How smart is artificial intelligence? (explained)

Introduction

The concept of artificial intelligence for many is still something abstract, while for others it is a wrong way of defining some standard functionalities: "it is not intelligence, but it is simply an #algorithm that is executed".

In reality, only the first wave of #AI could fall into this last definition, because already starting from the second one, that exploits Machine and Deep Learning that we are experiencing in recent years, it is something more. While the third wave, the one represented precisely by cognitive artificial intelligence, is something that goes further and that will lead computer systems to be able to add those ingredients typical of human beings, allowing a computer to make a "reasoning" that it will not be based only on statistical data.

It is clearly a difficult subject to deal with, not so much for the technical part (which I will not discuss here), but rather because it is something so far from what we are used to that just imagining and understanding that requires a considerable effort.

Currently, when we talk about AI we talk about Machine Learning, Deep Learning and Neural Networks. In short, a current artificial intelligence is very good at tasks such as the recognition of objects, humans, animals and images in general: a properly trained AI is able to understand what's inside an image that is proposed to it. It is also good at language recognition, hence translations, and anything else that involves analyzing data and correlating it into a recommender system.

In summary, given a visual input, data or data in other forms, the system must be able to give an output, which can provide information or can activate other functions and actions.

How the current artificial intelligence is created

To allow an AI to behave as just explained and perform certain actions, a system is no longer created with internal data configured so that an output is given based on the input. This was how early AI worked, and the reason why it doesn't work anymore is simple: you can't create a dataset that can identify all use cases. For example, if we wanted an AI to identify if a human being was present in a photo, we would have to provide data with all the photos of all present, past and future humans. Obviously it is not possible, so what is done is to create a system which, by accumulating experience, can tell if there is a human inside a photo based on the recognition of some recurrences.

Here we are talking about Machine Learning, Deep Learning and neural networks: an AI of this kind is trained by submitting various photos to it, asking if there is a human being in that image and indicating which answers are correct and which are not. Continuing this training, the AI slowly manages to refine the technique, it manages to collect more and more data, more statistics and recurrences, more details, which allow it to "understand" when there is a human being in a photo. Different AIs will have different datasets, but this is the approach for all of them.

What this AI can't do

Current AI, despite going beyond "give me a set of data, the operation I have to perform and I'll give you the result", lacks some elements typical of human beings. This means that, in one way or another, current AIs are statistical systems that have collected a very large amount of data, which correlate with each other and for which they make the "calculations necessary to understand, statistically, which answer to give or which operation make".

Current AIs base their intelligence on a pre-established set of data from which they "learn", find themselves in difficulty when they have to answer questions on topics where data is scarce, or none at all.

Try to feed an AI that has been trained to recognize human faces an image in which there is a bird: obviously it will not know what to tell you. To give a more concrete example, try to imagine an AI that has been trained to recognize any animal on earth, but at some point you show it a photo in which there is a species that has only been shot a very few times, or even a new species: it will hardly be able to give you an answer, or rather, he will probably tell you that in that photo there is the animal whose physiological characteristics are closest to those of the new species. If this thing happens to a human being, it is very probable that he would not say for example "in this photo there is a goldfish", but rather "it looks like a goldfish, but it probably isn't, it has traits in common and could belong to to the same family but be something else, maybe a new species”. In other words, what your mind does is reason, try to contextualize what it sees, imagine what it could be based on the information you have and based on experience.

Current AIs lack precisely this degree of abstraction, as they are unable to handle it: the only thing they do is a unique search in the data they possess, with the aim of weighing a statistical response.

Cognitive artificial intelligence

So we come to the next level of AI. How can a computer be allowed to perform this level of abstraction? The simple answer is “provide more data” and teach a machine how to use it, cross-reference and weigh it and get complex answers based on more differently structured data.

Future AI systems will have to "reason", i.e. cross-reference data, which include information on what it is, i.e. all the objective information that we have as it happens today, but also on why and when, i.e. contextualize, on "with whom" , that is, consider relationships between data and other models.

In practice, it is possible to see cognitive artificial intelligence as a mix of many models used today individually for individual AIs. Recognition of natural language, search the web, combined with computer vision, i.e. the ability to recognize the elements of an image, crossed with data analytics.

