This project aims to explore the intricate process of concept acquisition in humans and critically evaluate whether connectionist artificial intelligence systems, specifically large language models (LLMs), can acquire abstract concepts in a manner similar to human cognition. This exploration is conducted within the philosophical frameworks of abstraction and empiricism, highlighting the stark differences in how humans and AI models form concepts.
The question of whether artificial intelligence systems can truly understand abstract concepts as humans do is a central issue in cognitive science and AI research. Human concept acquisition is a complex process involving sensory experiences, introspection, and context, while AI systems rely on pattern recognition, statistical correlations, and pre-trained datasets. This project seeks to shed light on the differences and similarities between these processes, ultimately questioning whether LLMs can achieve genuine linguistic competence or if their abilities are limited to mere mimicry.
Can connectionist artificial intelligence systems, specifically large language models (LLMs), acquire abstract concepts in a manner akin to human cognition, or are they constrained by their reliance on pattern recognition?
The project involves creating a Python simulation to model the differences in concept acquisition between human cognition and AI systems. The simulation comprises two main components:
- Human Cognition Model
- AI Model Based on LLMs
These components are implemented as Python classes, each representing their respective approach to concept formation and linguistic competence.
The Human Cognition Class is designed to simulate the human process of concept acquisition and abstraction. This model incorporates various elements, such as:
- Sensory Inputs: Represents how humans use sensory data (e.g., visual, auditory, tactile) to form abstract ideas. Sensory integration is inspired by Chiou & Ralph's (2019) dynamic interplay between sensory-motor and hub regions, which enables coherent semantic representations.
- Situational Context: Highlights the context-dependent nature of human cognition, referencing Barsalou & Wiemer-Hastings' (2005) findings that abstract concepts are deeply influenced by social, cultural, and environmental contexts.
- Memory Retrieval: Uses memory to support concept acquisition. Memory is hierarchical, as Carey (2011) suggests, combining innate structures and experiences.
- Emotional Integration: Borghi et al. (2018) emphasize the role of emotions in concept formation. Emotions provide the motivation and affective grounding necessary for humans to internalize abstract concepts.
- Innate Structures: The model incorporates innate cognitive structures (Spelke, 1998), which are thought to provide foundational support for abstract thinking and reasoning.
These elements are brought together through functions like perceive(), generate_output(), and other methods (combine_concepts(), reflect()), representing human processes such as introspection and synthesis.
The AI Cognition Class simulates the process by which large language models (LLMs) acquire and process concepts. This model includes:
- Training Data: Represents the preprocessed text data that LLMs use to form concepts. LLMs learn through statistical correlations across vast datasets, as argued by Cherkassky & Lee (2024).
- Model Parameters: The internal architecture of the LLM, such as hidden layers and weights, which determines how the model interprets input data.
- Reinforcement Learning: AI's learning mechanism involves fine-tuning based on domain-specific data and reward signals. However, this approach is criticized for its lack of developmental evolution compared to human learning.
- Fine-Tuning and Pattern Recognition: Highlights AI’s reliance on pre-trained models and pattern recognition, which contrasts with the flexible and context-driven human approach to learning.
The class includes functions like parse_data(), generate_output(), and visual tools for generating concept maps. These help illustrate how AI systems create relationships between data points without genuine abstraction.
The Python simulation allows users to input abstract concepts (e.g., "freedom," "justice") and choose whether they want to explore how a human or an AI system would acquire and interpret that concept. The simulation then produces either:
- Textual Explanations: Detailed explanations of how the concept is understood by each system, connecting the underlying processes to relevant psychological and philosophical theories.
- Concept Maps: Graphical representations created using NetworkX and Matplotlib to visualize the different approaches to concept acquisition.
