#artificialintelligence #121: AI - reasoning - LLMs - knowledge graphs - neurosymbolic AI

Welcome to #artificialintelligence #121

Excuse the short gap between newsletters

I am still in DC and off for a brief holiday next week

Hence, this newsletter before I go on holiday

We are getting close to launch of #Salooki in late August / Sep

So, I wanted to share more about the core technology we are developing in our labs in USA and UK

The class of AI we are developing is called neurosymbolic AI. This is an emerging area and is based on the ideas of symbolic AI (which did not gain traction in the early days of AI). It can be thus seen as ‘hybrid AI’ ie a combination of symbolic AI and neural networks. Symbolic AI was the dominant paradigm of AI research from the mid-1950s until the mid-1990s. It was then superseded by neural networks which is the current dominant paradigm. Neurosymbolic AI is an emerging AI which combines the two worlds.??

Specifically, we use LLMs to accelerate the creation of knowledge graphs which are a component of symbolic AI.

Below is a relatively long discussion on the rationale

I used chatGPT to explain some of these ideas. I have long been developing these ideas but I still think this is an experimental domain (and a bit tangential to the current core focus of deep neural networks)

AGI and Reasoning - Synopsis

chatGPT takes us on the road to Artificial General Intelligence(AGI). AGI refers to highly autonomous systems that outperform humans at most economically valuable work.? The more capable AGI is, the more decisions can be delegated to AGI - and consequently, greater the productivity of humans and AI working together. However, a number of gaps remain in current technology before AGI is achieved.? One of these is the ability of AI to reason. Through our work, we hope to bridge this gap by the use of knowledge graphs(KG). Specifically, we want to use LLMs to generate KGs - thereby providing LLMs with an ability to perform some reasoning capabilities.?

The class of AI we are developing is called neurosymbolic AI. This is an emerging area and is based on the ideas of symbolic AI (which did not gain traction in the early days of AI). It can be thus seen as ‘hybrid AI’ ie a combination of symbolic AI and neural networks. Symbolic AI was the dominant paradigm of AI research from the mid-1950s until the mid-1990s. It was then superseded by neural networks which is the current dominant paradigm. Neurosymbolic AI is an emerging AI which combines the two worlds.??

We apply these ideas to the domain of interdisciplinary research. This is a large and complex problem which we aim to implement using graph databases, polyglot notebooks. Autonomous AI agents and Azure technology initially. The technology has multiple applications which we will continue to explore.??

Background

The paper "Sparks of Artificial Intelligence" by Bubeck et al. (2023) argues that large language models (LLMs) like GPT-4 are showing signs of artificial general intelligence (AGI), specifically in the area of reasoning. The authors present several examples of GPT-4's ability to reason, including:

• Solving novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology, and more, without needing any special prompting.
• Making inferences and drawing conclusions from incomplete or ambiguous information.
• Explaining its own reasoning process.
• Understanding the physical world and how objects interact with each other.

According to the authors, these findings suggest that LLMs are capable of a more sophisticated form of reasoning than previously thought possible. They argue that this is a significant step towards the development of AGI, and that it raises important questions about the nature of intelligence and the future of AI.

GPT-4 was able to solve a puzzle that required an intuitive understanding of the physical world. The puzzle involved stacking blocks in a way that would allow a laptop to be placed on top without the blocks collapsing. GPT-4 was able to solve the puzzle by reasoning about the weight of the laptop, the size of the blocks, and the way that the blocks would interact with each other.

GPT-4 was also able to make inferences and draw conclusions from incomplete or ambiguous information. For example, when asked to write a story about a man who was walking down the street and saw a woman being mugged, GPT-4 was able to infer that the man would help the woman, even though this was not explicitly stated in the prompt.

GPT-4 was also able to explain its own reasoning process. When asked to explain how it solved the puzzle involving the blocks, GPT-4 was able to provide a detailed explanation of its thought process. This suggests that GPT-4 has a good understanding of its own reasoning abilities.

