💻 computer science

Evolutionary feature selection for spiking neural network pattern classifiers

This paper extends an evolutionary procedure for simultaneous feature selection and training to the biologically realistic JASTAP spiking neural network model, demonstrating that this approach enables smaller, more noise-robust classifiers with maintained accuracy on the IRIS dataset.

Original authors: Michal Valko, Nuno C. Marques, Marco Castelani

Published 2026-04-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Michal Valko, Nuno C. Marques, Marco Castelani

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 are trying to teach a computer to recognize different types of flowers. Usually, we teach computers using a standard "recipe" called a Multi-Layer Perceptron (MLP). Think of this standard recipe like a very rigid, loud factory assembly line: it processes information in big, slow waves, and to get the job done right, you often need a huge factory with thousands of workers (neurons) shouting instructions at each other.

This paper introduces a new, more "biological" recipe called JASTAP. Instead of a loud factory, JASTAP is like a team of highly skilled, quiet messengers who communicate by sending precise, timed "pings" (spikes).

Here is a breakdown of what the researchers did and found, using simple analogies:

1. The New Messengers (JASTAP)

The standard computer models are like a continuous stream of water. The JASTAP model is like a series of distinct raindrops.

How it works: Instead of just saying "I'm excited," a JASTAP neuron says, "I'm excited at exactly this moment." It uses the timing of these "pings" to carry information.
The Benefit: Because these messengers are so precise, the researchers found they needed far fewer of them to do the same job as the loud factory. It's like needing a small, elite special forces team instead of a massive, noisy army to win a battle.

2. The Smart Selector (FeaSANNT)

Teaching a computer usually involves two hard tasks at once:

Feature Selection: Figuring out which clues actually matter (e.g., "Do we need to measure the petal width, or is the color enough?").
Training: Teaching the network how to use those clues.

The researchers used a "Smart Selector" tool called FeaSANNT. Imagine a coach who doesn't just train the players but also constantly re-evaluates the team roster. The coach tries different combinations of players and different strategies simultaneously using an "evolutionary" process (like natural selection: the best teams survive, the weak ones are cut).

The Innovation: They took this Smart Selector and taught it how to work with the new JASTAP "messengers" instead of the old factory workers.

3. The Test Drive (The Iris Flower Dataset)

The team tested this new system on the famous "Iris" dataset, which involves sorting three types of flowers based on measurements.

The Result: The new system (JASTAP + Smart Selector) achieved 100% accuracy on the test data.
The Efficiency: It did this with a tiny network (only 7 neurons) and by ignoring some of the extra data points, proving that you don't need a giant brain to solve this problem.

4. The Noise Test (The "Static" Challenge)

Real life is messy. Data often comes with "noise" or errors, like static on a radio.

The Experiment: The researchers added artificial "static" to the data. They didn't just add random errors; they added "Gamma noise," which mimics the kind of biological noise that happens in real brains.
The Result: The JASTAP system was incredibly tough. Even with up to 10% noise (a lot of static), it still got 100% of the answers right. The standard factory models often struggle when the data gets this messy, but the JASTAP messengers could filter out the static and hear the signal clearly.

5. The Catch (Speed)

While the new system is smaller and tougher, it has one downside: it takes longer to learn.

Why? Because the system simulates time and the precise timing of every "ping," it has to do a lot more math than the standard factory model. It's like the difference between a quick snapshot and a high-definition movie; the movie is more detailed and realistic, but it takes longer to render.

Summary

The paper claims that by switching from a standard, loud computer model to a precise, time-based biological model (JASTAP) and using a smart evolutionary coach (FeaSANNT) to pick the best features, we can build smaller, tougher, and more accurate pattern classifiers. They proved this works on flower data, handling noise better than older models, though it currently takes more time to train.

1. Problem Statement

The paper addresses the limitations of traditional Multi-Layer Perceptrons (MLPs) and standard backpropagation learning in specific classification scenarios. Key issues identified include:

Lack of Recurrence and Context: Standard MLPs struggle with problems requiring temporal representation, context, or cyclic connections because they lack recursive structures.
Scalability and Complexity: To approximate complex functions or handle context, MLPs often require large numbers of hidden units, leading to computationally expensive models.
Training Difficulties: Training Spiking Neural Networks (SNNs) like JASTAP using standard backpropagation is difficult due to non-differentiable spike functions and cyclic connections.
Feature Selection: Selecting optimal input features and network parameters simultaneously is a complex, combinatorial search problem often prone to local minima when using gradient-based methods.

The authors propose replacing the basic perceptron model with the JASTAP (a biologically realistic spiking neural network model) and utilizing an evolutionary algorithm to simultaneously solve feature selection and network parameter optimization.

