Print

Neuro-symbolic artificial intelligence

Fictional scenario: Rays of hope in a cup of tea

Miriam is a brilliant artist: her paintings and carvings have been showcased in many exhibitions over the past year. Despite her artistic success, she is deeply concerned about her rare disease. Doctors have told her that there are only a few hundred similar cases worldwide, which makes diagnosis challenging and suggests the condition may be underreported. They also informed her that research is still underway, it is difficult to find drugs to treat her condition and she would need personalised treatment in the event of a health emergency.

 

One day Miriam attends a dinner with former high school friends and discovers that Stella, not exactly her best friend back then, now works as a data scientist in a company that develops AI for healthcare solutions. Miriam decides to overcome her reticence and invites her to dinner to find out more.

 

During their conversation, Stella explains that while AI has made significant strides in areas like image recognition and data analysis, traditional neural networks face limitations when dealing with rare diseases. These models rely heavily on large datasets, which are unsuitable for uncommon conditions like Miriam's. Moreover, the available medical information is often unstructured, consisting of countless images and texts that traditional AI struggles to organize and interpret reliably as it requires a level of abstraction that goes beyond statistical capabilities.

 

A month later Stella attends a conference where she discovers a kind of new paradigm in AI and a bell rings! When she returns home, she immediately invites Miriam for tea and begins to enthusiastically sketch out a possible future scenario.

 

This innovative AI approach would excel by merging unstructured data, such as medical images and research papers, with structured knowledge bases like medical guidelines and ontologies. This integration would allow the AI to not only recognize patterns in the data but also apply logical abstract reasoning based on established medical knowledge. The AI would be able to analyse her unique medical data alongside existing treatments for similar conditions, identifying potential drug repurposing opportunities that traditional neural networks might overlook. By understanding the underlying biological mechanisms of her disease, the AI could suggest personalised treatment options tailored specifically to her needs.

 

Also, this could be very useful in emergencies by integrating patient records, verbal inputs from patients and witnesses, and real-time data from various sources with accuracy, thus allowing medical teams to make swift and informed decisions, enhancing the reliability and speed of their responses, which is particularly critical for patients with rare conditions as Miriam.

 

Despite the challenges ahead, Miriam feels a renewed sense of confidence in the future. A smile spreads across her face and meets Stella’s bright look. Rays of hope... in a cup of tea. 

Neuro-symbolic artificial intelligence

By Massimo Attoresi

Neuro-symbolic artificial intelligence (NSAI) refers to a field of research and applications, and a family of technologies that combine machine learning (ML) methods, particularly deep learning (DL), with symbolic approaches to computing and artificial intelligence.

The term ‘symbolic’ relates to approaches based on the explicit representation of knowledge, logics and rules, often using formal languages and the processing of those language items (symbols) via algorithms. For example, an equation in mathematics and physics or an expression in logics (e.g. a set A that is a subset of another set B) or instructions in a programming language are all framed with symbols. In symbolic AI, data scientists try to identify classes of objects (e.g. types of words, images) and link them with relationships and constraints using logic rules, making the knowledge machine-readable and usable to draw further logic inferences.

In nowadays ‘non-symbolic’ artificial neural systems, the representation of information is encoded by means of weighted connections among a large number of ‘neurons’ which are optimised to produce the desired output. While neural networks have demonstrated their ability to learn from unstructured datasets and their efficiency and scalability in processing large amounts of data in dynamic environments, these ‘non-symbolic’ approaches, have shown their weaknesses, particularly in identifying new patterns from complex datasets. These include the so-called ‘hallucinations’ (wrong inferences, as assessed by common sense or available knowledge), uncontrolled bias and lack of explainability in the results reached.

This has led to reconsidering the option of integrating DL-based approaches (‘non-symbolic’) with knowledge and logic-based ones (‘symbolic’), where respective strengths would be leveraged and weaknesses mitigated.

For example, in illness diagnosis, purely DL-based image classification would be able to identify an image pattern with a certain probability of being a specific illness, without further explanation. Adding medical knowledge and concepts would help diagnose diseases. This could include more information about the disease, how it relates to others in the same image, and statistical or organised health data for potential treatments. Looking at the patient's medical record could help make more accurate predictions.

In language applications, the new approach's results would not only be based on the statistical probability of finding a particular word sequence in a given textual context, but also benefit from semantic and syntactic recognition. Some researchers have defined NSAI as the third wave of AI.  Others see NSAI as the natural evolution of AI and a pathway towards artificial general intelligence, since its integration of learning-based and reasoning-based approaches would enable to understand, learn, and reason in a more human-like and versatile manner.

NSAI systems are hybrid models that combine DL and symbolic AI features in a variety of ways.