Overall, the findings in the paper "Sparks of Artificial Intelligence" suggest that LLMs are capable of a more sophisticated form of reasoning than previously thought possible. This is a significant step towards the development of AGI,?

What is reasoning in terms of AGI?

However, LLMs lack the ability to reason, in its complete sense. Reasoning is a crucial component of AGI, as it enables the system to process information, make inferences, and arrive at logical conclusions. AGI's reasoning capabilities are designed to mimic or surpass human-level reasoning in various domains. Reasoning has a number of sub-capabilities.

Deductive Reasoning: AGI systems are expected to excel in deductive reasoning, which involves drawing logical conclusions from given premises. They can perform tasks like theorem proving, logical inference, and formal reasoning. Examples include solving mathematical proofs, verifying logical consistency, or generating valid conclusions based on a set of rules.

Inductive Reasoning: AGI should possess strong inductive reasoning abilities to make generalizations or predictions based on patterns and observations. It involves inferring general rules or principles from specific instances. This capability enables AGI to learn from data, generalize knowledge, and make probabilistic judgments.

Abductive Reasoning: AGI systems are expected to have the ability to perform abductive reasoning, which involves generating the best possible explanation for a given set of observations or evidence. It helps AGI to make educated guesses or hypotheses based on incomplete or ambiguous information.

Analogical Reasoning: AGI's capacity for analogical reasoning allows it to find similarities or relationships between different domains or contexts. By recognizing patterns or similarities between known and unknown situations, AGI can transfer knowledge and apply existing solutions to new problems.

Meta-Reasoning: AGI systems may possess meta-reasoning capabilities, which involve reasoning about their own reasoning processes. This enables them to monitor and improve their own decision-making, detect biases, assess uncertainties, and modify their approaches accordingly.

Limitations of current AGI for reasoning

There are a number of limitations for LLMs to achieve the ability to reason.

Computational Complexity: AGI will likely require massive computational resources to perform complex reasoning tasks, especially those involving vast amounts of data or intricate logical processes. Scaling AGI to handle real-world problems at a human-level scale is a daunting challenge.

Lack of Common Sense Reasoning: Human reasoning often relies on common sense knowledge and intuition, allowing us to fill in gaps and make sense of ambiguous or incomplete information. Developing AGI with robust common sense reasoning remains an open problem as encoding this knowledge in a machine-readable format is complex.

Context Sensitivity: Reasoning often depends on understanding the context of a problem, which can be challenging for AGI systems. Ambiguities or changes in context can lead to incorrect conclusions or interpretations.

Handling Uncertainty: Human reasoning is adept at dealing with uncertain or probabilistic information, making informed decisions based on the available evidence. AGI needs to develop the ability to reason effectively in the face of uncertainty.

Bias and Ethical Challenges: AGI's reasoning capabilities heavily depend on the data it is trained on. Biases present in the training data can lead to biased reasoning, which could have significant ethical implications if AGI is applied in real-world decision-making scenarios.

Symbol Grounding Problem: AGI might face difficulties in understanding the meanings of symbols and concepts. Linking abstract symbols to real-world referents (symbol grounding) is crucial for reasoning effectively in diverse environments.

Creativity and Insight: Human reasoning often involves creativity, insight, and intuition in problem-solving. Developing AGI systems that can generate novel solutions and innovative approaches remains a challenging task.

Explainability and Interpretability: AGI systems capable of complex reasoning may be difficult to interpret or explain their decisions to humans. Ensuring transparency and interpretability in their reasoning processes is important for building trust and addressing safety concerns.

Transfer Learning: While AGI should be able to generalize across various domains, achieving seamless transfer learning from one domain to another remains a challenge.

