Cartoon me on the ferris wheel, waving goodbye and having a good time :)

Overall, we hope that our paper sheds some light on how we combine strategies for flexible navigation, and how different representations in the brain may support different strategies. + see paper for some additional results about how humans select landmarks! (6/6)

Plots showing that units are differentially correlated with the ikelihood of using ‘direction’ or ‘state’ actions, and lesioning these units differentially affects agents’ use of directions vs. states. Plot showing that ‘vector’ units are more like to have spatial activation patterns and ‘transition’ units are more likely to be modulated by landmarks.

We also identify ‘modules’ that appear causally important for implementing each strategy. These modules represent the environment differently: for eg, ‘vector’ units are more likely to have stable spatial representations, while ‘transition’ units are more likely to carry landmark information (5/6)

Activation patterns of example units as networks navigated the grid environment. Some units are responsive to location of landmarks, stable regions of space across environments, or a conjunction of both.

We examined the representations learnt by networks for hints about how these strategies might be implemented in the brain. We find units that represent the environment in different ways - resembling diverse spatial representations observed in mammalian navigation systems. (4/6)

Plots showing both humans and models perform best when they can freely arbitrate between strategies, and prefer to use ‘states’ at landmarks and other learnt locations.

We find that humans did best when they could freely arbitrate between strategies - preferring vector-based strategies overall but transition-based strategies near learnt landmarks. Interestingly, deep meta-learning models developed strikingly similar behavioural profiles. (3/6)

Cartoon image of me thinking about how to get to the ferris wheel. I consider the direction of the ferris wheel from my current location (‘vectors’) and the buildings I need to pass to get there (‘transitions’)

Schematic diagram of the task where participants learn a few landmarks (locations of objects on a grid) and then navigate the grid in two ways - choosing a direction (tapping on vector-based strategies) or choosing an object representing a neighbouring state (tapping on transition-based strategies)

We hypothesised that flexible navigation requires a mix of strategies, involving either a spatial sense of direction (‘vectors’) or associative knowledge between landmarks (‘transitions’). We designed a task to help dissociate between strategies while humans navigated unfamiliar grid worlds. (2/6)

Cartoon image of me looking at a map, with a stadium behind me and a hotel and ferris wheel across the river in the background. I am thinking about going to the ferris wheel

My first PhD paper - with @lhuntneuro.bsky.social and @summerfieldlab.bsky.social - is now out in @plosbiology.org! We ask: how do humans (and deep neural networks) navigate flexibly even in unfamiliar environments, such as a new city? Link: plos.io/45uSwNm 🧵 (1/6)