Neural sampling from cognitive maps enables goal-directed imagination and planning

This paper proposes an energy-efficient neural network model that integrates cognitive maps, stochastic computing, and compositional coding to enable goal-directed imagination and planning for novel problems using only local synaptic plasticity, offering a viable alternative to resource-intensive deep learning for edge devices.

The Gist

Imagine your brain is a super-efficient, low-power GPS that doesn't just tell you where you are, but also lets you daydream about where you could go, even if you've never been there before.

This paper introduces a new kind of artificial intelligence (AI) model called GCML (Generative Cognitive Map Learner). It tries to copy how human brains solve problems and plan for the future, but it does so in a way that is incredibly energy-efficient and doesn't require the massive computer power that today's AI (like the chatbots you use) needs.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: Modern AI is a "Gas-Guzzler"

Current AI systems are like massive, fuel-hungry trucks. They need huge amounts of electricity and years of training to learn a task. If you change the destination (the goal), the truck often has to stop and relearn the whole route from scratch.

Human brains, on the other hand, are like electric bicycles. They run on just 20 watts of power (about as much as a dim lightbulb). They learn on the fly, and if you suddenly say, "Actually, let's go to the park instead of the store," the brain instantly figures out a new path without needing a reboot.

2. The Three Secret Ingredients

The researchers found that the brain uses three specific "tools" to achieve this magic. They built a model that combines all three:

• Cognitive Maps (The Mental Atlas):
  Think of this as a mental map of your neighborhood. You don't just memorize a list of turns; you understand the relationships between places. If you know the store is two blocks north of the park, you can figure out how to get to the store even if you've never walked that specific path before. The brain builds this map by connecting "states" (where you are) with "actions" (what you did to get there).
• Stochastic Sampling (The "What If?" Daydream):
  This is the brain's ability to run simulations. Before you actually move, your brain runs a "mental movie" of different possible paths. It's like rolling a dice to see a few different outcomes. This randomness is actually a feature, not a bug—it helps the brain explore creative solutions rather than just sticking to the one obvious path.
• Compositional Coding (The Lego Set):
  This is how we understand complex things by breaking them into parts. Just like you can build a million different castles using the same set of Lego bricks, the brain can understand new, complex situations by combining familiar building blocks (like words, shapes, or concepts) in new ways.

3. How the GCML Model Works

The researchers built a digital version of this brain tool. Here is how it plays out in three different scenarios:

Scenario A: The Maze Runner (Spatial Navigation)

Imagine a rat in a maze. Before it moves, the rat's brain "replays" potential paths to the cheese.

• The Old Way: Standard AI would try to calculate the perfect straight line.
• The GCML Way: The model creates a "mental map" using a grid (like graph paper). It then adds a little bit of "noise" (randomness) to its thinking. This allows it to imagine many different winding paths to the goal.
• The Magic: Even if there is a new wall in the maze that the rat has never seen, the model can instantly imagine a path around it. It doesn't need to relearn the map; it just uses its sense of direction to detour.

Scenario B: The Abstract Problem Solver (The Graph)

Now, imagine a problem that isn't about space, but about logic—like finding the shortest route through a network of cities or solving a puzzle.

• The model treats this like a map. It learns that "Action A" moves you closer to "Goal B."
• Instead of giving you just one answer, it generates a menu of options. It might say, "Here is the shortest path, but here are three other paths that are almost as short but might be safer or cheaper."
• This is like having a travel agent who doesn't just book the cheapest flight, but shows you five different options so you can choose based on your mood.

Scenario C: The Master Builder (Compositional Tasks)

This is the most impressive part. The researchers tested the model on a task where it had to take a complex shape (a silhouette) and break it down into smaller building blocks (like a puzzle).

• The Challenge: This is a mathematically very hard problem (called "NP-hard"). Usually, computers need to try billions of combinations to solve it.
• The GCML Solution: The model learned the "rules of the game" using just a few examples. Then, it was given a completely new shape it had never seen before.
• The Result: Because the model understood the "grammar" of the shapes (how the blocks fit together), it could instantly imagine a way to break the new shape down. It didn't need to memorize every possible shape; it just used its "mental Lego skills" to figure it out.

4. Why This Matters for the Future

The biggest takeaway is efficiency.

• No Heavy Lifting: This model doesn't need "Backpropagation" (a complex, energy-intensive math trick used to train modern AI). Instead, it learns locally, like a brain, adjusting connections as it goes.
• Instant Adaptation: If you change the goal, the model doesn't need to retrain. It just updates its "mental map" and starts imagining new paths immediately.
• Edge Devices: Because it is so efficient, this technology could run on small, battery-powered devices (like a smartwatch or a robot vacuum) without needing to connect to a massive cloud server.

The Bottom Line

This paper suggests that we don't need to build bigger, hotter, and more expensive computers to get smarter AI. Instead, we should build AI that thinks more like a human: by building mental maps, daydreaming about possibilities, and using simple building blocks to solve complex, new problems. It's a shift from "brute force" computing to "intuitive" computing.

Technical Summary

1. Problem Statement

Current AI systems, particularly those based on Deep Reinforcement Learning (RL) and Large Language Models (LLMs), face significant challenges regarding energy efficiency, training data requirements, and adaptability. They often struggle to:

• Adapt instantly to new goals or changing environmental contingencies without retraining.
• Generalize to states or problems never encountered during training (e.g., novel barriers or compositional combinations).
• Operate efficiently on edge devices with limited power (biological brains operate on ~20W, whereas AI training requires massive energy).

The core question addressed is: What data structures and algorithms allow biological brains to perform flexible, goal-directed planning and imagination in both spatial and non-spatial (compositional) domains, and can these be implemented in energy-efficient neuromorphic hardware?