Self-Improvement and Alignment: AGI systems may aim to improve their own capabilities, but ensuring that they align with human values and do not act in unintended ways during the self-improvement process is a significant challenge in AGI research.

how do knowledge graphs help in reasoning for AI

Knowledge graphs play a crucial role in reasoning for AI by organizing and representing information in a structured and interconnected manner. A knowledge graph is a type of graph database that represents knowledge as nodes (entities) connected by edges (relationships). These graphs provide a powerful foundation for AI reasoning because they enable machines to understand, infer, and reason over vast amounts of data efficiently. Here's how knowledge graphs help in reasoning for AI:

Data Integration and Semantics: Knowledge graphs can integrate diverse and heterogeneous data from various sources, providing a unified representation of information. The explicit semantic relationships between entities allow AI systems to comprehend the meaning and context of the data, facilitating more precise reasoning.

Inference and Deduction: With knowledge graphs, AI systems can perform deductive reasoning by traversing relationships and drawing logical conclusions. They can infer new facts from existing ones, identify implicit connections, and deduce missing information through pattern recognition and transitive relationships.

Contextual Reasoning: Knowledge graphs enable AI systems to reason within specific contexts by considering relevant relationships. This contextual reasoning helps machines understand the implications of information in different scenarios and arrive at more contextually appropriate conclusions.

Common Sense Reasoning: By capturing common sense knowledge in the form of relationships between entities, knowledge graphs aid AI in reasoning more akin to human common sense. This assists AI in filling gaps, resolving ambiguities, and making reasonable decisions based on intuition.

Answering Complex Queries: AI systems can use knowledge graphs to efficiently answer complex queries that involve multiple layers of information. The graph structure allows for the traversal of relevant paths and retrieval of relevant information, resulting in more accurate and comprehensive responses.

Explainable AI: Knowledge graphs provide a transparent and interpretable representation of the reasoning process. The logic used in navigating the graph and deriving conclusions can be examined, making AI systems more accountable and understandable.

Transfer Learning: Knowledge graphs facilitate transfer learning by allowing AI systems to generalize knowledge across domains. The structured representation of information enables the transfer of insights gained from one domain to another, enhancing reasoning capabilities.

Real-time Reasoning: Knowledge graphs can be designed to support real-time updates, making them dynamic repositories of knowledge. AI systems can reason over the most up-to-date information, ensuring they are equipped to handle dynamic and evolving scenarios.

Scalability and Efficiency: Knowledge graphs are highly scalable, enabling efficient reasoning even with vast amounts of data. The graph structure allows for optimized storage, retrieval, and processing, making reasoning tasks more computationally tractable.

Integrating Human Knowledge: Knowledge graphs can be curated and enriched with human expertise, capturing expert knowledge in a structured format. AI systems can then leverage this curated knowledge for advanced reasoning tasks.

Introduction to knowledge graphs and Ontologies

Within the realm of knowledge representation, Knowledge graphs and ontologies are closely related but serve different purposes. Ontologies, are formal representations of knowledge that define concepts, relationships, and properties within a specific domain. They provide a shared vocabulary and understanding of the entities and their relationships in a particular knowledge domain. Ontologies focus on establishing a structured, hierarchical classification of concepts and defining the relationships and constraints that govern those concepts. For an Environmental Ontology i.e. an ontology for representing environmental concepts, example Classes could be Ecosystem, Species, Habitat, Pollutant etc and? example Properties could be inhabits, affects, isPollutedBy. Specific domains have ontologies for ex Gene Ontology?

A knowledge graph is a structured representation of knowledge that captures entities, relationships, and attributes in a graph-like structure. It models real-world entities and their interconnections, allowing for rich and flexible data representation. Knowledge graphs are often created by integrating data from various sources and can be dynamically updated as new information becomes available. They emphasize the organization and interlinking of data to create a comprehensive knowledge base.?

Thus, Ontologies provide a high-level conceptual framework, while knowledge graphs instantiate and populate that framework with real-world data and relationships.