The differences between human and AI conceptualization are highlighted through categories such as:
- Sensory Processing
- Context Interpretation
- Memory Retrieval
- Reinforcement Learning
- Emotional Integration
To date, significant progress has been made in constructing the theoretical and computational foundations of this project. This includes defining the two primary classes, gathering relevant literature, and aligning the code with key philosophical frameworks:
- Philosophical and Theoretical Basis: Developed a strong theoretical foundation bridging philosophical theories of abstraction, empiricism, and nativism with computational methodologies. This has been crucial in contrasting the cognitive capabilities of humans and AI.
- Technical Implementation: Outlined the structure for both classes, implemented methods for sensory integration, context interpretation, and emotional processing in the Human Cognition Class. Defined training and fine-tuning mechanisms for the AI Model Class.
Translating complex human cognitive processes into functional code has presented challenges:
- Complexity of Introspection and Emotional Integration: Accurately representing introspective processes and emotional integration without oversimplification is a key difficulty.
- Balancing Technical and Philosophical Depth: The project requires maintaining a balance between a detailed philosophical inquiry into human cognition and the technical robustness of the AI simulation.
Moving forward, the priorities for the project are:
- Finalizing the Coding Framework: Completing the implementation of methods for the Human Cognition and AI Cognition classes.
- Developing Use Cases: Creating specific user scenarios to demonstrate divergences between human and AI concept processing, illustrating the limitations of AI systems in grasping true abstraction.
- Integrating Analysis with Simulation: Expanding the written analysis to seamlessly incorporate results from the simulation, providing a comprehensive evaluation of how AI and humans differ in conceptualization.
Barsalou, L. W., & Wiemer-Hastings, K. (2005). Situating abstract concepts. Cambridge University Press, 129–163. https://doi.org/10.1017/cbo9780511499968.007.
Summary: Barsalou and Wiemer-Hastings argue that abstract concepts are dynamically grounded in situational and introspective contexts. They contrast abstract concepts, which rely heavily on situational variability, with concrete concepts, which exhibit stronger sensory grounding. Assessment: This source provides a critical foundation for articulating the situational and introspective elements of human abstraction. It highlights the absence of such grounding in LLMs and will serve as a cornerstone for critiquing AI’s inability to simulate conceptual context.
Borghi, A. M., Barca, L., Binkofski, F., & Tummolini, L. (2018). Abstract concepts, language and sociality: from acquisition to inner speech. Philosophical Transactions of the Royal Society B Biological Sciences, 373(1752), 20170134. https://doi.org/10.1098/rstb.2017.0134.
Summary: Borghi et al. argue that abstract concepts rely heavily on language and social interactions for acquisition and internalization. They highlight the role of inner speech in managing and integrating these concepts. Assessment: This source will be used to illustrate the fundamentally social nature of abstract concept formation in humans, contrasting it with the isolated statistical models used in LLMs.
Carey, S. (2011). Précis of The Origin of Concepts. Behavioral and Brain Sciences, 34(3), 113–124. https://doi.org/10.1017/s0140525x10000919.
Summary: Carey presents a hybrid model of concept acquisition, arguing that it combines innate structures with empirical experiences. The paper focuses on how humans generate novel abstract ideas through hierarchical learning. Assessment: This source provides theoretical grounding for the nativist perspective in my project, which I will contrast with AI’s reliance on preprocessed data and lack of innate structures.
Cherkassky, V., & Lee, E. H. (2024). A perspective on large language models, intelligent machines, and knowledge acquisition. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2408.06598.
Summary: This paper critiques LLMs’ epistemic abilities, emphasizing their inability to form true concepts due to reliance on pattern recognition. Cherkassy et al. propose that LLMs generate only a superficial mimicry of human knowledge. Assessment: This is a cornerstone source for arguing that LLMs lack true abstraction, as their “knowledge” is entirely statistical and devoid of conceptual grounding.
Chiou, R., & Ralph, M. a. L. (2019). Unveiling the dynamic interplay between the hub- and spoke-components of the brain’s semantic system and its impact on human behaviour. NeuroImage, 199, 114–126. https://doi.org/10.1016/j.neuroimage.2019.05.059.
