A5045424407- Motor Control and Adaptation
- Neural dynamics and brain function
- Action Observation and Synchronization
- Visual perception and processing mechanisms
- Ethics and Social Impacts of AI
- Muscle activation and electromyography studies
- Explainable Artificial Intelligence (XAI)
- Data Visualization and Analytics
- Face Recognition and Perception
- Neural Networks and Applications
- Child and Animal Learning Development
- Memory and Neural Mechanisms
- Natural Language Processing Techniques
- Topic Modeling
- Statistical Methods and Inference
- Sleep and Wakefulness Research
- Cognitive and developmental aspects of mathematical skills
- Sport Psychology and Performance
- Children's Physical and Motor Development
- Multimodal Machine Learning Applications
- Air Quality Monitoring and Forecasting
- Gaussian Processes and Bayesian Inference
- Psychology of Moral and Emotional Judgment
- Misinformation and Its Impacts
- Advanced Neural Network Applications
Google (United Kingdom)
2023-2024
DeepMind (United Kingdom)
2022-2024
University of Oxford
2019-2024
University of Cambridge
2016-2019
Recent work in large language modeling (LLMs) has used fine-tuning to align outputs with the preferences of a prototypical user. This assumes that human are static and homogeneous across individuals, so aligning single "generic" user will confer more general alignment. Here, we embrace heterogeneity consider different challenge: how might machine help people diverse views find agreement? We fine-tune 70 billion parameter LLM generate statements maximize expected approval for group...
Finding agreement through a free exchange of views is often difficult. Collective deliberation can be slow, difficult to scale, and unequally attentive different voices. In this study, we trained an artificial intelligence (AI) mediate human deliberation. Using participants' personal opinions critiques, the AI mediator iteratively generates refines statements that express common ground among group on social or political issues. Participants (N = 5734) preferred AI-generated those written by...
<h2>Summary</h2> A prerequisite for intelligent behavior is to understand how stimuli are related and generalize this knowledge across contexts. Generalization can be challenging when relational patterns shared contexts but exist on different physical scales. Here, we studied neural representations in humans recurrent networks performing a magnitude comparison task, which it was advantageous concepts of "more" or "less" between Using multivariate analysis human brain signals network hidden...
reasoning is a key ability for an intelligent system. Large language models (LMs) achieve above-chance performance on abstract tasks but exhibit many imperfections. However, human also imperfect. Human affected by our real-world knowledge and beliefs, shows notable "content effects"; humans reason more reliably when the semantic content of problem supports correct logical inferences. These content-entangled patterns are central to debates about fundamental nature intelligence. Here, we...
Humans can navigate flexibly to meet their goals. Here, we asked how the neural representation of allocentric space is distorted by goal-directed behavior. Participants navigated an agent two successive goal locations in a grid world environment comprising four interlinked rooms, with contextual cue indicating conditional dependence one location on another. Examining geometry which room and context were encoded fMRI signals, found that map-like representations emerged both hippocampus...
Motor imagery, that is the mental rehearsal of a motor skill, can lead to improvements when performing same skill. Here we show powerful and complementary role, in which imagery different movements after actually skill allows learning not possible without imagery. We leverage well-studied task subjects reach presence dynamic (force-field) perturbation. When two opposing perturbations are presented alternately for physical movement, there substantial interference, preventing any learning....
Recent work has shown the potential benefit of selective prediction systems that can learn to defer a human when predictions AI are unreliable, particularly improve reliability in high-stakes applications like healthcare or conservation. However, most prior assumes behavior remains unchanged they solve task as part human-AI team opposed by themselves. We show this is not case performing experiments quantify interaction context prediction. In particular, we study impact communicating...
Real-world data is high-dimensional: a book, image, or musical performance can easily contain hundreds of thousands elements even after compression. However, the most commonly used autoregressive models, Transformers, are prohibitively expensive to scale number inputs and layers needed capture this long-range structure. We develop Perceiver AR, an autoregressive, modality-agnostic architecture which uses cross-attention map small latents while also maintaining end-to-end causal masking. AR...
Abstract Humans can navigate flexibly to meet their goals. Here, we asked how the neural representation of allocentric space is distorted by goal-directed behaviour. Participants navigated an agent two successive goal locations in a grid world environment comprising four interlinked rooms, with contextual cue indicating conditional dependence one location on another. Examining geometry which room and context were encoded fMRI signals, found that map-like representations emerged both...
To understand a visual scene, observers need to both recognize objects and encode relational structure. For example, scene comprising three apples requires the observer concepts of "apple" "three." In primate brain, these functions rely on dual (ventral dorsal) processing streams. Object recognition in primates has been successfully modeled with deep neural networks, but how structure (including numerosity) is encoded remains poorly understood. Here, we built learning model, based...
Abstract After extended practice, motor adaptation reaches a limit in which learning appears to stop, despite the fact that residual errors persist. What prevents brain from eliminating errors? Here we found was causally dependent on second order statistics of perturbation; when variance high, impaired and large persisted. However, relied solely explicit strategy, both its dependence perturbation variability disappeared. In contrast, depended entirely, or part implicit learning, developed....
We propose a theory of structure learning in the primate brain. argue that parietal cortex is critical for about relations among objects and categories populate visual scene. suggest current deep models exhibit poor global scene understanding because they fail to perform relational inferences occur dorsal stream. review studies neural coding posterior (PPC), drawing conclusion this brain area represents potentially high-dimensional inputs on low-dimensional manifold encodes relative position...
Knowledge about a tool's dynamics can be acquired from the visual configuration of tool and through physical interaction. Here, we examine how information affects generalization dynamic learning during use. Subjects rotated virtual hammer-like object while varied separately for two rotational directions. This allowed us to quantify coupling adaptation between directions, that is, transferred one direction other. Two groups experienced same object. For group, object's was displayed, other,...
Deep neural networks have provided a computational framework for understanding object recognition, grounded in the neurophysiology of primate ventral stream, but fail to account how we process relational aspects scene. For example, deep at problems that involve enumerating number elements an array, problem humans relies on parietal cortex. Here, build 'dual-stream' network model which, equipped with both dorsal and streams, can generalise its counting ability wholly novel items ('zero-shot'...
Abstract Motor imagery, that is the mental rehearsal of a motor skill, can lead to improvements when performing same skill. Here we show powerful and complementary role, in which imagery movements after actually skill allows learning not possible without imagery. We leverage well-studied task subjects reach presence dynamic (force-field) perturbation. When two opposing perturbations are presented alternately for physical movement, there substantial interference, preventing any learning....
Summary A prerequisite for intelligent behaviour is to understand how stimuli are related and generalise this knowledge across contexts. Generalisation can be challenging when relational patterns shared contexts but exist on different physical scales. Here, we studied neural representations in humans recurrent networks performing a magnitude comparison task, which it was advantageous concepts of “more” or “less” between Using multivariate analysis human brain signals network hidden unit...
Recent work has shown the potential benefit of selective prediction systems that can learn to defer a human when predictions AI are unreliable, particularly improve reliability in high-stakes applications like healthcare or conservation. However, most prior assumes behavior remains unchanged they solve task as part human-AI team opposed by themselves. We show this is not case performing experiments quantify interaction context prediction. In particular, we study impact communicating...