Examples of KG include Google Knowledge Graph, Wikidata,? DBpedia etc There are a number of ways in which you can implement KGs commercially through Graph databases like Apache Jena, Neo4,? Amazon Neptune,Stardog, MongoDB, Apache Cassandra, Azure Cosmos db etc. Graph databases are based on the idea of Resource Description Framework (RDF) data. The RDF standard supports querying using SPARQL. More broadly, Knowledge graph engineering refers to the process of designing, developing, and maintaining knowledge graphs. KG engineering includes steps like Data Integration and Cleaning; Ontology Development; Entity and Relationship Extraction; Graph Storage and Querying and Graph Visualization and Exploration.?

Our objective

We are essentially looking at LLMs to accelerate the creation of knowledge graphs especially considering that given a body of knowledge , we can create multiple ontologies of the same subject using LLMs.

For example: consider a domain like smart cities. We could have multiple ontologies: City Ontology: An ontology representing urban infrastructure, services, and governance; Energy Ontology: An ontology representing energy-related concepts and relationships. Etc.?

Each of these could be created using LLMs. By combining these ontologies, in a single knowledge graph, a holistic understanding of smart cities can be achieved, encompassing aspects of urban planning, energy management, and transportation systems. Once created, they can be used in research, applications, learning etc. We are particularly interested in using these ideas to solve hard and complex scientific problems spanning interdisciplinary domains.??

LLMs and knowledge Graphs

There are essentially two ways in which you can use LLMs and knowledge graphs

1. Use the LLM to build a KG and?
2. Use the LLM as a natural language interface to the KG

You can use the LLM at each stage of the KG: Identify the relevant entities; Explore entity relationships and attributes; Extract question patterns; Generate questions; Incorporate context and user input;? Fill in Placeholder Values; Incorporate Context and Diversity.?

Applications to Research and Learning

By accelerating the creation of knowledge graphs, we can? impact a number of areas in research, science, learning, complex problem solving via creative thinking especially in interdisciplinary areas - where we need new ways to speed up education and research. KGs accelerated by LLMs can provide an element of reasoning which accelerates this research.?

1. Multiple ontologies can create multiple perspectives as discussed before?
2. We can use knowledge graphs for reasoning. This includes causal reasoning; deductive Reasoning based on logical rules; Semantic Reasoning; Contextual Reasoning; Ontological Reasoning and abductive Reasoning i.e. inferring the best explanation or hypothesis for a given set of observations or facts.?
3. Bloom's Taxonomy and ontologies.? Bloom's Taxonomy (see our work here in education) is a hierarchical framework used in education to classify and categorize different levels of cognitive learning. While Bloom's Taxonomy focuses on classifying cognitive skills, ontologies can be used to organize and categorize knowledge itself. Ontologies can support the implementation and application of Bloom's Taxonomy by providing a structured framework for representing knowledge at different levels of complexity.For instance, an ontology might represent concepts and relationships at different levels of Bloom's Taxonomy. It could include basic concepts for remembering and understanding, more complex concepts for applying and analyzing, and higher-level concepts for evaluating and creating.
4. Ontologies play a crucial role in interdisciplinary research by providing a shared vocabulary and structured representation of knowledge that facilitates communication, collaboration, and integration of different disciplines. Areas of focus include conceptual Integration;Cross-Disciplinary Hypothesis Generation and Reasoning and Education and Training. These could apply in domains such as? Drug Discovery and Development; Protein Folding and Structure Prediction; Bioinformatics and Genomics; Climate Modeling and Environmental Science; Material Science and Nanotechnology.
5. Knowledge graphs could play a role in AI product management

Relation to neurosymbolic AI

The class of AI we are developing is called neurosymbolic AI. This is an emerging area and is based on the ideas of symbolic AI (which did not gain traction in the early days of AI). It can be thus seen as ‘hybrid AI’ ie a combination of symbolic AI and neural networks. Symbolic AI was the dominant paradigm of AI research from the mid-1950s until the mid-1990s. It was then superseded by neural networks which is the current dominant paradigm. Neurosymbolic AI is an emerging AI which combines the two worlds.??