Summary: This paper explores the neural mechanisms underlying the human semantic system, focusing on the interaction between the “hub” (anterior temporal lobe) and “spoke” (sensory-motor regions) components. Chiou et al. argue that this dynamic interplay enables the integration of diverse inputs into coherent semantic representations. Assessment: This source will be instrumental in illustrating how human cognition integrates multimodal sensory data to form abstract concepts. It will also support the argument that human semantic processing relies on a highly interactive neural network, whereas AI models are constrained by static, ungrounded data representations.
Constantinescu, I., Pimentel, T., Cotterell, R., & Warstadt, A. (2024). Do language models have a critical period for language acquisition? arXiv (Cornell University). https://doi.org/10.48550/arxiv.2407.19325.
Summary: This study investigates whether LLMs mimic humans’ critical period for language acquisition, finding that LLMs exhibit no such developmental constraints. Assessment: This source will support the argument that LLMs, unlike humans, do not undergo a developmental trajectory, reinforcing their fundamental differences from human cognitive processes.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2016). Building machines that learn and think like people. Behavioral and Brain Sciences, 40. https://doi.org/10.1017/s0140525x16001837.
Summary: This paper outlines principles for developing AI systems that emulate human cognitive processes, emphasizing learning from small data, compositionality, causality, and grounded understanding. Lake et al. argue that current machine learning systems, including LLMs, fail to replicate the flexibility and conceptual depth of human thought. They propose a shift toward incorporating structured cognitive models to achieve human-like AI. Assessment: This source will provide critical insights into the limitations of LLMs, specifically their reliance on large datasets and lack of causal reasoning. It will be used to argue for the fundamental differences between human conceptual learning and AI pattern recognition, supporting the position that LLMs do not achieve true abstraction.
Laurence, S., & Margolis, E. (2012). Abstraction and the origin of general ideas. Philosopher’s Imprint, 12(19), 1–22. https://quod.lib.umich.edu/cgi/p/pod/dod-idx/abstraction-and-the-origin-of-general-ideas.pdf?c=phimp;idno=3521354.0012.019;format=pdf.
Summary: Laurence and Margolis analyze how abstraction enables the formation of general ideas, emphasizing cognitive mechanisms that allow humans to go beyond sensory experiences. Assessment: This will be used to show that human abstraction processes are fundamentally different from LLMs, which cannot generalize beyond data patterns.
Magid, R. W., Sheskin, M., & Schulz, L. E. (2015). Imagination and the Generation of New Ideas. Cognitive Development, 34, 99–110. https://doi.org/10.1016/j.cogdev.2014.12.008.
Summary: This article explores the role of imagination in the creation of new ideas and concepts, particularly in childhood development. Assessment: I will use this to argue that imagination, a key component of human cognition, is entirely absent in LLMs.
MacDonald, B., & Witten, I. (1989). A framework for knowledge acquisition through techniques of concept learning. IEEE Transactions on Systems Man and Cybernetics, 19(3), 499–512. https://doi.org/10.1109/21.31057.
Summary: MacDonald et al. present methods for concept learning in AI systems, emphasizing algorithmic approaches and their limitations in replicating human conceptual understanding. Assessment: This will provide context for historical challenges in AI’s attempts to replicate human-like knowledge acquisition.
Marcus, G. (2018). Innateness, AlphaZero, and artificial intelligence. arXiv (Cornell University). https://doi.org/10.48550/arxiv.1801.05667.
Summary: Marcus critiques modern AI systems, arguing that they lack the innate scaffolding present in human cognition. He highlights the limitations of deep learning models, such as AlphaZero, in understanding abstract principles and posits that integrating innate structures into AI design could enhance their capabilities. The paper contrasts AI’s reliance on statistical pattern recognition with the innate, flexible reasoning humans use to form abstract concepts. Assessment: This source will bolster arguments about the necessity of innate structures for abstract thought, positioning human cognition as fundamentally different from AI systems. It will also support critiques of the empiricist approach in current AI, aligning with philosophical arguments about the limitations of antecedently supplied labelled training datasets categorized according to various grouping principles.
