“For the Uninitiated” part 2, i.e. the level of intelligence of Artificial Intelligence.
Artificial intelligence is a field that has grown into numerous myths. It inspires scientists as much as it does pop culture creators. In today’s episode “For the Uninitiated” about the essence of Artificial Intelligence (AI) and the current level of machine perception will speak Grzegorz Gwardys – Promity expert, dealing with machine learning on a daily basis.
The (in)famous artificial intelligence … The theme, inspiring science fiction creators for years, still arousing emotions. For some, extreme and dark, related to the vision of intelligent machines that take control of people … What is Artificial Intelligence actually and did we begin to abuse this term a little?
Grzegorz Gwardys: As soon as I hear about those intelligent machines, I am immediately reminded of the movie “Terminator”. The likeness of the title hero is used by every second article on Artificial Intelligence … Just like those photos of smoking chimneys, which always “decorate” texts devoted to emissions from coal power plants. However, there is a difference: while only water vapor can come out of such chimneys, current AI solutions are far from bloodthirsty machines from Hollywood super productions.
Unfortunately, the very term AI raises a whole lot of problems and pitfalls. While defining Artificial Intelligence as a certain set of various sciences and fields of knowledge (Wiki lists here a whole list: fuzzy logic, evolutionary programming, neural networks, artificial life, robotics) is correct at a certain level of generality, it does not allow understanding the essence of the issue. And this comes down to determining where Artificial Intelligence begins.
Currently, all devices based on AI shine, they allow you to take better pictures, recognize speech or personalize them. Here it is worth to remember the concept of Artificial Intelligence Effect, formulated by Pamela McCorduck – the author of several books on this subject. According to McCorduck, if the problem is solved, it is no longer part of AI. Well, the computer’s victory over Gary Kasparov in 1997 was not Artificial Intelligence, but a program searching for the best solution possible! What is AI then? The AI, which, according to the pioneers of this discipline from the 1960s, was supposed to surpass human intelligence by the end of the 20th century, and the chess victory seemed like a fulfilled prophecy?
The IT guru, Alan Turing, comes to the aid by saying that if a human is unable to distinguish between a machine’s response and a human response, then such a machine should be considered intelligent. The so-called Turing Test can also be questioned, but in order not to go any further into details than necessary, let’s assume that Artificial Intelligence is a set of techniques enabling the machine to imitate human intelligence, regardless of the set of rules or neural networks. So called mimicry made by machines.
Since the attack of bloodthirsty super-machines does not threaten the world (at least for now), and the forecasts of the pioneers of Artificial Intelligence have been painfully verified by the reality, how is it with this perception of machines? Are they intelligent or not?
Grzegorz Gwardys: We are currently living in an era of Weak Artificial Intelligence (Narrow AI), which is not aware, does not feel or is driven by emotions. So, imitation of human intelligence is limited to acting in a pre-defined range. For the so-called Strong Artificial Intelligence (General AI), exceeding the capabilities of the human mind, i.e. the one we know from the movies, we have to wait.
If we are talking about such a narrow scope, wouldn’t the rules mentioned above be enough?
Grzegorz Gwardys: Over the years, it turned out that the rules are not enough. Let’s return to the Turing Test for a moment and expand it slightly. Let’s assume that the question asked relates to a displayed photo. What is obvious to us, i.e. our ability to identify and associate objects, words and other concepts and entities, is no longer obvious for machines. Therefore, the machine must be taught to recognize them. We are talking here about teaching, because what rule should be created so that the machine could say that the picture shows, for example, a cat, or a dog or a bicycle?
This is very difficult, so image processing techniques and the AI sub-fields were mixed, this is called Machine Learning (ML). Machine Learning is nothing but machines learning perception. ML techniques use so-called labeled data to produce statistical models predicting the likelihood of what is shown, e.g. in the photo. Therefore, the data here are photographs, and the labels contained in an additional text file are information about the objects presented in these photographs. It is also talked about annotation or annotating data.