Relationship between knowledge graphs and symbolic AI

Knowledge graphs and symbolic AI are closely related concepts, and knowledge graphs can be considered a specific implementation of symbolic AI.

Symbolic AI, also known as classical or rule-based AI, is an approach to artificial intelligence that uses explicit symbols and rules to represent knowledge and perform reasoning. In symbolic AI, information is represented in a structured, symbolic form, and algorithms are designed to manipulate these symbols according to predefined rules to perform various tasks.

Knowledge graphs, as mentioned earlier, are a type of graph database that represents knowledge as nodes (entities) connected by edges (relationships). Each node and edge in the graph corresponds to a symbol in symbolic AI, and the relationships between them represent rules or logical connections.

The relationship between knowledge graphs and symbolic AI can be summarized as follows:

Symbolic Representation: Both symbolic AI and knowledge graphs use symbolic representation to encode information. In symbolic AI, symbols are used to represent entities, concepts, and relationships, while in knowledge graphs, nodes and edges represent entities and relationships, respectively.

Rule-based Reasoning: Symbolic AI employs rule-based reasoning, where explicit rules and logical inference mechanisms are used to draw conclusions and make decisions. Similarly, knowledge graphs enable rule-based reasoning by traversing relationships and making inferences based on the graph's structure.

Logical Operations: Both symbolic AI and knowledge graphs allow for logical operations, such as conjunction, disjunction, negation, and quantification. These logical operations are fundamental for reasoning and knowledge representation in both paradigms.

Knowledge Representation and Inference: Knowledge graphs are a specific way of representing knowledge, and they offer a structured framework for performing inference and reasoning. The graph structure allows for efficient knowledge retrieval and reasoning over large amounts of interconnected data.

Explainability: Both symbolic AI and knowledge graphs provide a high degree of explainability. The explicit representation of information and reasoning steps allows for a clear understanding of how conclusions are reached, making the AI system more interpretable.

While knowledge graphs provide a powerful means of representing and reasoning over structured knowledge, symbolic AI encompasses a broader set of techniques, including logic-based reasoning, expert systems, and production rules. Knowledge graphs can be seen as a particular instantiation of symbolic AI, leveraging graph-based structures to represent and reason over knowledge effectively.

Overcoming the drawbacks of symbolic AI?

Symbolic AI, despite its advantages, also has several drawbacks that have limited its scalability and applicability in certain domains. Some of the major drawbacks of symbolic AI include:

Knowledge Engineering Bottleneck: Symbolic AI heavily relies on explicit knowledge representation and predefined rules. Building and maintaining a knowledge base can be a time-consuming and expensive process. Knowledge engineers must manually encode domain knowledge, making it challenging to scale the system to handle vast amounts of data or complex domains.

Brittle Knowledge Representation: Symbolic AI systems are sensitive to small changes in input or knowledge base, which can lead to unexpected and erroneous behavior. They may not handle uncertainty well and struggle with incomplete or noisy data, reducing robustness in real-world scenarios.

Lack of Generalization: Symbolic AI systems often lack the ability to generalize across different domains. As the system's knowledge is explicitly encoded for specific tasks, it may not transfer well to new or unseen scenarios.

Difficulty with Learning: Traditional symbolic AI methods face challenges in learning from data. They often require extensive manual intervention to update rules or modify the knowledge base based on new experiences or changing environments.

Scalability: As the knowledge base grows, the complexity of reasoning and search increases exponentially. Symbolic AI can become computationally expensive and struggle to handle large-scale problems efficiently.

Lack of Common Sense Reasoning: Capturing common-sense knowledge in symbolic representations is a difficult task. Symbolic AI systems may struggle with handling ambiguous or implicit information that humans easily comprehend.

Intractable Representations: Complex real-world problems might not have concise and tractable symbolic representations. Representing the nuances of human language, emotions, or perception in a symbolic form is challenging.

Combining Uncertainty: Symbolic AI struggles to handle uncertainty and probabilistic reasoning effectively, as most of its methods rely on deterministic rules and logical inference.

