Computers are suitable for working with symbols. Developers knew about this fact at the dawn of artificial intelligence. In the 1950s, scientists tried to compare their intuitions about solving the problem with symbols and algorithms available on emerging first computers. Such a comparison has led to unprecedented success, including the ability to prove mathematical theories and communicate with computers automatically.
Now the new golden era of artificial intelligence has begun. Large global corporations like Google, Inbenta or Facebook annually invest billions of dollars into AI research and create unique digital products. But algorithms that underlie such know-how are very different from the systems developed in the 1950s.
Such an engineering style (now called good old-fashioned AI or simply GOFAI) has been replaced by a technology more directly inspired by the brain. Nowadays, artificial neural networks are trained directly on the database to see, speak, play and plan. With no explicit symbols in sight, neural networks seem to prove the fallacy of the system GOFAI.
Automate 84% of user questions
AI Engine can transform your data into knowledge, and answer any question your users asks, complexity automatically
However, some scholars do not agree with this view. Despite the impressive achievements of neural networks, they have disadvantages closely related to the advantages of symbolic AI. Does it mean that AI needs to bring back symbols?
The role of symbols in artificial intelligence
We often use symbols to define things (a table, a dress, etc.), people (a salesman, a cop), abstract concepts (money transfers), and something that doesn’t have a physical shell (website, social media page, etc.). Sometimes signs describe actions or states; they can be grouped in a hierarchy.
As for symbolic artificial intelligence, it considers symbols as a visual pattern, says a character or string of characters, which have a specific meaning, and this sign points to something else. It may be the variable x, pointing to an unknown quantity or word, for instance, «a rose», which indicates a red flower with petals twisted and layered on top of each other.
The ability to communicate using symbols is one of the main things which make us smart. Therefore, symbols have also played an essential role in developing artificial intelligence.
Early researchers believed they could precisely describe any aspect of learning and that a machine could model it. Consequently, symbolic AI took center stage and became an essential part of research projects. Scientists created tools to define and manage symbols.
Many concepts and tools you know in computer science are the results of such integration. Symbolic AI programs rely on creating explicit structures and behavior rules.
Benefits of Symbolic AI
Symbolic artificial intelligence has been rapidly developing at the dawn of AI and computing. They allow us to easily visualize the logic of rule-based programs, communicate them, and fix any problems.
Symbolic AI is the best option for settings with clear rules; you can easily take input and transform it into symbols. Rule-based systems still make up the majority of computer programs, including those to provide the creation of deep learning apps.
However, symbolic AI doesn’t work when dealing with chaos. Let’s talk about computer vision – the science of how computers recognize the sense of pictures and videos. Image: You have a photo of the dog and plan to build a program to detect images containing your dog. You create a rule-based program that uses new photos as input, compares the pixels to the original photo, and draws conclusions about whether your dog is in the picture.
Such software will work if you only provide an exact copy of the original photo. Even small changes in the image will give a negative answer; if you photograph the dog from a different angle, the program will not work.
One option is to take a picture of the dog from all possible angles to compare each new picture with all the images. Even if you take a million photos of a cat, you still won’t be able to account for all possible situations, a simple change in lighting can cause the program to crash.
The dog example may seem strange, but it clearly shows the main problems of symbolic artificial intelligence. You can’t define rules for a chaotic dataset we encounter in real life. However, it can’t be translated to direct rules, including speech recognition and natural language processing.
Specialists have tried many times to create complex symbolic AI systems that can cover many rules from one industry, e.g., to make a medical diagnosis. They require extensive efforts of specialists in a particular industry and software developers and work only in limited use cases.
Areas where symbolic AI has been successful
Nowadays, many AI mechanisms operate based on symbolic AI, the most common systems:
Constraint satisfaction – it’s the procedure of issue resolution by satisfying specified conditions or restrictions. For instance, you must color a map using only green, yellow and red, but you cannot use the same color on two adjacent lots.
Natural language processing (NLP) is the sphere of AI, which allows machines to recognize human speech and provide communications between people and computers. The most famous examples of NLP are intelligent assistants, like Alexa from Amazon and Siri from Apple, predictive text applications, chatbots and search engines.
Logical Inferences. Symbolic AI relies on rules; they are helpful for logical reasoning. The machine analyzes given practices and evidence to conclude.
It’s not a complete list of industries where computer engineers use symbolic AI.
