Frank van der Velde

Frank van der Velde

This article presents a neural language architecture of sentence structure in the brain, in which sentences are temporal connection paths that interconnect neural structures underlying their words. Words remain 'in-situ', hence they are always content-addressable. Arbitrary and novel sentences (with novel words) can be created with 'neural blackboards' for words and sentences. Hence, the unlimited productivity of natural language can be achieved with a 'fixed' small world like network structure. The article focuses on the neural blackboard for sentences. The architecture uses only one 'connection matrix' for binding all structural relations between words in sentences. Its ability to represent arbitrary (English) sentences is discussed in detail, based on a comprehensive analysis of them. The architecture simulates intra-cranial brain activity observed during sentence processing and fMRI observations related to sentence complexity and ambiguity. The simulations indicate that the observed effects relate to global control over the architecture, not to the sentence structures involved, which predicts higher activity differences related to complexity and ambiguity with higher comprehension capacity. Other aspects discussed are the 'intrinsic' sentence structures provided by connection paths and their relation to scope and inflection, the use of a dependency parser for control of binding, long-distance dependencies and gaps, question answering, ambiguity resolution based on backward processing without explicit backtracking, garden paths, and performance difficulties related to embeddings.

Frank van der Velde

Recently, a number of articles have argued that deep learning models such as GPT could also capture key aspects of language processing in the human mind and brain. However, I will argue that these models are not suitable as neural models of human language. Firstly, because they fail on fundamental boundary conditions, such as the amount of learning they require. This would in fact imply that the mechanisms of GPT and brain language processing are fundamentally different. Secondly, because they do not possess the logistics of access needed for compositional and productive human language processing. Neural architectures could possess logistics of access based on small-world like network structures, in which processing does not consist of symbol manipulation but of controlling the flow of activation. In this view, two complementary approaches would be needed to investigate the relation between brain and cognition. Investigating learning methods could reveal how 'learned cognition' as found in deep learning could develop in the brain. However, neural architectures with logistics of access should also be developed to account for 'productive cognition' as required for natural or artificial human language processing. Later on, these approaches could perhaps be combined to see how such architectures could develop by learning and development from a simpler basis.