🤖 AI

The Whale That Outswam Evolution: Swarm Intelligence Maximises Memory in Connectome Reservoirs

This study demonstrates that applying bio-inspired, gradient-free optimisation algorithms—particularly the Whale Optimisation Algorithm—to the edge weights of biological connectomes significantly enhances reservoir computing performance across six species, proving that biological weight values serve as a crucial inductive bias that topology alone cannot replicate.

Original authors: Anmol Guragain, Savvas Kakalis, Juan Ignacio Godino-Llorente

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anmol Guragain, Savvas Kakalis, Juan Ignacio Godino-Llorente

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

Imagine you have a giant, complex machine made of millions of tiny switches (neurons) connected by wires. In the world of computers, we usually build these machines by randomly connecting the wires and hoping they work. But nature has spent millions of years building its own machines—brains—and the way it wired them is likely much smarter than our random guesses.

This paper asks a simple question: If we take nature's wiring diagram (a "connectome") and just tweak the strength of the connections between the switches, can we make the brain-machine even better at remembering things and predicting the future?

Here is the breakdown of their experiment, explained with everyday analogies:

1. The Setup: The "Brain" and the "Tuner"

The researchers took the wiring diagrams of six different animals, ranging from a tiny worm (C. elegans) to a fruit fly, a mouse, a rat, a monkey, and a human.

The Machine: They built a computer model called a "Reservoir" that mimics how these brains work. Think of this as a complex echo chamber where sound bounces around.
The Goal: They wanted to see if they could tune the "volume knobs" (the weights of the connections) on these echo chambers to make them better at two things:
1. Memory: Remembering a list of random numbers.
2. Prediction: Guessing what happens next in a chaotic system (like predicting the weather or a bouncing ball).

2. The Four "Tuners" (Optimizers)

To adjust the volume knobs, they didn't use a human or a standard computer algorithm. Instead, they used four bio-inspired "swarm" algorithms. Imagine these as four different teams of explorers trying to find the perfect setting for the knobs:

Particle Swarm (PSO): Like a flock of birds. They fly around, and if one bird finds a good spot, the whole flock moves toward it.
Differential Evolution (DE): Like a group of chefs mixing ingredients. They take different recipes, mix them together, and keep the tastiest one.
Grey Wolf (GWO): Like a pack of wolves hunting. They organize themselves into leaders and followers to surround the prey (the best solution).
Whale (WOA): Like humpback whales bubble-net feeding. They swim in a spiral, tightening their circle around the prey to catch it.

3. The Big Discovery: Evolution vs. Optimization

The researchers ran these "tuners" on the animal brains. Here is what they found:

Nature is a Great Start, but Not the Finish Line: The un-tuned biological brains were already decent at their jobs. However, when the algorithms tweaked the connection strengths, all four teams got much better results.
The Whale Wins: The Whale Optimizer (WOA) was the clear champion. It improved the memory capacity of the worm's brain by 17 times (from a score of 1.39 to 23.91). For the human brain model, it reduced prediction errors by nearly 90%.
The "Secret Sauce" is the Starting Point: This is the most critical finding. The researchers tried starting the Whale algorithm with random connections (ignoring the biological wiring) instead of the real animal wiring.
- Result: The random start failed miserably. It couldn't find a good solution even after thousands of tries.
- The Lesson: The pattern of the wires (topology) isn't enough. The specific values of the connections that evolution gave us are a "head start" (an inductive bias). You have to start with nature's blueprint to get the best results; you can't just start from scratch.

4. The Analogy: The Orchestra

Think of the animal brain as a symphony orchestra:

The Topology (Wiring): This is the seating chart. Who sits next to whom.
The Weights (Strengths): This is how loudly each musician plays.
The Experiment: The researchers kept the seating chart exactly as nature designed it (the biological connectome). They then asked, "If we just adjust the volume knobs for each musician, can we make the music sound better?"
The Result: Yes! The "Whale" conductor was able to adjust the volumes so perfectly that the orchestra played 17 times better than before.
The Warning: If you tried to assemble a new orchestra with random seating and random volume settings, no amount of tuning would make it sound as good as the one nature built. You must start with nature's seating chart.

Summary

The paper concludes that nature's wiring diagrams are incredibly valuable, but they aren't perfect. By using smart, nature-inspired algorithms (especially the Whale algorithm) to fine-tune the connection strengths, we can unlock massive improvements in how these brain-machines process information. However, you cannot skip the step of using the real biological starting point; without it, the optimization fails.

Key Takeaway: Evolution gave us a great foundation, but with the right "tuning," we can make these biological blueprints perform superhuman feats of memory and prediction.

