. Rather than having clear compositional rules, it relies on the expectations of the listener, and how it meets or defies these. OpenAI has lately given it an impressive shot with its Jukebox, and you can take a listen for yourself. But it sounds like it still has a way to go…
3. Produce results free from human biasThink of:Selecting promising job candidatesEstimating the probability of a repeat offenderSince we often have to train an AI system, our biases can easily creep into the results they produce without us noticing. This has produced especially dire results in the fields of criminal law and human resources.
AI systems used by the US police force to predict criminal behaviour have for example shown racial bias, as uncovered by AI Now, an institute specialised in understanding the social implications of AI. This is due to the data that the system is based on being racially biased as well.
And in human resources likewise, there have been cases of AI systems favouring men over women despite the same qualifications, especially in male-dominated fields. This happens because data on previously successful hires is being used to train these systems, but these include mostly resumes from men.
Avoiding biased outcomes is one of AI’s biggest challenges, and increasingly critical as the number of fields AI is used in diversifies. It is important to keep this in mind when applying AI in a field in which bias is common and could have a large impact. Take a critical look at the data you are using to train your AI system, thoroughly test the outcomes, and always use these in combination with the judgment of a professional.
Things are ever-changing in the field of AIWhile this is where AI is currently at, the excitement around it and an active open source community cause developments to happen at breakneck speed. It is, therefore, crucial to stay up to date if you want to be part of the AI revolution.
Therefore make sure to educate and keep updating yourself on developments in the field by following high-quality resources. I can recommend the following sources for a deeper dive into the topic, and staying up to date: