Research
Continual Learning
The Continual Learning Working Group spearheaded our efforts to address core limitations of modern vision and language models, focusing on integrating biologically and cognitively inspired memory mechanisms to enable lifelong learning without catastrophic forgetting. Key directions included managing repeated specialization within a single architecture, developing auxiliary memory modules for retaining prior knowledge while incorporating new data, and exploring mechanisms such as episodic memory and cortical consolidation to improve adaptability beyond conventional approaches. In parallel, the group developed meta-learning strategies for efficient task adaptation and benchmark suites that better capture the complexity of real-world continual learning scenarios.
These efforts converged in the project Analysis and Metrics for Continual Learning and Continual Meta-Learning. A total of four continual learning projects were funded in Year 3, exploring evaluation, biological mechanisms, environmental structure, and embodied systems: the flagship project introduces new metrics for adaptive generalization, while complementary projects study neural mechanisms of stable learning, characterize adaptation in dynamic environments, and extend these principles to resource-constrained, embodied AI. Collectively, this portfolio reframes learning as a continuous, adaptive process grounded in both artificial and natural intelligence.
Biological Learning Rules
The Biological Learning working group has spearheaded discussions to define benchmarks, align research directions, and catalyze collaborative projects in biologically plausible learning. The group focused on multimodal unsupervised representation learning without backpropagation, exploring bio-inspired algorithms such as Hebbian learning, predictive coding, and feedback alignment, supported by shared datasets and evaluation metrics. A key outcome is the development of a Hyper-Modal Representation Learning Benchmark, a community-driven testbed for evaluating biologically inspired learning across multimodal, large-scale, and agentic settings. The overall goal is to advance generalizable representations grounded in how humans learn from multimodal experience, establishing a coordinated ARNI-wide effort for biologically grounded AI. We funded two complementary projects, one focused on biologically plausible sequence learning architectures and the other on theoretical principles of learning and decision-making. These well-integrated efforts advance ARNI’s mission to develop generalizable, memory-driven, and flexible AI grounded in biological intelligence.
Foundation Models for Neuroscience
ARNI’s work on Frontier Models for Neuroscience continues to focus on developing scalable, integrative modeling frameworks that unify neural activity, behavior, and environment. Our community worked on a coordinated set of flagship efforts centered on neurofoundation models, establishing a new paradigm for modeling brain systems at scale. Led by the Frontier Models for Neuroscience & Behavior working group, these projects combine large-scale data integration, multimodal representation learning, and standardized evaluation to enable generalizable, task-flexible models of neural computation. At the core of this effort is OmniMouse, a multimodal, multi-task foundation model trained on one of the largest neural datasets to date, which establishes state-of-the-art performance across prediction, decoding, and forecasting tasks while revealing a data-limited scaling regime. Complementary projects extend this foundation through cognitively inspired architectures and theoretical frameworks for valuing neural data, together defining the principles, infrastructure, and benchmarks for neurofoundation models. These efforts position ARNI as an emerging leader in establishing the technical and conceptual foundations for next-generation AI systems grounded in brain and behavior.
Language & Vision
The Vision and Language thrust advances ARNI’s mission by integrating perception, language, and interaction to develop AI systems that are both human-centered and grounded in real-world complexity. ARNI funded four projects that span foundational studies of neural language processing, multimodal interaction systems, safe and interpretable dialogue, and robust perception under data constraints, combining insights from cognitive science, neuroscience, and AI. This year, the Vision and Language Working Group, is bringing together faculty and trainees across ARNI to identify shared challenges in multimodal representation and foster deeper interdisciplinary collaboration. A central focus is advancing use-inspired research that addresses societal needs, particularly in assistive technologies. This is exemplified by the flagship collaborative project on interactive audio-haptic perception for blind and low-vision users, with plans to expand similar efforts in Years 4–5. These projects establish a unified framework for vision-language systems that are not only more capable, but also more interpretable, accessible, and deployable in real-world settings. In response to previous site visit feedback on limited cross-thrust integration for the Vision and Language thrust, ARNI’s continual learning group in collaboration with vision and language faculty spearheaded new efforts around shared challenges that culminated to the the Continual Learning flagship project, which integrates biologically inspired memory mechanisms into LLMs to enable continual adaptation and lifelong learning.
Multi-resource-cost Optimization of Neural Network Models
Neural network models are typically set up with a fixed architecture that defines the number of nodes and the connectivity, and are unrolled for a fixed number of timesteps to obtain a computational graph for backpropagation. This amounts to fixing the resources that a physical implementation in a biological brain or dedicated engineered system would require in terms of space (to accommodate nodes and connections), time (to execute the steps), and energy. The fixed architecture of neural network models allows us to limit the resource requirements and discover what level of performance is possible through optimization. However, it makes it difficult to explore the tradeoffs between the multiple resources. For example, would a smaller network that runs for more timesteps give preferable results according to a joint cost of nodes, connections, time, energy, and error? It would be useful to be able to flexibly trade off resources against each other and against task performance as part of the optimization of a single model, rather than having to train many models (each with a fixed vector of costs) to explore the space of solutions. We will develop (1) ways to quantify space, time, and energy costs of neural network models and (2) differentiable objectives that enable efficient joint minimization of the costs of multiple resources. Such methods could help us understand biological neural mechanisms that emerge from particular profiles of resource costs and behavioral affordances and also to engineer more efficient AI for resource-limited devices.
Neuromechanisms of Intelligence
The Neural Mechanisms of Intelligence thrust investigates the computational principles underlying biological intelligence and translates them into new paradigms for AI. ARNI supports a diverse portfolio of projects spanning connectome-informed architectures, circuit dynamics, modular learning, and embodied control, providing mechanistic and theoretical insights that complement our work on foundation models. These projects move beyond prediction to uncover how structure, dynamics, and constraints in neural systems give rise to efficient, flexible, and generalizable computation.
Across scales, from synaptic connectivity and circuit organization to behavior and embodiment, this work identifies key principles such as sparse and structured representations, modularity, hierarchical organization, and resource-efficient computation. By grounding AI development in experimentally validated models of neural systems and biologically plausible mechanisms, this thrust establishes a principled framework for understanding intelligence as an emergent property of structured, adaptive, and embodied systems.