Technical Summary: The Whale That Outswam Evolution

Problem Statement
Reservoir computing (RC) relies on fixed, high-dimensional recurrent networks to map inputs into a rich feature space, requiring only a trained linear readout. While standard practice initializes these networks with random, unstructured connectivity, biological neural circuits have been refined by millions of years of evolution to support efficient information processing. A central open question remains: do biological connectomes encode computational structures that random reservoirs lack, and if so, can principled, gradient-free optimization further enhance the performance of these evolutionarily shaped networks? Previous work has shown biological connectomes outperform random reservoirs, and meta-heuristics can improve weights without gradients, but no study has systematically compared multiple bio-inspired optimizers across diverse species and tasks under controlled conditions.

Methodology
The authors present a systematic benchmark evaluating four gradient-free, bio-inspired optimizers applied to the edge weights of Echo-State Networks (ESNs) constructed from six biological connectomes. The study spans six species representing six orders of magnitude in neural complexity: C. elegans (279 neurons), Drosophila (49 nodes), mouse (112), rat (73), macaque (29 regions with continuous FLNe synaptic strengths), and human structural MRI connectivity (83 parcels).

The experimental pipeline involves:

Fixed Topology: The sparsity pattern of each biological connectome is frozen.
Optimization: Four optimizers—Particle Swarm Optimization (PSO), Differential Evolution (DE), Grey Wolf Optimizer (GWO), and Whale Optimization Algorithm (WOA)—tune the non-zero edge weights.
Initialization Conditions:
- Bio-initialised (A): Populations are initialized as Gaussian perturbations of the biological weights.
- Random-initialised (B): A control condition (PSO-B) uses uniform random initialization on the same topology to isolate the contribution of biological weight values.
Evaluation: Performance is assessed on four canonical benchmarks: Memory Capacity (MC), Lorenz attractor prediction, NARMA-10 system identification, and Mackey–Glass chaotic time-series prediction.
Constraints: All optimizers use a budget of 20 particles and 50 iterations (1,000 evaluations). The spectral radius is fixed at $\rho=0.97$ to maximize linear memory retention.

Key Contributions

Systematic Cross-Species Benchmark: This is the first study to compare multiple bio-inspired optimizers across multiple connectomes and tasks with held-out evaluation.
Demonstration of Inductive Bias: The study establishes that biological weight values constitute an essential inductive bias. Random initialization on the same topology consistently underperforms the unoptimized biological baseline, indicating that the specific weight values encoded by evolution are critical and cannot be recovered by topology alone within the given evaluation budget.
Identification of Superior Optimizer: The Whale Optimization Algorithm (WOA) is identified as the most effective optimizer across all species and tasks.

Results

Performance Gains: All four optimizers consistently outperformed unoptimized biological baselines when initialized from biological weights. WOA achieved the largest gains, with an average improvement of 214% across all species and tasks.
- Memory Capacity: WOA achieved up to a 17 $\times$ improvement (e.g., C. elegans increased from 1.39 to 23.91).
- Nonlinear Prediction: On the Mackey–Glass task, WOA reduced the Normalized Root Mean Squared Error (NRMSE) by up to 89% (human connectome), approaching the numerical integration error of the underlying system.
Optimizer Comparison: WOA consistently dominated, followed by GWO, DE, and PSO. PSO showed the most modest gains and, in the case of the macaque connectome with random initialization, actually degraded performance by 17%, highlighting the risk of premature convergence in dense, well-configured networks without biological seeding.
Task Specificity: Gains were most pronounced on Memory Capacity and Mackey–Glass tasks. NARMA-10 showed more modest improvements, likely because it requires tightly coupled nonlinear computation and extended memory simultaneously, a regime where weight tuning alone is inherently limited compared to architectural changes.
Species Variability: While C. elegans showed the largest relative gain (17 $\times$ ) due to its large reservoir size (279 nodes) and heterogeneous weights, the macaque achieved the highest absolute baseline and significant gains despite having fewer nodes (29), driven by its exceptional synaptic density (72.7%). This suggests optimisability is driven by graph density, weight heterogeneity, and reservoir size rather than phylogenetic complexity alone.

Significance
The paper positions bio-inspired, biologically-initialized optimization as a principled and broadly effective strategy for connectome reservoir computing. The findings suggest that the high-dimensional, non-convex weight landscape of a connectome-based ESN contains a well-positioned basin of attraction near the biological solution. Optimization navigates toward deeper minima within this basin, whereas random initialization fails to locate it. Consequently, the authors argue that when biological connectome data is available, it should always be used to seed optimization populations, regardless of the specific optimizer chosen. This work validates the utility of evolutionary structures as a starting point for artificial temporal computation and motivates future extensions to larger connectomes and richer optimization regimes.