🤖 machine learning

Grow-Prune-Freeze Networks: Adaptive & Continual Learning Technique for Olfactory Navigation

This paper introduces the Grow-Prune-Freeze (GPF) network framework, grounded in non-linear random matrix theory, which enables adaptive continual learning for olfactory navigation by dynamically growing, pruning, and freezing network layers to achieve a 94% success rate in turbulent plume tasks while demonstrating potential generalization to other machine learning domains.

Original authors: Kordel K. France, Ovidiu Daescu

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kordel K. France, Ovidiu Daescu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Smelly" Robot

Imagine you are building a robot that needs to find a hidden source of smell (like a leaky gas pipe or a flower) in a windy, chaotic room. This is called olfactory navigation.

The problem is that smell is messy. The wind blows the scent in unpredictable swirls (turbulence), and the signal is often weak or broken up. Unlike vision, where we have massive libraries of photos to teach robots what things look like, we don't have huge, standardized libraries of "smell data."

Because the environment changes so fast, a robot that is trained once and then left alone (a "static" robot) will fail. It needs to learn while it is moving, adapting to the wind and the smell in real-time. But if it learns too much new stuff, it often forgets what it learned earlier (a problem called "catastrophic forgetting").

The Solution: The "Grow-Prune-Freeze" (GPF) Network

The authors created a special type of AI brain called a Grow-Prune-Freeze (GPF) network. Think of this network not as a rigid computer chip, but as a living, breathing garden that the robot tends to while it works.

Here is how the three parts of the garden work:

1. Grow (Adding New Rooms)

The Analogy: Imagine the robot is trying to solve a puzzle. At first, it has a small table (a simple neural network). As the puzzle gets harder and the wind gets more chaotic, the table isn't big enough.
What the paper says: When the robot realizes it's stuck (the "loss" stops improving), the GPF network automatically adds a new layer (a new room) to its brain. This gives it more space to understand complex patterns in the smell. It doesn't guess how big the brain needs to be; it grows only as much as the job requires.

2. Prune (Weeding the Garden)

The Analogy: As the robot adds more rooms, it might start filling them with junk or useless furniture. If the garden gets too overgrown, the robot can't move fast enough.
What the paper says: The network constantly checks its own connections. If a specific part of the brain isn't being used often or isn't helping the robot find the smell, it gets pruned (cut away). This keeps the robot's brain small, fast, and efficient, ensuring it doesn't get bogged down by unnecessary data.

3. Freeze (Locking in the Memories)

The Analogy: Imagine you are learning to ride a bike. Once you figure out how to balance, you don't want to "unlearn" that skill just because you are now learning how to ride on a bumpy trail. You want to lock that basic skill in place.
What the paper says: As the robot learns, some parts of its brain become very stable and reliable. The GPF network freezes these layers. This means the robot stops changing those specific parts, protecting the knowledge it already has. This prevents "catastrophic forgetting," ensuring the robot doesn't lose its old skills while learning new ones.

The Secret Sauce: "Math Magic" (Random Matrix Theory)

The authors didn't just guess that this would work; they used advanced math called Random Matrix Theory to prove it.

The Analogy: Think of the robot's brain as a complex musical instrument. If you keep adding strings (layers) and cutting others (pruning), the instrument might go out of tune.
What the paper says: The authors showed that by using their specific Grow-Prune-Freeze method, the "tuning" of the instrument (the mathematical properties of the weights) stays perfect. Even as the brain grows and shrinks, the underlying math remains stable. This ensures that the robot's learning doesn't collapse under its own weight.

The Results: A Robot That Can Smell Its Way Home

The team tested this on a simulated robot trying to find a smell source in a turbulent wind.

The Challenge: The robot had to navigate a "turbulent plume" (a messy, swirling cloud of scent) where the smell would disappear and reappear randomly.
The Result: The GPF robot was incredibly successful. It found the source 94% of the time on new, unseen trials.
Comparison: A standard robot (one that didn't grow or prune) struggled much more. The GPF robot learned to "surge" forward when it smelled the scent, "cast" side-to-side when it lost it, and "loiter" when the wind was tricky—just like a real animal would.

Does it work for other things?

The paper also tested this "garden" method on other tasks, like recognizing images (CIFAR-10), playing video games (Atari's Ms. Pac-Man), and even writing text (GPT-2).

The Finding: In all these cases, the GPF network performed well, often stabilizing better than standard networks. It showed that this "grow, prune, and freeze" idea isn't just for smell; it's a general way to make AI that can learn continuously without breaking.

Summary

The paper introduces a method for robots to build their own brains while they work. Instead of being a fixed size, the brain grows when it needs more power, prunes to stay efficient, and freezes to protect what it has learned. This allows a robot to navigate a messy, changing world (like a windy smell) without forgetting how to do its job, backed by solid mathematical proof that the system stays stable.

Technical Summary: Grow-Prune-Freeze Networks for Olfactory Navigation

Problem Statement

The paper addresses the challenge of training robotic agents for olfactory navigation, a task characterized by sparse, intermittent signals, non-stationary environments, and a lack of standardized, large-scale datasets (unlike vision or language). Traditional deep learning architectures are static and rely on massive, pre-collected datasets, making them ill-suited for "big world" robotics where agents must adapt in real-time to evolving molecular patterns and turbulent plumes. Furthermore, standard continual learning approaches often suffer from catastrophic forgetting (loss of prior knowledge) or plasticity loss (inability to learn new tasks), particularly in partially observable environments where the agent must infer context from limited sensory history.

