Artificial Intelligence: how smart are you?

Artificial Intelligence (AI) is the intelligence of machines… Sounds like a self-referenced definition…. Let’s try again.

Artificial Intelligence makes the machines act as if they had the intelligence humans have… Really?... The same intelligence for any scenario? May they have abstract intelligence as well? Imagination?

Some time ago, when my son was 5 years old, we used to play with Siri during our car travels, asking questions and getting amused with the answers, a fantastic example of how natural language in machines provide them with almost anything you can expect from the conversation with a human, including humour… being the idea to get the longest conversation possible.

At the end of one of these conversations in the car, and after some minutes of silence, my son asked “Dad, does Siri have nightmares when it sleeps?”.

It reminded me of that book Philip K. Dick published in 1968 “Do Androids Dream of Electric Sheep?”, later retitled “Blade Runner” in the Ridley Scott movie in 1982.

The thing is that when you provide a machine with a voice, and implement a set of complex algorithms for a specific purpose, like for example, answering to spoken questions about topics like weather, favourite songs… even telling a joke when asked for, we are usually tempted to lose perspective on how the intelligence of a machine is different from the human intelligence.

IBM defines Artificial Intelligence (AI) as anything that makes machines act smarter. They propose that rather than Artificial Intelligence, you may think of AI as Augmented Intelligence, and that we shouldn’t intend for AI to replace human experts, rather, extend human capabilities and perform tasks that neither humans nor machines could do on their own.

That makes a lot of sense, but still we are not rid of the word “intelligence” or “smart” when trying to define AI, so let’s try to define it by defining Intelligence first. Any dictionary gives you not one but several ways for defining intelligence:

• The ability to learn, understand, and make judgments or have opinions that are based on reason. (Cambridge English Dictionary)
• The ability to acquire and apply knowledge and skills. (Oxford Dictionary)
• The ability to think, understand, and learn things quickly and well (Collins Dictionary)
• The ability to learn or understand or to deal with new or trying situations. The skilled use of reason (Merriam-Webster Dictionary)

So, we may settle that Intelligence is an ability, a skill, that provides the capability for learning from experience or observations, turn that learning into knowledge, and apply that knowledge to gain better judgement, understanding or decision-making about future similar events.

So it is true that machines don’t think, or dream, but we may provide them with the ability to learn, by incorporating a set of technologies for extracting knowledge from data, finding patterns in a set of this data and propose mathematical algorithms to explain it by its own, obtaining this way knowledge in a faster way than humans do, and suggesting cross-references of Big Data that may result in unforeseen results.

Internet has given us access to more information, faster. Distributed computing and the IoT (Internet of Things) have given rise to massive amounts of data (see my previous article about the subject in Industry 4.0: What Daedalus did), and social networks make much of that data unstructured. With Augmented Intelligence, humans can make informed decisions and extend their capabilities while letting the machines do the hard work.

In the visualization of the Key Performance Indicators of your company, machines help you deal with a lot of information coming from different sources. What you are seeing is inputs and outputs, even one next to the other in the screen.

You can see for example that a number of specific clients have generated a specific amount of sales. But what is the relation between the input and the output?. What transformed the existence of these clients into your cashflow this month? We give machines the ability to examine examples and create learning models based on desired inputs and outputs. And we do this in a variety of ways, namely:

• Supervised Learning: accomplishing a task by providing training, input and output patterns in the form of “labeled” data to the machine. The data has already been tagged with the correct answers by the data scientists. For example, if you provide your machine with a camera and you want it to identify whether a fruit is an apple or a banana, you can create a labeled data set (training data) containing the correct answer Apple (output) when the shape is round and the color is red (inputs), and the correct answer Banana (output) when the shape is curved cylinder and the color is yellow (inputs). The test data will be a new object which you will ask the machine to identify based on the knowledge to classify the fruit according to the input colours and shapes. There are plenty of examples, the simplest would be a cloud of dots in a graph which Microsoft Excel may analyse to provide you with a mathematical equation (polynomial, exponential or other) that accomplishes most of the outputs from the inputs, y=f(x), what is called a linear regression model. Having this mathematical equations, the machine may provide you new outputs from new inputs just by applying it. Another example would be to provide the machine with a data set consisting on the inputs 1 and 2 and the outputs 2 and 4. The machine may suggest that the algorithm is y=2x. However, the bigger the data set is, the more accurate the algorithm is. In this example, if the data set would have consisted of the inputs 1, 2, 3 and the outputs 2, 4 and 8, the machine would have suggested a different algorithm, which in fact is y=2^x. If you are interested to learn more about Supervised Learning you may find this link useful https://www.upgrad.com/blog/types-of-supervised-learning/
• Unsupervised Learning: self-learning technique in which the machine has to discover the features of the input population by its own, and no prior set of categories are used. In this case the machine is not given any output (it is unknown), so what can it do with only input values?. The machine can categorise these input values several ways, finding different set of patterns for the classification, what is called clustering. For example, given your clients, the machine may categorise them by age, gender, sold products per type, or amount of sales,... Some of these clusters may be useful or not for the purpose of a problem to solve. It is later when we can present a new client to the machine that it will give you what cluster is more appropriate for the characteristics of the client, then we may extract the corresponding conclussions. This is often used for Clients segmentation, or understanding different client groups around which to build marketing or other business strategies, but also Genetics, for example clustering DNA patterns to analyze evolutionary biology, Recommender systems, which involve grouping together users with similar viewing patterns in order to recommend similar content, Anomaly detection, including fraud detection or detecting defective mechanical parts (i.e., predictive maintenance).You can learn more about unsupervised learning in https://www.upgrad.com/blog/how-does-unsupervised-machine-learning-work/
• Reinforcement Learning: the goal is for the machine (called agent in this type of artificial intelligence) to learn a good strategy from experimental trials and relatively simple feedback received (rewards). With the optimal strategy, the machine is capable to actively adapt to the environment to maximize future rewards. How the environment reacts to certain actions is defined by a model which we may or may not know. What is relevant is how the agent will adapt (change) the algorithms depending on the states. Compared to the example in the supervised learning, in reinforcement learning the machine (agent) would assign a low probability of getting an output 8 from an input 3 (since the initial working model is y=2x for which the input 3 gives an output 6). However if the environment provided the agent with an output 8 for an input 3, it would get a low reward (an output with low probabilty of occurrence) which would reinforce the agent to reconsider the current model (algorithm) so that it gets couples of input/ouput that correspond to higher probabilities, thus changing it to y=2^x which does give an output 8, what the environment is returning for an input 3. Reinforcement learning solves a particular kind of problem where decision making is sequential, and the goal is long-term, such as game playing, robotics, resource management, or logistics. You will find and extended explanation about Reinforcement Learning in https://blog.floydhub.com/an-introduction-to-q-learning-reinforcement-learning/

https://starship-knowledge.com/supervised-vs-unsupervised-vs-reinforcement

While we may provide a machine with learning capabilities three different ways as was just described above, AI can also be categorised on application, and strength in different ways:

Weak or Narrow AI is AI that applies to a specific domain. For example, language translators, virtual assistants, autonomous vehicles, AI-based web searches, recommendation engines and intelligent spam filters. Applied AI can perform specific tasks, but not learn new ones, making decisions based on programmed algorithms and training data.

Strong AI or Generalised AI is AI that can interact and handle a wide variety of independent and unrelated tasks. It can learn new tasks to solve new problems, and does so by teaching itself new strategies. Generalised Intelligence is the combination of many AI strategies that learn from experience and can function at a human level of intelligence.

Super AI or Conscious AI is AI with human-level awareness, which would require it to be self-aware. Because we are not yet able to adequately define what consciousness is, it is unlikely that we will be able to create conscious AI in the near future.

Although the science fiction version of AI, like the one suggested by Blade Runner, may be a distant possibility, we are already seeing more and more AI involved, in the decisions we make every day.

Thus, if you followed me through this short trip, you may agree to answer the question in the title “AI: how smart are you?” as follows: we are very smart, but we are even smarter with Artificial Intelligence; and Artificial Intelligence is smart, but it is even smarter when, based on the knowledge it acquires, it is you who make fast and key decisions that are beneficial for your organization and goals.