Topic: zero-shot learning

One of the most essential language skills that humans possess is the ability to correctly execute actions based on linguistic instructions. Here we use the latest advances in Natural Language Processing (NLP) to create a neural model of this cognitive ability. Models are trained on a set of 50 psychophysical tasks which require cognitive competencies well studied in the literature. Information about tasks demands is embedded using a transformer architecture pre-trained on an NLP task. Our best performing model, clipNet, uses 33M parameters pre-trained on embedding images and their correspondin

In a recent experiment, Shamash et al. demonstrated that mice spontaneously adopt a subgoal strategy when escaping to a shelter hidden behind a wall. Their optimized trajectory, learned in 1-3 trials, involves running straight to a subgoal located at the edge of the wall, and then to the shelter. Strikingly, even if the wall is removed, animals consistently show curved escape trajectories, revealing that they continue to use the subgoal even though this involves a detour. This behavior can disappear depending on the details of the environment configuration during initial exploration, and can s

Recent studies have demonstrated that brain areas involved in motor skills display common representations of latent trajectories for both stereotypical and generalizable dynamical cortical behavior. However, current research often focuses on analyzing single tasks and individual sessions, missing the opportunity to exploit the structural patterns across datasets. By fitting a unified model to a collection of neural data with shared features, we can increase data usage efficiency and enhance our understanding of common neural computation structures, despite the challenge of distinct populations

To realize goal-directed behavior an executive control mechanism needs to coordinate when and which pieces of working memory (WM) information are operated on in the specific task context. For that, fundamental WM flexibility is desirable so that a given task learned using a specific set of WM items can be immediately performed on novel items too without any extra training (generalization in the spirit of zero-shot learning). Here we apply the computational paradigm of Spatial computing as a candidate mechanism for such generalization in a Delayed Match-To-Sample Task (DMTS), using an attracto

Humans can make an on-the-fly rapid inference when presented with a novel word, using prior knowledge and context. While this is an everyday occurrence necessary for effective communication, the mechanisms of rapid inference and learning during reading are not well understood. One reason for this complexity is that the processes of inference are shrouded by those associated with processing any familiar word. To isolate potential mechanisms of inference alone, we designed a zero-shot learning task where a novel word was presented in a highly constrained context allowing for easy inference of it