Mice in a labyrinth: Rapid learning, sudden insight, and efficient exploration

Matthew Rosenberg, Tony Zhang, Pietro Perona, Markus Meister

TL;DR

Mice can learn complex navigation tasks with remarkable speed and efficiency, exhibiting "sudden insight" rather than gradual improvement. Their exploration strategy relies on simple, local turning rules, not a global memory of visited locations, suggesting a more innate and efficient learning mechanism than previously thought.

What Problem Does This Paper Address?

The paper tackles the question of how animals, and by extension artificial systems, can acquire complex behaviors with very few experiences. Traditional studies often use simplified tasks (like 2AFC) where learning is slow and requires extensive training. The authors argue that these tasks don't reflect the complexity of real-world navigation and may not reveal the underlying mechanisms of rapid, "few-shot" learning. They aim to study a more naturalistic, complex behavior – maze navigation – in mice, without human intervention, to understand how they learn efficiently and exhibit sudden insights.

Methods

The researchers designed a complex, 6-level binary tree labyrinth. They observed mice (both water-deprived for reward and sated controls) freely exploring this maze for 7 hours overnight, using infrared video and automated tracking. Key methods include:

  • Unconstrained Exploration: Mice were given free access to the maze without shaping or training.
  • Node Sequence Analysis: Trajectories were simplified into sequences of visited nodes (decision points and endpoints) to analyze navigation patterns.
  • Learning Curve Analysis: The number of reward experiences was correlated with the accuracy and efficiency of reaching the water port.
  • Sudden Insight Detection: Statistical models (step vs. ramp models) were used to identify discontinuous changes in behavior (reward rate, direct path frequency).
  • Exploration Efficiency Measurement: A metric (E = 32/N32) was developed to quantify how efficiently mice surveyed the maze's end nodes.
  • Turning Bias Analysis: Local turning rules at T-junctions were quantified and compared to random walks and more complex models.
  • Predictive Modeling: Fixed-depth and variable-depth Markov chains were used to predict mouse decisions based on historical trajectories.

Datasets

The study used data collected from 20 mice (10 rewarded, 9 unrewarded, one outlier excluded) over 7-hour observation periods. This constitutes a novel dataset of unconstrained, complex maze navigation.

Key Insights and Ideas

  • "Sudden Insight" is Real: Mice don't just gradually improve; they exhibit discontinuous leaps in performance, suggesting moments of sudden understanding about the maze structure.
  • Local Rules, Global Efficiency: The efficient exploration of the maze is explained by simple, local turning biases at junctions, not by a global memory of visited places. This is a key insight into how complex search can be achieved with minimal cognitive overhead.
  • Exploration is Intrinsically Rewarding: Even without external rewards, mice spend a significant amount of time exploring the maze, suggesting an intrinsic drive for information gathering.
  • Rapid Learning in Complex Tasks: Mice learn to navigate to a reward in a complex maze with a learning rate orders of magnitude higher than in simplified 2AFC tasks.
  • Innate Navigation Strategies: The consistency of turning biases across different animals and conditions suggests these rules might be genetically determined or innate.

Key Results

  • Mice learned to execute correct 10-bit choices (navigating to the water port) after only ~10 reward experiences, a learning rate ~1000x higher than in 2AFC tasks.
  • Sudden changes in behavior (reward rate and direct path frequency) were observed in at least 5 out of 10 water-deprived mice, often occurring minutes apart.
  • Mice explored the maze efficiently, discovering half of the end nodes with about 76 visits, performing about half as well as an optimal agent but twice as well as a random walker.
  • Exploration efficiency was remarkably consistent across animals and unaffected by reward motivation.
  • Local turning biases (forward bias, alternating turns) were consistent across mice and explained ~87% of the efficiency of a biased random walk model.
  • Predictive models showed that mouse decisions were constrained by recent history (around 4 preceding nodes), reducing uncertainty from 1.585 bits to ~1.24 bits per decision.

Limitations and Future Work

  • Understanding "Sudden Insight": While the study identifies moments of sudden insight, the precise neural mechanisms and triggers remain to be elucidated. Future neurophysiological studies are needed.
  • Mechanisms of Navigation: The paper rules out some navigation mechanisms (e.g., visual beacons, odor trails, vector-based navigation) but acknowledges that the exact sensory cues and algorithms used for direct path navigation are still unknown and require further directed experiments.
  • Generalizability of Exploration Rules: While local turning biases explain much of the exploration, some components of efficient search remain to be understood.
  • Intrinsic Reward of Exploration: The strong drive to explore, even without external rewards, is noted, but its precise nature and how it interacts with learning requires further investigation.
  • Model Limitations: Even sophisticated Markov chain models couldn't perfectly predict mouse behavior, leaving a residual uncertainty that might reflect "free will" or other unmodeled factors. The authors encourage others to improve upon their predictive models using their published data.

Summary updated

Abstract

Abstract Animals learn certain complex tasks remarkably fast, sometimes after a single experience. What behavioral algorithms support this efficiency? Many contemporary studies based on two-alternative-forced-choice (2AFC) tasks observe only slow or incomplete learning. As an alternative, we study the unconstrained behavior of mice in a complex labyrinth and measure the dynamics of learning and the behaviors that enable it. A mouse in the labyrinth makes ~2000 navigation decisions per hour. The animal quickly discovers the location of a reward in the maze and executes correct 10-bit choices after only 10 reward experiences – a learning rate 1000-fold higher than in 2AFC experiments. Many mice improve discontinuously from one minute to the next, suggesting moments of sudden insight about the structure of the labyrinth. The underlying search algorithm does not require a global memory of places visited and is largely explained by purely local turning rules.
BibTeX
@inproceedings{rosenberg2021mice,
  author = {Matthew Rosenberg and Tony Zhang and Pietro Perona and Markus Meister},
  title = {{Mice in a labyrinth: Rapid learning, sudden insight, and efficient exploration}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.66175},
}
Figure