General concept of NSAI (from reference 1)
General concept of NSAI (Yu et al., 2023)


Recently a few taxonomies have been proposed. Here we refer to a specific one (see 1st recommended reading) that classifies NSAI system in:

  • Learning for reasoning - The aim is to use neural networks and DL to extract symbolic knowledge from unstructured data such as texts, images and video, with the aim to integrate it into symbolic reasoning and decision-making tasks.
  • Reasoning for learning - The aim is to incorporate symbolic knowledge into the training process of neural network-based systems to improve performance and interpretability. For example, in knowledge transfer models that are used to generalise a model and make connections between different domains, symbolic knowledge (e.g. semantic information) can guide the learning process in the new domain.
  • Learning-reasoning - The neural network and symbolic systems interact bidirectionally. Both components work together to achieve a balance in the problem-solving process. The neural network generates hypotheses or predictions on rules and relationship, which are then used by the symbolic component to perform logical reasoning. The results can be sent back to the neural network for improvement.

NSAI systems have been proposed in many domains. They include computer vision, natural language processing (language understanding, generating, and reasoning), recommendation systems and self-driving cars, where the complexity needs to be conjugated with clear, already known logical relationships between objects in the environment.

Development status

Research on NSAI has a long tradition but only in recent years has seen a steep increase, as shown by relevant literature. Researchers have started exploring the potential of applying NSAI in areas such as healthcare, to extract relevant information from medical literature or combining inferences on clinical data with medical general and personalised knowledge, and advanced robotics, to enhance robot intelligence and decision-making capabilities.

Yet, to our knowledge, NSAI has not gained significant presence in the market so far and even the evolution of natural language processing which already shows capacity to reflect and analyse (which is an area that could benefit from the new approach) is not for the moment integrated with ‘symbolic’ approaches.

A critical question stays in how to effectively combine neural and symbolic components without diluting their respective strengths, a task that requires innovative architectural designs and learning paradigms. Despite promising research exists, the quest for an efficient general integration strategy continues.

In NSAI, the symbolic components often meet efficiency challenges.  The construction of logic rules in symbolic approaches typically relies on manual efforts from domain experts. In these cases, neural networks are proposed to handle tasks that are computationally difficult in traditional symbolic systems. The automatic identification of rules in data, as well as the design of more robust and efficient symbolic representation learning methods, represent an important future research direction in the field.

Last but not the least, the future of NSAI is tied hand in glove to the evolution of neural networks, from whose potential and limits NSAI draws its reason to exist and, somehow paradoxically, its synergies. Recent developments in LLMs seem to be narrowing the accuracy gap with symbolic AI, as they are better able to deal with mathematical and logical challenges.

Whether NSAI represents the future of AI, or whether a purist neural-network-based future should be pursued, is a matter of debate among researchers. This is further sparked by the discussion over whether AI paradigms should follow (and how) the architecture and functioning of the brain. Artificial neural networks have been conceived as a sort of abstraction of how our brain physically works, as composed by interconnected neurons and, on a more abstract layer, symbolic representation and logics can be identified as the way we explicit the perception of our reasoning. A consequent question arises as to whether these two views on the way our brain works can be related to each other and how, and as to whether operationally they can complement each other. NSAI is an attempt in that direction.

Potential impact on individuals

Neural and symbolic approaches to AI complement each other as to their strengths and weaknesses, including their impact on data protection principles and individuals’ rights and freedoms. For example, symbolic AI can enhance transparency and accountability, while reducing the use of personal data, which is crucial for protecting individual rights in AI-driven decision-making processes.

Relying solely on neural network based approaches can have limitations or unsatisfactory results. DL has shown its limitations, often producing results that contradict not only specific domain expertise, but even known facts and common sense. This leads to accuracy problems, which in certain contexts cannot be tolerated. This is why embedding symbolic knowledge within DL can help to provide logical constraints and feedback within the learning process of neural networks.

NSAI comes also into the picture in reducing possible bias due to statistical misrepresentation by integrating existing knowledge and logic in situations where it is necessary to identify new classes of elements (e.g. rare animals, rare diseases, new objects, new concepts) with one or a few, or even zero labelled examples. All this leads to greater accuracy, for example in the outcome of a medical diagnosis or in identifying obstacles and objects in the path of an autonomous vehicle.

At the same time, finding relationships and rules in unstructured data through the use of deep learning multiplies the application domains in which symbolic knowledge can be used, thus avoiding or mitigating the weaknesses of statistical inference and increasing accuracy.

Furthermore, the use of the symbolic component as a replacement of DL, when suitable, can reduce the amount of data (including personal data) to train the model.

Another critical consideration is the compatibility of models built purely on neural networks, with the principles of explainable AI. Neural networks are unable to provide explicit logics and algorithms. Integrating the symbolic dimension offers new opportunities in terms of reasoning and interpretability. For example, through deductive reasoning and automatic theorem proving, symbolic systems can generate additional information and illustrate the reasoning process employed by the model, making it easier to understand how decisions are taken and on what basis. This would contribute to enhanced transparency of controllers as to the decision-making process, thus improving accountability.

However, NSAI technologies should not preclude the implementation of effective 'human oversight' of AI systems. The fact that the system is capable of ‘reasoning’ does not negate the need to ensure that privacy and ethical considerations are taken into account and that the adverse effects of system malfunction are limited, particularly in areas that have a direct and significant impact on individuals - such as medical diagnosis or treatment.

Suggestions for further reading