Let's take an example again, to understand the difference between current AI and what could happen with future AI. If today we apply an AI system within medical diagnosis, providing for example images of lung plates, or any other organ, the AI, trained with a lot of data, will be able to make a diagnosis, since it searches certain patterns that lead to a certain disease or case problem. The future AI will do all of this, but it will cross-reference this data with the historicity of patient data, with what is happening in the world (for example a pandemic with a virus that attacks certain organs), with the patient's habits (whether he smokes or not ), with the latest data and studies available on the web and in other sources on the consequences of smoking, with the information that the system can retrieve on the patient and his or her degrees of kinship. Once this is done, he will prescribe a cure, which will not be based solely on "if you have this disease, take this drug", but will also take into account information on how that drug works, on the possibility that it may or may not interact with other drugs that the patient may be already assuming and with real-time data that the AI can retrieve from the patient itself. In other words, it will reason as a doctor thinks, therefore not focusing solely on what it sees on the x-ray, but on everything that surrounds it and all the information it can have from multiple sources, reasoning on the weight and on the way in which it must consider each given.

How do you achieve this level of complexity?

AI works on data, but after all, the way we humans make decisions, the process we use, is also based on data: we have historical data, experience, contemporary data (i.e. what we see and perceive with the various senses at this moment) and the possibility of recovering other data if the ones we already possess are not sufficient to make a decision. We take all this data, in their forms, cross-reference them and make a decision.

Let's take another example to better understand: we are on the street at night and our brain tells us, from experience, that we see less at night, we can react less quickly, those who want to commit a crime take advantage of the night to not be seen, therefore the night hides more dangers than the day. In front of us there is a road with several very dark areas in which we cannot see well, but what we actually see is a sort of indistinguishable movement and we hear suspicious sounds; we are recovering this information, which when cross-referenced with our historical information tells us that perhaps we are in a dangerous situation. What do we do now? If the information we have is convincing enough, we probably turn around. If it's not, we look for other information: we move a little to better see what's in the shadows, we shout something like "who is there", we do some action that can trigger a reaction, we use causality, and thus recover other information.

What has just been described is potentially an approach that can be applied in future AI systems, and which is defined as 3LK, three level of knowledge.

The first level is known as "instant knowledge" and allows for a rapid response to an external event. It is what we previously defined as historical knowledge, i.e. a very complete set of data, with very defined rules, which represents a direct output to a given input. For example, the sound of barking instantly tells us that there is a dog.

The second level is defined as "standby knowledge", translatable if we want with "waiting knowledge". This is the set of information that we perceive in the present and that requires a series of processes and thoughts to be processed. They are the knowledge that we have by interacting with the world.

The third level is called "retrieved external knowledge", translated into something like "external knowledge or knowledge retrieved externally". The goal of all this analysis we have done is to understand how to react in the situation in which we find ourselves.

Let's take an example of these three levels. Imagine you are driving a car: you are moving from home to work, on a road that you have been traveling for years and you know well. Where to go, the maneuvers to make, are all instantaneous actions because you have ingrained knowledge about the environment and purpose, and this represents "instant knowledge". If, on the other hand, you change road because there is a detour, you have to pay more attention, you have to consider all the visual inputs, because the road is different and you have to be able to respond to the pitfalls of the road you are on, and this is the " standby knowledge”. While you have made this detour, you have gone too far, and you no longer know which way to go, and here you have to take out your smartphone and put on the satellite navigator to find your way again: this is the "external knowledge".

By combining all these types of information it is possible for us humans and for an AI to respond to new unexpected situations, from those in which we have a lot of data, to those we have never been in before, breaking the limitation of current AIs that can only rely on a pre-established dataset on which they have been trained.

The technical approach

The problem with neural networks underlying current AI is that based on datasets, to increase intelligence it is necessary to increase the dataset and the processing power of the data itself, because there is always more information to process. This clearly leads to ever-increasing hardware requirements, increasing costs, increasing the energy consumption and losing efficiency. Obviously a new hardware approach is needed.

When will the new AI arrive?

Cognitive AI is actually already among us, albeit in the early stages of evolution. Most probably by 2025 current AI systems will have or will be making the leap to this more advanced AI.

Where the new AI will be applied

Simply put, everywhere. After all, we are talking about allowing cognitive AIs to recover and cross-reference data of different kinds, in such a way as not to give an output weighed on the basis of the input received and the historicity of inputs of the same type, but to adapt the answer. Man-machine interaction will become more "personal", more accurate, more natural, because machines with this AI will begin to "reason" not only based on input data and on what they have been taught, but also on contemporary data and personal, which AI can fetch across multiple domains and sources.

Let's take the image of a guitarist as an example. A current AI “looks” at the photo and says “there is a guitar” or “there is a guitar and a man”. A future AI will say, for example, "There's a man in a hat playing guitar, the man is smiling, so he enjoys playing guitar, and he's wearing a cowboy hat, he's probably playing country music." .

Cognitive AI recognizes objects, basic, contemporary and contextual information (smiles, so it is in a state of mind that is probably related to what it is doing), as well as external information (usually, the guitar association and cowboy hat is found in those who play country music). In practice, AI makes the same reasoning that any of us would.