The essence of machine learning is basically reflected in its name. And what is the so-called Deep Learning (DL)?
Grzegorz Gwardys: Just as Machine Learning (ML) is a sub-field of AI, Deep Learning is a sub-field of ML. Earlier I mentioned combining Image Processing techniques with Machine Learning. For classic ML algorithms, it would be very difficult to process the whole photo. That is why the whole range of Image Processing techniques began to be used so that individual image features (from edges and colors, to sophisticated image descriptors), and not the entire image, are the input for ML algorithms. In the case of Deep Learning, we are talking about neural networks. These are mathematical structures that actually resemble a synapse in their structure. Hence the name. The neural network is nothing more than a program trained (i.e. taught) for specific purposes and tasks: e.g. to recognize characteristic points of the face or to distinguish a cat from a dog in a photo. The disadvantage of Deep Learning is the necessity to have a huge amount of labeled data. On the other hand, the unquestionable advantage: the ability to take these data as they are, i.e. without devising or developing traits. While training the neural network, the traits that we had to invent before were now produced by the network itself. This is indeed a revolution because we have overcome the need for domain knowledge and the arduous production of needed traits, using reserves of accumulated data. I think that is why business is so deeply interested in Deep Learning, which is what is often referred to as AI in marketing.
A bright future for AI / DL then?
Grzegorz Gwardys: I am more realistic than optimistic. Accumulation of data is one thing, and ensuring their quality is a completely different pair of shoes. If someone has 10 million files with the .JPG extension and does not know what is in these pictures (because there are no labels), then first you have to carry out the tedious process of data preparation, so that they can be used to teach neural networks. The appetite of various businesses for this type of solutions is enormous, but at the same time expectations are excessive and unrealistic. For example: current image classification solutions work well for a thousand of different types of objects. Business, for that matter, can expect 10,000 types and 100% accuracy. On the other hand, it’s good because the cooperation between the scientific community and industry is tightening, but we must remember that AI (or rather Machine Learning and Deep Learning) is not a panacea for everything. For now, we live in an era of Weak Artificial Intelligence, where the scope of activity is narrowed. Therefore, mutual understanding is needed: business needs to be understood by technical specialists and current AI capabilities need to be accepted by the business.
The Internet is overwhelmed by reports of artificial intelligence composing music or painting images, and Facebook is bursting at the seams with “smart” makeup apps, assessing age and changing hairstyle … How does an expert in machine learning and a realist approach practical applications of AI?
Grzegorz Gwardys: It is a fact that the use of AI techniques is becoming more and more popular. I perceive the aforementioned composing music or painting pictures as marketing tricks than the actual manifestation of the creative abilities of machines. But there are really many areas of practical applications of AI. For years, using AI techniques, the search engines have been optimized and personalized recommendations have been created for individual users based on their search results. Deep Learning techniques are used by commonly used image recognition systems (including faces or human figures), sound recognition systems (where speech recognition systems are the most popular), as well as automatic translation systems. This category also includes programs that examine photos and documents for prohibited content. Generally, wherever we talk about a large amount of unstructured data (such as an image or speech) you can see the potential of Deep Learning, which, as has been said before, learns the features. Personally, I root for all medical solutions. For example, in the paper “Thoracic Disease Identification and Localization with Limited Supervision” it was shown that AI can help identify a disease in real clinical environment, where data is scarce and doctors require justification of the disease. On the other hand, the authors emphasize that only part of the radiologist’s work can be automated through Artificial Intelligence and in the near future we are talking about supporting the work of doctors, not replacing them. I think this observation also applies to other industries, because the current methods of Artificial Intelligence, Weak Artificial Intelligence, operate only on a section of reality.
We will come back to the topic of Artificial Intelligence more than once. Today’s episode ends with an optimistic conclusion that there is no threat from intelligent super-machines, which at least in the near future will not take over the world and will not lead to the destruction of the mankind. Everyone interested in the discussed subject is encouraged to read the remaining blog entries.