Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Visual and Goal Vector Signals Interact to Shape Behavior and Spatial Representations

Sandhiya Vijayabaskaran¹, Sen Cheng²; ¹Ruhr-Universität Bochum, ²Ruhr-Universtät Bochum

Presenter: Sandhiya Vijayabaskaran

Spatial navigation relies on a variety of signals across different sensory modalities to guide movement towards a goal. While these signals can sometimes be redundant, they are crucial in the face of uncertainty, where navigating agents may have to switch between these signals or integrate them to determine the moving direction. We model these interactions in a deep reinforcement learning agent that uses two signals, vision and goal-vectors, to navigate. By analysing the agent's behavior and spatial representations, we show that it can successfully navigate using each signal independently or by integrating both. We show that this flexibility enables the agent to successfully cope with changing environments or with signals becoming contaminated with noise. Interestingly, our model also highlights a trade-off --- when integration is unnecessary, such as in an unchanging environment, relying on a single stable signal improves navigation. We use this insight to explain counterintuitive experimental results. Additionally, we show that the place-cell-like spatial representations emerging in the network are shaped by both signals, albeit to varying degrees.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF