580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How do neural population models characterize effective connectivity and dynamics in brain activity?
This research theme investigates mathematical and computational models—especially neural mass, field, and conductance-based models—to represent the activity of neuronal populations and their synaptic connectivity. These models aim to capture the mesoscopic scale of brain dynamics underlying electrophysiological signals (EEG/MEG) and to infer the properties of effective connectivity in distributed brain networks. Their relevance lies in enabling mechanistic interpretations of neural activity and bridging neurobiology with observed data, which is critical for understanding both normal and pathological brain functions.
Key finding: This paper reviews neural mass, neural field, and conductance-based models used within dynamic causal modeling (DCM) for electrophysiological data (e.g., EEG, MEG). It demonstrates how convolution-based neural mass models... Read more
Key finding: This study formulates a normative framework combining efficient coding with generative decoding in neural populations, representing the encoding-decoding interaction as a variational autoencoder. It jointly optimizes stimulus... Read more
Key finding: Introducing a novel model named neurohydrodynamics, the paper extends the Cohen-Grossberg equations governing neural population dynamics by incorporating reaction-diffusion processes inspired by Bohm’s quantum hydrodynamics.... Read more
2. How do connectionist architectures facilitate transfer learning, analogical reasoning, and structural mapping in cognitive tasks?
This theme explores connectionist (neural network) models that learn structural relationships across different but related tasks, enabling transfer of knowledge and analogical inference. The focus is on how weight sharing, shared hidden representations, and multitask learning enable networks to encode identical or analogous elements, accelerating learning and generalization in novel tasks. These models provide computational insights into human cognitive functions involving analogy, transfer learning, and schema formation, essential for understanding complex learning and reasoning.
Key finding: Through simulations involving feedforward networks with shared hidden layer weights, the paper demonstrates that networks trained on superficially different but structurally analogous tasks develop parallel hidden unit... Read more
Key finding: This article critically reviews transfer learning and analogical inference within machine learning and cognitive science. It highlights how structural relational representations support knowledge transfer across tasks and... Read more
Key finding: The paper presents a connectionist planning schema that performs backchaining by exploiting episodic memory traces represented as Precondition-Action-Consequences (PAC) triples. The model iteratively recalls events whose... Read more
3. What roles do connectionist models play in linguistic processing, concept representation, and language acquisition?
This theme focuses on connectionist approaches to modeling language phenomena, including morphological acquisition, rule learning, lexical representation, linguistic relativity, and cognitive semantics. Such models challenge traditional symbolic frameworks by demonstrating emergent rule-like behavior, graded representations, and integrated learning of lexical and grammatical forms. The research investigates how distributed representations and learning dynamics underlie concept formation, word acquisition, and the interplay between linguistic and non-linguistic cognition, contributing to theories of the mental lexicon and language processing architectures.
Key finding: This study develops a modular connectionist network combining a connectionist short-term memory (STM) for symbolic rule processing with an associative memory-based lexicon to model the German verb paradigm. The model... Read more
Key finding: Using the Playpen connectionist model, the paper shows that the interaction between linguistic and non-linguistic categories depends on the correlation patterns in the world and how they relate to linguistic categories... Read more
Key finding: The paper proposes a geometric framework for mental processes and concept representation grounded in intermediate-level models linking first-person mental states to neural dynamics. Concepts are represented as geometric... Read more
Key finding: Arguing against modular, symbolist frameworks, this chapter advocates for a distributed, integrative model of the mental lexicon grounded in psycho-computational principles where words emerge from interactions between general... Read more
Key finding: Through connectionist simulations, the paper demonstrates that both case marking and strict word order facilitate the learnability of syntactic mappings by sequential learning devices. Networks trained on typologically common... Read more
4. How do deep learning models compare to biological neural systems in vision and cognition, and what are their limitations?
This theme critically examines the relationship and differences between deep artificial neural networks and biological neural systems, especially in vision. It explores the extent to which neural network architectures and operational principles correspond to brain mechanisms, challenges the assumption that deep models fully capture human cognition, and discusses the implications for explainability and neuroscientific relevance. It highlights fundamental mismatches, such as differing physical substrates, representational abstractions, and algorithmic processes, and the ensuing challenges in interpreting deep models as cognitive or neural analogs.
Key finding: The article argues that while inspiration from the brain historically influenced artificial neural network designs, recent deep learning models have diverged significantly in structure and algorithmic mechanisms from... Read more
Key finding: By reviewing empirical psychological findings, the paper shows that despite impressive benchmark image classification performance, deep neural networks (DNNs) fail to replicate core aspects of human vision and cognition. DNNs... Read more