2025 Theses Doctoral

Context-dependent computations in navigation and decision-making

Sun, Minni

Animals constantly adapt to dynamic external environments, and one effective policy is to select from a toolkit of strategies or skills depending on the current context. For instance, in the classical model of insect odor navigation, instantaneous odor serves as a contextual signal for animals to adjust their navigation strategy: surging upwind when encountering the odor and searching randomly when losing the odor. More generally, the contextual signal could integrate external information with internal state to determine the appropriate strategy.

In Chapter 2, I discuss the context-dependent use of directional memories in fruit fly odor navigation. Using a virtual reality olfactory paradigm developed by the Ruta lab that allows precise closed-loop control over odor encounters, we found that flies track an appetitive odor corridor by following its boundary, alternating between rapid counterturns to exit the plume and directed returns to its edge. This "edge-tracking" behavior allows flies to advance along a plume even when it is not aligned to the wind.

Our behavior analyses show that edge tracking cannot be attributed to reflexive strategies based on instantaneous odor cues. Instead, flies alternate between leaving and returning states, with transition rates depending on the presence or absence of odor. State-dependent directional biases then guide flies away from or back toward the edge. Flies rely on distinct returning biases when tracking edges in different directions, suggesting that they use memory of past entry directions to bias subsequent return. I developed a switching state-space model implementing this hypothesis. The model, fitted to data, captures the ability of flies to track various edge directions, and predicts that a fly's returning bias can be altered by a few operant training sessions in which appetitive odor is turned on whenever flies spontaneously walk in a certain direction. We verified this prediction in flies through behavioral experiments. Components of our model can be mapped directly onto the neural circuits of the Drosophila fan-shaped body, building the foundation for investigating the underlying neural mechanisms. As a first step in this direction, we found that FC2 neurons within the fan-shaped body signal state-dependent directional goals that guide a fly's return to the odor edge.

In Chapter 3, I discuss how inferring computational contexts can facilitate continual learning in recurrent neural networks. The ability to continually learn different tasks is an important feature of cognitive flexibility, yet it remains a challenging problem in neural network models. Most existing models encode task identities as one-hot embeddings, failing to capture the similarity or compositional structures of different tasks. We show that a large family of cognitive tasks studied in neuroscience can be systematically described by a probabilistic generative model, in which compositionality stems from a shared vocabulary of discrete computational contexts. These contexts can then dictate what computations to perform across different trial epochs. We developed an online learning algorithm that allows this task model to build its vocabulary incrementally as it encounters new tasks and to infer time-varying contexts given task identity and trial input.

I then designed a recurrent neural network with low-rank components selectively gated by the computational context inferred by the task model. Contextual inference facilitates the creation, learning, and reuse of low-rank RNN components as new tasks are introduced sequentially, enabling continual learning without catastrophic forgetting. Using an example task set, I demonstrate the competitive performance of this learning framework, its potential for forward and backward transfer, as well as rapid compositional generalization to unseen tasks.

Other than discrete contexts, animals can adapt their strategies to continuous contextual signals. In Chapter 4, I show that adapting to a continuous context, i.e., implicit task demand, leads to improved performance in a visual tracking task. Humans are adept at tracking moving objects in complex scenes, even when targets are visually identical to distractors and can only be tracked through spatial continuity, a task known as multiple object tracking (MOT). Behavior studies on this task show that tracking precision flexibly adapts to task demand, with precision increasing when distractors are in close proximity. I ask whether task-optimized neural networks can reproduce this flexibility, and if so, what type of objective function is required.

I compared a classification objective, which mirrors the decision format of MOT by labeling objects as targets or distractors, with a regression objective, which enforces slot-based outputs for target locations. Only classification-based models reproduced the demand-driven tracking precision observed in humans, and they generalized more effectively to larger target sets and to clustered targets. These results suggest that the classification objective aligns more closely with the behavioral goals of visual tracking. By contrast, the regression objective, designed for veridical estimation, is advantageous for extrapolating predictable trajectories, making it better suited for modeling visual–motor tasks such as object interception.