Methodology: Grow-Prune-Freeze (GPF) Framework

The authors propose Grow-Prune-Freeze (GPF) networks, an adaptive framework that dynamically modifies a neural network's architecture during runtime based on environmental complexity and learning performance. The framework integrates three core mechanisms:

Growing: The network starts with a single hidden layer. New layers are added only when the validation loss stagnates (exceeds a patience threshold $\rho_k$ without improvement), allowing the model to increase capacity as the task complexity (e.g., plume turbulence) demands it.
Pruning: To maintain efficiency and prevent over-parameterization, the network prunes "soft" neurons. A belief value ( $h \in [0, 1)$ ) is assigned to each neuron, incrementing when its weight magnitude or activation remains consistently high. Neurons with low belief values and small weights are removed. This is grounded in magnitude-based pruning (LeCun et al., 1989) but augmented by belief values to protect sparse but relevant patterns.
Freezing: To prevent catastrophic forgetting, layers that stabilize (weights change minimally over a patience period $\rho_f$ ) are frozen. This preserves the temporal dynamics learned from early encounters while allowing newer layers to remain plastic.

Theoretical Grounding

The methodology is grounded in Non-Linear Random Matrix Theory (RMT). The authors extend the work of Pennington & Worah (2017), which established spectral invariance for single-hidden-layer networks, to $n$ -layer continual-learning models.

They demonstrate that as layers are added, the eigenvalue composition of the network weights is preserved.
By freezing stable layers, the framework ensures the spectral distribution of the Gram matrix remains in its ideal shape (Marchenko-Pastur distribution), ensuring dynamical isometry and stable gradient propagation even as the network depth increases.
Theoretical analysis (Theorem 1) proves that under homogeneous layer assumptions, the Stieltjes transform of the empirical spectral distribution converges to a unique fixed point, ensuring stability during indefinite adaptation.

Policy and Training

The GPF framework is evaluated using Expected SARSA, a temporal difference learning algorithm chosen for its robustness in partially observable environments with sparse rewards (compared to greedy Q-learning). The agent operates in a simulated turbulent odor plume using a Farrell–Murlis filament model. The state space is discretized into 392 unique states based on bilateral antenna concentration bins and wind direction octants.

Key Contributions

Adaptive Framework: A novel RMT-grounded framework for adaptive reinforcement learning agents in partially observable, non-stationary environments. It combines growth, pruning, and freezing to balance plasticity and stability.
Theoretical Extension: An extension of Pennington & Worah (2017) and Martin & Mahoney (2021) to show that spectral eigenvalue invariance holds for multi-layer networks that grow indefinitely, providing a theoretical basis for resource-efficient continual learning.
Tuning Criteria: Specific criteria for stabilizing continual learning, including patience variables for layer growth ( $\rho_k$ ), pruning ( $\rho_p$ ), and freezing ( $\rho_f$ ), alongside a belief-value mechanism for neuron hardening.
Empirical Validation: Demonstration of GPFs achieving high success rates in olfactory navigation and evidence of generalization to other domains (Atari, CIFAR-10, GPT-2).

Experimental Results

Olfactory Navigation (Primary Task)

Performance: The GPF agent achieved a 98% success rate on the training distribution and 94% on held-out trials across diverse random initializations.
Architecture Evolution: The network started with one layer and grew to four layers through three growth events.
- After the first grow-prune pass, the network retained 47.9% of weights.
- Subsequent passes were more aggressive, retaining 19.6% and 9.9% of weights respectively, resulting in a highly sparse four-layer architecture.
Comparison: The GPF agent significantly outperformed a static baseline (single hidden layer) and showed superior stability compared to Q-learning agents, which struggled with sparse rewards and non-stationary plumes.

Generalization Experiments

The authors tested GPFs on three additional tasks to demonstrate generalizability, though they explicitly state they do not claim state-of-the-art performance across all benchmarks:

Image Classification (CIFAR-10): GPF models showed more stable validation loss convergence compared to static baselines, with eigenvalue spectra adhering to the Marchenko-Pastur distribution.
Reinforcement Learning (Atari - Ms. Pacman): GPF agents exhibited lower volatility in validation rewards compared to static policies.
Language Modeling (GPT-2 on TinyShakespeare): The GPF variant achieved steady convergence and ceased growing after reaching a maximum layer count, whereas the static baseline showed volatile loss.

Significance and Claims

The paper claims that GPF networks offer a practical solution for edge-deployable robotic learning where data is scarce and environments are dynamic. By grounding the architecture in RMT, the authors provide a mechanism for agents to allocate representational capacity only where environmental signals are complex, while preserving learned knowledge through freezing.

The authors are modest in their claims, noting that:

The primary motivation and validation are for olfactory robotics, a modality lacking standardized benchmarks.
The generalization results (Atari, CIFAR, GPT-2) are presented as supporting evidence for the framework's potential, not as definitive state-of-the-art benchmarks.
Future work is required to validate GPFs across a broader battery of tasks and to integrate them with federated learning and more complex computational fluid dynamics (CFD) simulators.

The work concludes that GPFs successfully address the "plasticity loss" identified in recent literature (Dohare et al., 2024) by enabling agents to grow, prune, and freeze in response to real-time data, offering a pathway toward truly adaptive, continual learning systems for robotics.