2. Methodology

The authors propose a novel neural network model called the Generative Cognitive Map Learner (GCML). This model integrates three key biological tools:

1. Cognitive Maps: Data structures that organize experiences by encoding relational information between states and actions.
2. Stochastic Computing: Using noise as a computational resource to sample diverse trajectories.
3. Compositional Coding: Representing complex states as combinations of simpler components.

Core Architecture & Learning Mechanisms

• Forward Model (Predictive Coding): The model learns embeddings ($Q$ for observations, $V$ for actions) into a high-dimensional state space. It predicts the next state ($\hat{s}_{t+1}$) based on the current state and action: $\hat{s}_{t+1} = Qo_t + V a_t$. This is learned via self-supervised local synaptic plasticity (Delta rules) during exploration.
• Inverse Model (Goal-Directedness): A weight matrix ($W$) is learned to map state differences ($s^* - s_t$) to action commands ($a_t$) that reduce this difference. This acts as a "sense of direction" toward any goal $s^*$.
  • Learning Rule: Hebbian plasticity ($\Delta W \propto a_t (s_{t+1} - s_t)^T$).
• Generative Sampling (Imagination): To generate plans without external feedback, the GCML replaces actual environmental observations with internal predictions (bootstrapping).
  • Stochasticity: Gaussian noise ($\epsilon$) is added to the action selection process ($a_t = W(s^* - \hat{s}_t) + \epsilon$). This transforms the deterministic CML into a probabilistic generative model capable of sampling multiple diverse trajectories.
• Affordance Gating: A mechanism ($G$) ensures that only executable actions are selected (e.g., not moving into a wall or removing a block that isn't there).

Implementation Details

• Learning: Requires only local, self-supervised synaptic plasticity. No backpropagation (BP) or backpropagation through time (BPTT) is needed.
• Hardware Suitability: Designed for in-memory computing (e.g., memristor crossbars) and neuromorphic chips, enabling low-latency, on-chip learning.

3. Key Contributions

1. Unified Framework: The first model to unify spatial navigation, abstract graph problem solving, and compositional tiling tasks under a single "goal-directed sampling" paradigm.
2. Inverse Model for Cognitive Maps: Introduces a simple linear inverse model ($W$) that provides a "sense of direction" to any goal, enabling generalization to unseen states.
3. Noise as a Resource: Demonstrates that adding noise to the inverse model allows for stochastic sampling of diverse, near-optimal paths (solving the $k$-shortest path problem) and robust exploration.
4. Compositional Generalization: Shows that the model can solve NP-hard tiling tasks (decomposing silhouettes) and generalize to completely novel start/goal configurations (e.g., 8-block silhouettes after training on 5-block silhouettes) without retraining.
5. Biological Plausibility & Efficiency: The model replicates rodent hippocampal replay data (diverse, goal-directed trajectories) and operates with minimal computational overhead, suitable for edge AI.

4. Results & Experiments

The model was validated in three distinct domains:

A. 2D Spatial Navigation (Rodent Hippocampus)

• Task: Generating imagined trajectories from a start point to a goal, including rerouting around novel barriers.
• Findings: The GCML reproduced the diversity of "replay" trajectories observed in rodent brains (Pfeiffer & Foster, 2013).
  • It successfully navigated unexplored regions of the map (generalization).
  • It dynamically rerouted around obstacles without remapping the entire cognitive map, relying on repulsive forces from "barrier cells."
  • Trajectory diversity was controlled by the noise level, matching biological observations.

B. Abstract Concept Spaces (Graph Navigation)

• Task: Finding paths in a random graph (32 nodes) and solving the $k$-shortest path problem.
• Findings:
  • The GCML generated a "menu" of heuristic solutions (multiple paths) rather than a single deterministic path.
  • It outperformed standard RL in adaptability: when the goal changed, GCML required zero replanning time, whereas traditional search algorithms (A*, K*) required extensive re-computation.
  • It achieved near-optimal path lengths with significantly fewer visited nodes (lower computational effort) compared to K*, mA*, and BELA* algorithms.

C. Compositional Domains (Silhouette Decomposition)

• Task: Decomposing 2D silhouettes into a set of building blocks (an NP-hard tiling problem).
• Findings:
  • Trained on 5-block silhouettes, the GCML successfully decomposed 8-block silhouettes (never seen during training) with near-perfect success rates.
  • It generalized to partial decomposition (stopping at a non-empty goal) and reconstruction (adding blocks to reach a goal).
  • Comparison: GCML significantly outperformed Deep RL (D3QN) and Model Predictive Control (MPC) in success rate and sample efficiency, especially when the number of sampling proposals was low (<20). RL struggled with sparse rewards, and MPC suffered from high computational latency and horizon limitations.

5. Significance

• Energy Efficiency: By eliminating the need for deep backpropagation and large-scale training, this approach paves the way for energy-efficient edge AI capable of complex planning.
• Flexibility: It solves the "catastrophic forgetting" and "retraining" problems of RL. The model can instantly adapt to new goals or environments using local plasticity.
• Biological Insight: The model provides a mechanistic explanation for how the brain uses hippocampal replay for imagination and planning, suggesting that the brain learns an inverse model alongside its cognitive map to enable "foresight."
• Algorithmic Innovation: It challenges the dominance of search-based and deep learning methods for planning, offering a sampling-based heuristic that is computationally cheaper and more robust to novel contingencies.

In summary, the paper demonstrates that neural sampling from cognitive maps, driven by simple local learning rules and stochasticity, is a powerful, biologically plausible, and energy-efficient mechanism for solving complex planning and problem-solving tasks in both spatial and abstract domains.