2. Methodology

A. The JASTAP Model

JASTAP is a spiking neuron model designed to simulate biologically realistic functions without the simplifications often found in other SNNs. Key components include:

Synapses: Defined by input type (internal/external), synaptic weight ( $SW \in [-1, 1]$ ), latency, and a Post-Synaptic Potential (PSP) waveform. The PSP is modeled using a Gamma distribution-like function controlled by time constants $t_1$ and $t_2$ .
Membrane Potential (MP): Calculated as the sum of PSPs, limited by a non-linear function ( $\frac{2}{\pi} \cdot \arctan(\sum PSP)$ ).
Firing Mechanism: A neuron fires when MP exceeds a threshold ( $\theta$ ). The spike frequency is regulated by an absolute refractory period, determining the inter-spike interval ( $I_a$ ).
Temporal Coding: Unlike rate-coded networks, JASTAP uses temporal coding where information is encoded in the timing of spikes.

B. The FeaSANNT Framework

The authors extend FeaSANNT, an evolutionary procedure originally designed for MLPs, to work with JASTAP. This framework solves two combinatorial problems simultaneously:

Feature Selection: A binary mask determines which input features are active.
Parameter Learning: Evolutionary algorithms (Genetic Algorithms) optimize network parameters, including:
- Synaptic weights.
- Synaptic latencies.
- Neuron thresholds.
- Input encoding parameters.

C. Data Encoding and Fitness

Input Encoding: Real-valued data is converted into temporal spike patterns. Data values are linearly mapped to inter-spike intervals (5–15 ms). To test noise robustness, Gamma-distributed noise is added to these intervals.
Output Decoding: The network decision is made when an output neuron fires (a "hot-spot" decision).
Fitness Function: A simple accuracy ratio was found insufficient. The authors developed a composite fitness function that rewards:
- Early/Non-silent responses.
- Heterogeneity: Ability to respond correctly to multiple classes.
- Selectivity: High accuracy on specific classes even if others are ignored.
- Multi-class capability: Bonus for correctly distinguishing pairs of classes.

3. Key Contributions

Integration of JASTAP with Evolutionary Learning: Successfully adapted the FeaSANNT evolutionary framework to train the biologically realistic JASTAP model, bypassing the need for gradient-based backpropagation.
Simultaneous Feature and Parameter Optimization: Demonstrated that evolutionary search can effectively select relevant features while tuning complex spiking network parameters (weights, latencies, thresholds) in a single process.
Noise Robustness via Biological Plausibility: Introduced a method to inject biologically relevant Gamma-distributed noise into the temporal spike trains, proving that the model can handle noise inherent in biological systems without performance degradation.
Compact Network Architecture: Showed that JASTAP can achieve high classification accuracy with significantly fewer neurons than traditional MLPs by leveraging temporal dynamics and recurrence.

4. Experimental Results

The experiments were conducted on the IRIS dataset (150 samples, 4 features, 3 classes).

Accuracy vs. Complexity:
- MLP (Backpropagation): Achieved 96.2% accuracy using all 4 features and a hidden layer.
- FeaSANNT (MLP): Achieved 94.7% accuracy using only 2 features (selected by the algorithm).
- FeaSTAP (JASTAP + FeaSANNT): Achieved 100% accuracy using no hidden layer (only 7 neurons: 4 inputs, 3 outputs) and mutual connections between input neurons.
Noise Handling:
- The JASTAP model was tested with Gamma-distributed noise added to inter-spike intervals.
- Result: Classification accuracy remained at 100% even with noise levels up to 10% (approx. 1ms jitter). Performance only degraded when noise exceeded this biological threshold.
Feature Selection:
- While the algorithm selected features, the reduction was less dramatic than in previous MLP studies (often retaining 3-4 features). The authors attribute this to the network's ability to utilize "irrelevant" spikes to overcome firing thresholds, a unique property of the spiking mechanism.

5. Significance and Conclusion

Efficiency: The study demonstrates that biologically realistic spiking models can outperform traditional MLPs in terms of architectural compactness (fewer neurons) while maintaining or exceeding accuracy.
Robustness: The model's ability to handle noisy, temporal data suggests strong potential for real-world applications where data is imperfect and context-dependent.
Future Directions: The authors highlight that while JASTAP offers superior representation for recurrence and context, the computational cost of simulation is high. Future work aims to parallelize the training process and refine the model to represent logic-based knowledge (Neuro-symbolic integration) by fixing certain architectures to encode background knowledge.

In summary, this paper validates that evolutionary feature selection combined with biologically realistic spiking neural networks is a viable and powerful alternative to standard perceptron-based classifiers, offering superior noise tolerance and structural efficiency.