A new model for AI
Related Articles
Two types of motifs enhance human recall and generalization of long sequences
Whether it is listening to a piece of music, learning a new language, or solving a mathematical equation, people often acquire abstract notions in the sense of motifs and variables—manifested in musical themes, grammatical categories, or mathematical symbols. How do we create abstract representations of sequences? Are these abstract representations useful for memory recall? In addition to learning transition probabilities, chunking, and tracking ordinal positions, we propose that humans also use abstractions to arrive at efficient representations of sequences. We propose and study two abstraction categories: projectional motifs and variable motifs. Projectional motifs find a common theme underlying distinct sequence instances. Variable motifs contain symbols representing sequence entities that can change. In two sequence recall experiments, we train participants to remember sequences with projectional and variable motifs, respectively, and examine whether motif training benefits the recall of novel sequences sharing the same motif. Our result suggests that training projectional and variables motifs improve transfer recall accuracy, relative to control groups. We show that a model that chunks sequences in an abstract motif space may learn and transfer more efficiently, compared to models that learn chunks or associations on a superficial level. Our study suggests that humans construct efficient sequential memory representations according to the two types of abstraction we propose, and creating these abstractions benefits learning and out-of-distribution generalization. Our study paves the way for a deeper understanding of human abstraction learning and generalization.
Segment Anything for Microscopy
Accurate segmentation of objects in microscopy images remains a bottleneck for many researchers despite the number of tools developed for this purpose. Here, we present Segment Anything for Microscopy (μSAM), a tool for segmentation and tracking in multidimensional microscopy data. It is based on Segment Anything, a vision foundation model for image segmentation. We extend it by fine-tuning generalist models for light and electron microscopy that clearly improve segmentation quality for a wide range of imaging conditions. We also implement interactive and automatic segmentation in a napari plugin that can speed up diverse segmentation tasks and provides a unified solution for microscopy annotation across different microscopy modalities. Our work constitutes the application of vision foundation models in microscopy, laying the groundwork for solving image analysis tasks in this domain with a small set of powerful deep learning models.
Constructing future behavior in the hippocampal formation through composition and replay
The hippocampus is critical for memory, imagination and constructive reasoning. Recent models have suggested that its neuronal responses can be well explained by state spaces that model the transitions between experiences. Here we use simulations and hippocampal recordings to reconcile these views. We show that if state spaces are constructed compositionally from existing building blocks, or primitives, hippocampal responses can be interpreted as compositional memories, binding these primitives together. Critically, this enables agents to behave optimally in new environments with no new learning, inferring behavior directly from the composition. We predict a role for hippocampal replay in building and consolidating these compositional memories. We test these predictions in two datasets by showing that replay events from newly discovered landmarks induce and strengthen new remote firing fields. When the landmark is moved, replay builds a new firing field at the same vector to the new location. Together, these findings provide a framework for reasoning about compositional memories and demonstrate that such memories are formed in hippocampal replay.
Evolutionary optimization of model merging recipes
Large language models (LLMs) have become increasingly capable, but their development often requires substantial computational resources. Although model merging has emerged as a cost-effective promising approach for creating new models by combining existing ones, it currently relies on human intuition and domain knowledge, limiting its potential. Here we propose an evolutionary approach that overcomes this limitation by automatically discovering effective combinations of diverse open-source models, harnessing their collective intelligence without requiring extensive additional training data or compute. Our approach operates in both parameter space and data flow space, allowing optimization beyond just the weights of the individual models. This approach even facilitates cross-domain merging, generating models such as a Japanese LLM with math reasoning capabilities. Surprisingly, our Japanese math LLM achieved state-of-the-art performance on a variety of established Japanese LLM benchmarks, even surpassing models with substantially more parameters, despite not being explicitly trained for such tasks. Furthermore, a culturally aware Japanese vision–language model generated through our approach demonstrates its effectiveness in describing Japanese culture-specific content, outperforming previous Japanese vision–language models. This work not only contributes new state-of-the-art models back to the open-source community but also introduces a new paradigm for automated model composition, paving the way for exploring alternative, efficient approaches to foundation model development.
Bayesian p-curve mixture models as a tool to dissociate effect size and effect prevalence
Much research in the behavioral sciences aims to characterize the “typical” person. A statistically significant group-averaged effect size is often interpreted as evidence that the typical person shows an effect, but that is only true under certain distributional assumptions for which explicit evidence is rarely presented. Mean effect size varies with both within-participant effect size and population prevalence (proportion of population showing effect). Few studies consider how prevalence affects mean effect size estimates and existing estimators of prevalence are, conversely, confounded by uncertainty about effect size. We introduce a widely applicable Bayesian method, the p-curve mixture model, that jointly estimates prevalence and effect size by probabilistically clustering participant-level data based on their likelihood under a null distribution. Our approach, for which we provide a software tool, outperforms existing prevalence estimation methods when effect size is uncertain and is sensitive to differences in prevalence or effect size across groups or conditions.
Responses