Neuroscience-AI 2025 Symposium

The fields of neuroscience and AI have a long and intertwined history. From the study of simple and complex cells in visual areas of the brain to the success of deep neural networks in many real-world applications, experimental and theoretical neuroscience has contributed significantly to designing smarter machines. In turn, AI models help us better understand brain computations that underlie biological intelligence. In my talk, I will present several efforts to decipher brain function by building computational models and quantifying model behaviors with human benchmarks in several cognitive tasks, such as visual search, context reasoning, continual learning, working memory, and generalization. I will also discuss how these models not only help us understand the neural mechanisms driving human cognition but also inspire the development of more robust AI systems that replicate human-like decision-making, attention, reasoning, learning, and memory. Through these interdisciplinary efforts, we can pave the way for more intelligent, robust, and adaptable AI systems that mirror the complexities of biological intelligence.

In the last decade or so neuroscience has gone from predominantly single neuron recordings to large scale recordings of up to a million neurons. This transformation in data collection allows us to question more broadly what is the nature of the population code of large numbers of neurons and how these relate to behavior. Linear methods have shown poor predictive power. Common dimensionality reduction approaches fail to provide an understandable representation of the population code. The most common method is almost certainly, principal component analysis (PCA). However, to our knowledge up until now, existing dimensionality reduction methods produce latent variables that are experimentally not testable as the latent components do not have a direct correspondence to either brain areas or neurons. Here in the following work, we develop a couple novel dimensionality reduction algorithms that originate from a dynamical systems theory framework, the generalized Takens theorem, and its application for causal inference, the convergent cross-mapping (CCM) algorithm, to produce a predictive data driven geometric models based on low dimensional manifolds that generically map sensory input to brain activity to behavior without latent variables. Here every variable corresponds to either a single or a population of identifiable neurons or brain area. The dimensionally reduced representation is a manifold mapping that can predict future behaviors of the animal based on neural activity and is sensitive to sensory input where each orthogonal axis in the ambient space of the manifold corresponds to an observable neuron, population of neurons, or brain area. As such it allows for experimental verification as it produces falsifiable predictions of the contribution of candidate brain components that contribute to behavior. In addition, the direct 1:1 mapping of observables to orthogonal axes provides explainability for the relative contributions of each population of neurons in a state dependent manner. We name these first of these methods Causal Compression as it is grounded in CCM causal inference as its key element for the identification of orthogonal bases for attractor reconstruction. In the following example we show examples from Drosophila whole brain lightfield microscopy recordings, rat tetrodes recordings from 3 brain areas as well as whole brain human fMRI. We show that one can map brain activity to motor output using a manifold of surprisingly low dimensionality to predict future behaviors of drosophila, rats and humans. Manifolds discovered by causal compression can also be used to simulate large systems such as the brain if made into a network of manifolds. Here our method Generative manifold networks (GMN) generates a simulation of realistic whole brain activity of Drosophila as well as the motor output of the animal. We show that is scalable to single neuron resolution for over 100,000 neurons of a larval zebrafish.

Foundation models offer the potential to transform many aspects of our society, and are already impacting areas such as software programming and generative art. In this talk, I will describe progress by my group and our collaborators in developing foundation-model enabled frameworks for scientific domains, including in neuroscience. Some tasks require developing so-called "LLM Agents" that can automatically reason over complex workflows, such as data analysis or hypothesis search. Other tasks require developing novel foundation or foundation-like models, such as for processing neural signals or behavioral videos.

Despite recent rapid advances in large generative models (such as large language models, image/video diffusion models, etc.), these models still fall short in terms of being reliable, generalizable and efficient. In that regard, neuroscience and psychology can be natural sources of insights, as "human-level intelligence" is still widely considered to be a gold standard for a number of aspects. In this talk, I will present some early results demonstrating the usefulness of approaches to AI inspired by neuroscience and psychology, along with ideas for future work.

Last century's AI was inspired by last century's neuroscience and psychology. This century's AI should be based on 21st century neuroscience and psychology. A potentially paradigm-changing view of “21st century” neuroscience and psychology presented by a group of University of Virginia scientists based on concrete new scientific data is reviewed. Novel phenomena such as near-death experiences (NDEs), published in established scientific journals such as The Lancet, present major challenges to the current view of the relationship between the mind and brain. They particularly challenge the current computational paradigm in AI which attempts to replicate the functioning of the brain, and the computational paradigm in cognitive science in general which attempts to explain how mental phenomena are generated by the brain’s operations. In this talk, we will use the work of David Marr, who was both a neuroscientist and computer vision pioneer, to illustrate the fundamental issues arising from this new view of the relationship between the mind and brain.

Dense Associative Memories (Dense AMs) are energy-based networks deeply rooted in the ideas of statistical physics and Ising-like spin models. In contrast to conventional Hopfield Networks, which were popular in the 1980s, DenseAMs have a very large memory storage capacity - possibly exponential in the size of the network. This aspect makes them appealing tools for many problems in AI and neurobiology. In this talk I will describe two theories of how DenseAMs might be built in biological “hardware”. According to the first theory, DenseAMs arise as effective theories after integrating out a large number of neuronal degrees of freedom. According to the second theory, astrocytes, a particular type of glia cells serve as core computational units enabling large memory storage capabilities. This second theory challenges a common point of view in the neuroscience community that astrocytes play the role of passive house-keeping support structures in the brain. In contrast, it suggests that astrocytes are actively involved in brain computation and memory storage and retrieval.

How does the brain represent the world around us? In recent years, the unprecedented ability of deep neural networks (DNNs) to predict neural responses in the human visual cortex has led to a lot of excitement about these models potential to capture human visual perception. However, studies demonstrating representational alignment of visual DNNs with humans typically use brain responses to static, isolated objects, while real-life visual perception requires processing of complex, dynamic scenes. In this talk, I will discuss work aiming to move the field beyond object recognition by focusing on scene perception, video perception, and data-driven alignment via brain-guided image generation.

The Florence project is designing communication technology to assist people living with dementia. Florence comprises a personalized knowledge bank, which supports an ecosystem of communication devices designed to enhance quality of life at home. What makes Florence different is the participatory design process with a team of living experience experts – people living with dementia – which is producing an integrated system that builds on people’s strengths with hardware, software, tutoring and data custodianship built into every component. Effective AI in this human-centred technology is an enabler of people’s everyday goals, serving as a calm computing element that contributes to quality of life.