Expert-Dependent: Developing expert knowledge to encode rules requires human expertise, which might not always be readily available or may vary in quality.

Lack of Adaptability: Symbolic AI systems are generally static once built, and adapting them to new situations or evolving domains can be cumbersome.

While symbolic AI has been successful in specific domains, such as expert systems and certain reasoning tasks, addressing these limitations remains a challenge. As a result, researchers have explored other AI approaches, such as statistical and machine learning methods, to complement or replace pure symbolic AI in addressing more complex and data-intensive problems. Hybrid approaches that combine symbolic reasoning with statistical learning have shown promise in overcoming some of the drawbacks of pure symbolic AI.

neurosymbolic AI - overcoming the challenges of symbolic AI

Neurosymbolic AI, also known as hybrid AI or connectionist-symbolic integration, is an approach that combines elements of both neural networks (also known as deep learning) and symbolic AI to address the limitations of each paradigm and leverage their complementary strengths.

In classical symbolic AI, knowledge is represented explicitly as symbols and relationships, and reasoning is performed through rule-based systems and logical operations. Symbolic AI excels at logical reasoning, representing structured knowledge, and providing explainable results. However, it struggles with handling large-scale data, learning from raw sensory input, and dealing with uncertainty.

On the other hand, neural networks are powerful machine learning models that can learn patterns and representations from vast amounts of data. They excel in tasks like image and speech recognition, natural language processing, and reinforcement learning. However, neural networks are often seen as black boxes, lacking the explainability and transparency of symbolic AI.

Neurosymbolic AI aims to combine the strengths of both paradigms to create a more robust and flexible AI system. This integration can occur at various levels:

Knowledge Representation: Neurosymbolic AI seeks to represent knowledge in a hybrid manner, combining symbolic representations with learned embeddings or continuous vector representations. This enables the system to benefit from the structured and explicit knowledge of symbolic AI while leveraging the data-driven capabilities of neural networks.

Reasoning and Inference: Neurosymbolic AI integrates neural networks with symbolic reasoning engines to perform hybrid reasoning and inference. Neural networks can provide probabilistic outputs or learned features to guide the symbolic reasoning process, and the symbolic component can handle logical operations and explicit knowledge manipulation.

Learning and Adaptation: By incorporating neural networks, neurosymbolic AI can learn from data and adapt to changing environments, addressing the brittleness of purely symbolic systems. The learning component helps neurosymbolic AI to handle large-scale and complex data without relying solely on manually encoded rules.

Explainability: Neurosymbolic AI strives to retain the explainability and interpretability of symbolic AI by combining neural networks with rule-based systems. This facilitates the generation of more understandable and transparent results, which is crucial in applications where explainability is vital.

Neurosymbolic AI has shown promise in various AI tasks, including knowledge graph completion, program synthesis, question answering, and human-robot interaction. It provides a path towards more flexible and interpretable AI systems that can handle complex real-world problems while benefiting from both the knowledge representation capabilities of symbolic AI and the learning capacity of neural networks. As the field of AI progresses, neurosymbolic approaches continue to be an active area of research and development.

Conclusion

We are working on the following areas:

1. Looking at LLMs to accelerate the creation of knowledge graphs?
2. This enables us to achieve the following: given a body of knowledge? we can create multiple ontologies of the same subject using LLMs.
3. By accelerating the creation of knowledge graphs, we can? impact a number of areas in research, science, learning, complex problem solving via creative thinking especially in interdisciplinary areas.?

KG - LLM - neurosymbolic/hybrid AI is still an experimental area in my view - but I think it holds promise for the future of AI. Welcome thoughts

If you want to study with me please see?https://lnkd.in/euStW5a2

My newsletter?https://lnkd.in/dPq-Yy7

My mailing list for our new ventures in AI?https://lnkd.in/e6yxF5rd

Image source: https://pixabay.com/illustrations/human-soul-universe-stars-starry-5831238/