Neural networks vs. symbolic AI
Neural networks appeared around the same time as symbolic AI, but they were not used since non-symbolic systems required significant computing power, which was not available. In recent decades, thanks to the greater availability of information and increased computing power, deep learning has gained popularity and began to supplant the symbolic systems of AI.
The main advantage of neural networks is working with chaotic and unstructured information; back to the dog example. Instead of manually looking for dog pixels, people can train the algorithm on different images of such animals. Then the neural network creates a static model for dog pictures. When you show the system a new image, it will check the possibility there is a dog.
Deep learning and neural networks are great at tasks that symbolic AI cannot do. They revolutionized computer vision apps, i.e., facial recognition or cancer detection. Deep learning led to progress in solving language problems.
Deep neural networks are also helpful for reinforcement learning, AI models, which determine their behavior via trial and error. Developers use this type of artificial intelligence to create complex games, like StarCraft, Dota, and others.
Despite the benefits of deep learning and neural networks, they also have some disadvantages compared to symbolic AI:
Deep learning algorithms are opaque, and understanding their work baffles even creators. Such a mess makes it harder to troubleshoot their inner workings.
Neural networks are data-hungry. Unlike symbolic AI, neural networks lack the concept of symbols and the hierarchical structure of knowledge. It’s challenging to comprehend how the system concluded.
Deep learning and neural networks require manual coding of knowledge and rules for the learning process, creating additional problems.
All these limitations make it challenging to use non-symbolic AI to solve problems related to logic and reasoning in the field of natural sciences or mathematics.
Problems with symbolic artificial intelligence
In the 1990s, specialists planned to abandon symbolic AI techniques when they realized that they couldn’t handle the problems of knowing common sense. Since symbolic AI works with explicit representations, developers didn’t consider implicit knowledge, such as «sugar is sweet» or «the mother is always older than her child». There is too much implicit knowledge around us to ignore it.
Although symbolic artificial intelligence demonstrates good reasoning abilities, it is difficult for him to instill the ability to learn. Since such an algorithm can’t learn by itself, developers had to add new rules and data constantly. It turned out that the more information the machine receives, the less accurate its results become.
The current positions of symbolic AI
Some programmers believe that the best years of symbolic AI are over, but such a statement is far from the truth. Actually, rule-based AI systems are still very essential in modern applications. Leading experts in the industry are confident that symbolic tools always will be a necessary component of artificial intelligence.
There are many attempts to combine symbolic AI and neural networks. One of the most famous examples is the Neuro-Symbolic Concept Learner, a hybrid AI algorithm developed by the MIT-IBM Watson AI Lab. NSCL successfully utilizes rule-based programs and neural networks to solve visual problems without direct supervision. Such a model learns by watching images, recognize paired questions and answers. Unlike systems that use only symbolic artificial intelligence, NSCL models do not face the problem of analyzing provided photos.
It’s just one example. Soon, more projects may use symbolic AI in a broader concept with neural networks to conduct rigorous analysis and compare large amounts of data to determine correlations to training systems. It’s easy to imagine a future where artificial intelligence algorithms have innate abilities to learn and think clearly. Nowadays, we must accept that symbolic AI is the best way to solve problems that require knowledge representation and logical processes.
The actual examples of using symbolic and hybrid AI
In many cases, utilizing unstructured databases from different documents, social networks, and emails and turning them into actionable information is helpful. Below are some instances of how companies use symbolism and hybrid to address weaknesses.
Insurance – specialists have to deal with unstructured data from different formats in this sphere. Through the correct use of symbolic technologies, insurance organizations can extract key data to support policy reviews and risk assessments. It reduces the risk and workflow redundancy, allowing underwriters to consider more claims.
Media and Publishing. After the transition to digital technologies, media and publishing became more competitive. The strong interaction matters a lot because now customers can decide which content they want to read first. The media successfully process, categorize, and tags more than 1.5 million news daily using symbolic AI tools. Such manipulations help identify keywords and the most exciting topics for readers at scale.
Banking. Despite technological progress, banking remains in digitization. The use of outdated technologies simultaneously limits the processing of many calls, emails, and virtual data requests. From the knowledge-based FAQ section to email automation and classification, an effective AI platform can overcome obstacles through the service supply chain.
Even though symbolic artificial intelligence fails in some areas, it has ensured the rapid development of science and technology to create intelligent machines and software. It is still used effectively now. Experts are exploring the possibility of combining symbolic AI and neural networks to achieve advances in artificial intelligence.