TLDR: A new research paper argues that advanced neural networks, particularly large language models, now exhibit key cognitive abilities like compositionality, productivity, and human-like inductive biases, which were previously considered exclusive to symbolic systems in the human mind. This challenges the long-standing view that human cognition *must* be symbolic at the algorithmic level, suggesting a need for new research to understand how these subsymbolic systems achieve such complex behaviors.
For decades, the debate has raged in artificial intelligence and cognitive science: do human minds operate like symbolic systems, processing information through rules and representations, or are they more akin to the intricate, interconnected web of a neural network? A new research paper titled “Whither symbols in the era of advanced neural networks?” by Thomas L. Griffiths, Brenden M. Lake, R. Thomas McCoy, Ellie Pavlick, and Taylor W. Webb, argues that recent breakthroughs in neural networks are blurring these lines, challenging long-held assumptions about human cognition.
Historically, proponents of symbolic systems pointed to human abilities like combining ideas, generating novel thoughts, and learning quickly as evidence that our minds must be symbolic. They argued that neural networks, with their statistical learning approach, couldn’t replicate these “signatures” of intelligence, particularly productivity and compositionality.
However, the authors contend that the latest generation of neural networks, especially large language models (LLMs) and vision-language models (VLMs), are demonstrating these very capabilities. Thanks to massive increases in model size, vast amounts of training data, and innovative architectures like the Transformer, these AI systems are showing impressive feats of reasoning, learning, and language use.
Compositionality: Building Complex Ideas
Compositionality refers to the ability to flexibly construct complex thoughts from simpler elements. Think about understanding “dogs chase cats” and then immediately understanding “cats chase dogs.” Traditional neural networks struggled with this, leading to “piecemeal cognition.” The paper highlights that modern LLMs can compose concepts that are unlikely to have appeared together in their training data, such as writing a Shakespearean poem about prime numbers. Vision-language models also show this, generating images like “a teddy bear on a skateboard in Times Square.” While not flawless, the capacity for compositionality in neural networks is rapidly improving, often enhanced by a training technique called meta-learning.
Productivity: Generating Novelty
Closely related to compositionality, productivity is the capacity to produce or process structures never encountered before. Humans constantly create new sentences. The paper notes that LLMs routinely generate novel sentences and syntactic structures, with a significant portion of GPT-2 generated sentences having unique syntactic forms. While some limitations exist, particularly with extremely long sequences, advancements like specialized positional embeddings and meta-learning are enabling neural networks to achieve a substantial degree of productivity, even if they sometimes “hallucinate” factual errors.
Inductive Biases: Learning from Limited Data
A major criticism of early neural networks was their data inefficiency – requiring enormous amounts of data compared to humans. This was seen as a lack of human-like “inductive biases,” which are the assumptions that guide a learner to favor certain solutions. The paper argues that pre-trained LLMs, especially those leveraging in-context learning and meta-learning, now exhibit rapid few-shot learning, human-level analogical reasoning, and logical concept learning. This suggests that pre-training can imbue neural networks with powerful, rule-like behaviors that were previously thought to require explicit symbolic mechanisms.
Also Read:
- Unlocking AI’s Potential: The Overlooked Power of Cognitive Imagination
- Advanced Graph Neural Networks Surpass C2 Logic in Expressive Power
A New Research Path
The authors propose a new research agenda for cognitive science, moving beyond the classic “symbols vs. networks” debate. This agenda includes:
- Designing more diagnostic tasks that provide fair comparisons between humans and machines.
- Developing a deeper mechanistic understanding of how neural networks internally represent and process information, using interpretability tools.
- Training neural networks in more developmentally plausible ways, with human-scale inputs, to understand how symbol-like representations emerge.
- Building “cognitive” neural networks that can precisely predict and explain human behavior.
In conclusion, the paper suggests that while symbolic systems remain crucial for characterizing the abstract problems human minds solve (the “computational level”), modern neural networks are increasingly providing a compelling account of how these solutions are approximated (the “algorithmic level”). The impressive capabilities of these subsymbolic systems challenge the historical arguments for symbolic explanations of human cognition. Future research will need to explore new evidence or understand how neural networks might be implementing symbolic systems internally. You can read the full paper here.