McDonald, J. L., & MacWhinney, B. (1991). Levels of learning: A comparison of concept formation and language acquisition. Journal of Memory and Language, 30(4), 407–430. https://doi.org/10.1016/0749-596x(91)90014-b.
Summary: The article emphasizes hierarchical learning, comparing how humans form concepts and acquire language in structured stages. Assessment: This will support my argument that humans progress through learning stages, unlike LLMs, which do not evolve over time.
Miller, S., Jungheim, M., & Ptok, M. (2014). Erstspracherwerbsforschung und Spracherwerbstheorien. HNO, 62(4), 242–248. https://doi.org/10.1007/s00106-014-2855-z.
Summary: This paper discusses theories of first language acquisition, including nativist, empiricist, and interactionist perspectives. Assessment: It will provide comparative insights into how language acquisition theories differ from the mechanisms employed by LLMs.
Rohlfing, K. J., Wrede, B., Vollmer, A., & Oudeyer, P. (2016). An Alternative to Mapping a Word onto a Concept in Language Acquisition: Pragmatic Frames. Frontiers in Psychology, 7. https://doi.org/10.3389/fpsyg.2016.00470.
Summary: This paper introduces pragmatic frames, arguing that social and situational contexts are crucial for word-to-concept mapping. Assessment: This will highlight the limitations of LLMs, which lack access to pragmatic and social contexts.
Schulz, L. (2012). The origins of inquiry: inductive inference and exploration in early childhood. Trends in Cognitive Sciences, 16(7), 382–389. https://doi.org/10.1016/j.tics.2012.06.004.
Summary: Schulz discusses how children use exploration and inductive reasoning to generate new ideas, emphasizing active learning processes. Assessment: This supports the argument that human learning is dynamic and exploratory, contrasting with the static nature of LLM training.
Spelke, E. S. (1998). Nativism, empiricism, and the origins of knowledge. Infant Behavior and Development, 21(2), 181–200. http://harvardlds.org/wp-content/uploads/2017/01/spelke1998-1.pdf.
Summary: Spelke advocates for a nativist view, proposing that humans possess innate structures that guide early knowledge formation. Assessment: This will contrast human nativist capacities with the data-driven empiricism of LLMs.
Wei, H. (2020). The evolution of concept-acquisition based on developmental psychology. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2011.13089.
Summary: This paper outlines developmental trajectories in concept acquisition, focusing on how abstraction evolves over time. Assessment: It will emphasize the absence of temporal evolution in LLM learning processes.
Witten, I. H., & MacDonald, B. A. (1988). Using concept learning for knowledge acquisition. International Journal of Man-Machine Studies, 29(2), 171–196. https://doi.org/10.1016/s0020-7373(88)80045-2.
Summary: Witten et al. critique early AI approaches to concept learning, noting their reliance on rigid algorithms. Assessment: This will provide historical context for the limitations of computational models in replicating human cognition.
Xing, Y., Shi, X., Shen, F., Zhao, J., Pan, J., & Tan, A. (2018). Perception Coordination Network: a neuro framework for multimodal concept acquisition and binding. IEEE Transactions on Neural Networks and Learning Systems, 30(4), 1104–1122. https://doi.org/10.1109/tnnls.2018.2861680.
Summary: This paper presents a neural framework for integrating sensory modalities to form concepts. Assessment: I will use this to contrast multimodal integration in human cognition with the unimodal data processing of LLMs.
This project aims to provide a nuanced understanding of concept acquisition in humans and AI systems, focusing on whether AI can truly replicate human-like abstraction or if its abilities are fundamentally constrained by its reliance on statistical learning. The project will culminate in a Python-coded simulation that compares human and AI conceptualization through multiple lenses, offering insights into the capabilities and limitations of artificial intelligence in understanding complex, abstract ideas.
The outcomes of this study could contribute to ongoing discussions in AI ethics, philosophy of mind, and cognitive science, particularly regarding the limitations of machine intelligence and the irreplaceable nature of human cognitive processes.