📌 Poster Session:
⏰ When: TODAY, Thu, Dec 12, 4:30 p.m. – 7:30 p.m. PST
📍 Where: East Exhibit Hall A-C, #3705
📄 What: Geometry of Naturalistic Object Representations in Recurrent Neural Network Models of Working Memory

Hope to see you there!
@bashivan.bsky.social @takuito.bsky.social

👉 Check it out: arxiv.org/abs/2411.02685
📅 We’ll be at NeurIPS! Join us for our poster presentation on Thu 12 Dec, 7:30 p.m. EST — 10:30 p.m. EST.

#AI #CognitiveScience #WorkingMemory #DeepLearning #RepresentationGeometry #MultiTask

Our findings bridge cognitive science & AI, revealing how high-dimensional object information is encoded, retained, and recalled in recurrent models of working memory.

🎯 With training, RNNs implemented chronological memory subspaces allowing them to track object information using rotational dynamics—supporting resource-based models of working memory.

📐 Surprisingly, object features are less orthogonalized in RNN representations compared to perceptual space.

🧠 We found that multi-task RNNs (unlike single-task ones) retain both task-relevant & irrelevant info but reusable representations only emerged in simple gateless architectures.

🖥️ To answer this question, we trained multi-task RNNs (vanilla, GRU, LSTM) on 9 N-back tasks using naturalistic 3D object stimuli to study encoding, retention, & retrieval dynamics.

It’s unclear how high-dimensional naturalistic sensory information is encoded, retained and recalled in these models to accommodate various task demands.

Previous models of working memory have mainly focused on using abstract stimuli (Mante et al., 2013, Yang et al., 2019, Driscoll et al., 2024, Fascianelli et al., 2024, Piwek & Stokes, 2023 etc)