Artificial Intelligence

Tal Linzen, an associate professor of linguistics and data science at New York University and a research scientist at Google, will speak in MIT's Siegel Family Quest for Intelligence seminar series. His work combines behavioral experiments and computational methods to study language learning and comprehension, alongside large-language-model post-training, evaluation, and interpretability.

ISTA Lecture by Klaus Robert Müller (TU Berlin & Korea University, Seoul) on how machine learning and AI enable scientific research, particularly in medicine and chemistry, and on explainability techniques for extracting understanding from machine learning models.

Stanford's annual Precision Mental Health Symposium explores how neuroscience, psychiatry, data science, and artificial intelligence can translate brain research into real-world mental-health care. Themes include circuits and cognition, biomarkers and neuroimaging, rapid-acting interventions, neuromodulation, and precision therapeutics.

The MIT Aging Brain Initiative brings together work in molecular imaging, cognition, chemistry, bioengineering, neurodegeneration, and artificial intelligence to address brain aging and Alzheimer's prevention.

Useful internal representations should explain the patterns of regularities and dependencies among observations. Probabilistic graphical models promise a principled way to uncover latent factors as such, but they are hard to scale to handle high-dimensional sensory observations and complicated dependencies structures. Neural-networks, on the other hand, excel at approximating complicated high-dimensional functions, but their internal representations do not easily lend themselves to a probabilistic interpretation. Despite some successes, a general unified approach is still missing for integrating the two approaches. I will describe a novel approach towards merging adaptive neural-network components into a probabilistic framework, based on three core ideas. The first is to train a set of networks to collectively perform inference, leveraging the ability of pattern-recognition methods to amortise complicated transformations. The second is to constrain the way in which the outputs of these networks are interpreted, transformed, and combined together. These constraints, together with the learning objective itself, are derived directly from probabilistic considerations encoded in a graphical model. Finally, the third core idea is that of recognition-parametrisation, allowing the inference ("recognition") procedure to directly define the model itself, without requiring an explicit "generative" decoder. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-03-04. Recording duration: 00:48:26.

Dense Associative Memories (Dense AMs) are energy-based neural networks that share many desirable features of celebrated Hopfield Networks but have superior information storage capabilities. 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 only passive house-keeping support structures in the brain. In contrast, it suggests that astrocytes might be actively involved in brain computation and memory storage and retrieval. This story is an illustration of how computational principles originating in physics may provide insights into novel AI architectures and brain computation. VVTNS New Year Opening Lecture. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2025-01-08. Recording duration: 00:49:16.

NSF postdoctoral fellowships supporting early-career scientists who combine artificial intelligence with biological research and training to advance biotechnology innovation.

EMBL's ARISE2 programme offers more than 20 fully funded, three-year postdoctoral fellowships for scientists developing research-infrastructure technologies across the life sciences, including genomics, structural biology, imaging, cell and tissue research, environmental research, translation, and data science or artificial intelligence.

NSF support for interdisciplinary collaborations developing mathematical and theoretical foundations for explainable, reliable, sustainable and trustworthy artificial intelligence.

UKRI and EPSRC are funding speculative, high-risk fundamental research with the potential to produce a step change in the explainability of future AI systems. The call welcomes work on mechanistic interpretability, output reasoning, process examination, imposed explainability, uncertainty quantification, and novel approaches beyond those examples.

EMBL's ARISE2 programme offers more than 20 fully funded, three-year postdoctoral fellowships for scientists developing research-infrastructure technologies across the life sciences, including genomics, structural biology, imaging, cell and tissue research, environmental research, translation, and data science or artificial intelligence.

Develop an AI-assisted digital platform for stroke rehabilitation at the Singapore-ETH Centre. The project combines a mobile app and conversational agent with a soft robotic hand orthosis, supporting rehabilitation adherence at home. Work includes gathering patient and clinician requirements, building and validating the app, integrating rehabilitation hardware, and evaluating the intervention in a six-week clinical study with NTU and Tan Tock Seng Hospital. This is a full-time, fixed-term PhD position. The current vacancy accepts online applications without a stated closing date; submit a cover letter, CV with two references, and university transcripts. The employer offers up to two days of home working per week.

Deep networks are often thought of as black boxes. Their ability to encompass vast swathes of knowledge indeed makes them hard to explain. Yet many of their behaviours — generalization under overparametrization, grokking, OOD failures, neural scaling laws — recur across architectures and scales, and each, however surprising, can be reproduced and explained in controlled settings. I will review these efforts to identify and explain the universal phenomena of deep learning, and suggest that an LLM may amount to a sum of such tractable sub-phenomena, interpolative in nature. Finally, leaving scientific rigor aside, I'll argue that what separates this prosaic picture from the apparent magic of LLMs may well be the industrial scale of compute and human labour behind it, and that AGI in its deeper extrapolative sense may be much further away than claimed.

Dense Associative Memories (Dense AMs) are energy-based neural networks that share many desirable features of celebrated Hopfield Networks but have superior information storage capabilities. 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 only passive house-keeping support structures in the brain. In contrast, it suggests that astrocytes might be actively involved in brain computation and memory storage and retrieval. This story is an illustration of how computational principles originating in physics may provide insights into novel AI architectures and brain computation. VVTNS New Year Opening Lecture. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2025-01-08. Recording duration: 00:49:16.