PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolutionary feature selection for spiking neural network pattern classifiers

Michal Valko, Nuno C. Marques, Marco Castelani

arXiv:2604.26654v1cs.NE

TL;DR

The paper asks whether the biologically realistic JASTAP model can replace basic perceptrons for classification while supporting feature selection and parameter learning. It extends FeaSANNT's evolutionary procedure to JASTAP and reports preliminary IRIS results showing smaller networks and noise tolerance without degraded classification accuracy.

  • Problem

    Basic perceptrons can require many units, cannot handle cyclic connections, and offer limited support for recurrence, context, and background-information encoding.

  • Method

    The paper applies FeaSANNT's evolutionary procedure for simultaneous feature selection and parameter learning to JASTAP neural networks.

  • Results

    Preliminary IRIS results use smaller JASTAP networks without compromising accuracy and show no classification-accuracy degradation with artificially added training noise.

  • Takeaways & Limitations

    JASTAP can serve as a classification model with smaller networks and can accommodate artificially generated noise within the reported preliminary setting.

Abstract

from arXiv · show

This paper presents an application of the biologically realistic JASTAP neural network model to classification tasks. The JASTAP neural network model is presented as an alternative to the basic multi-layer perceptron model. An evolutionary procedure previously applied to the simultaneous solution of feature selection and neural network training on standard multi-layer perceptrons is extended with JASTAP model. Preliminary results on IRIS standard data set give evidence that this extension allows the use of smaller neural networks that can handle noisier data without any degradation in classification accuracy.

I. INTRODUCTION

The paper motivates replacing basic perceptrons with biologically realistic JASTAP networks for classification, especially where recurrence, context, or background knowledge matter. It extends evolutionary feature selection and parameter learning to JASTAP and introduces the model through an IRIS demonstration.

  • Motivation: JASTAP is proposed as an alternative to basic perceptrons for classification problems needing better recurrence and context representation.The model is biologically plausible, parameterized, and can retain flexibility by avoiding a priori simplifications.
  • Learning approach: Standard backpropagation remains an open issue for JASTAP and simpler spiking models because of varied weights and cyclic connections.The paper therefore uses a more generic evolutionary learning method.
  • Learning approach: FeaSANNT simultaneously searches for selected features and learned network parameters, extending earlier work from traditional neural-network classifiers to JASTAP.This extension targets the same class of classification problems.
  • Motivation: MLP models often require many units and cannot handle cyclic connections, limiting context or time representation.The paper also identifies difficulty encoding background information or algorithms in adjustable neural-network structures.
  • Model context: JASTAP belongs to spiking models that use temporal coding and can perform nonlinear classification with fewer units than classical rate-coded networks.The model is presented as a biologically realistic neural-network framework whose classification application is demonstrated on IRIS data.

B. Model Description

JASTAP networks consist of spiking neurons connected through synapses whose temporal and waveform properties shape signal processing. The model specifies membrane potential, firing threshold, spike frequency, synaptic structure, and optional plasticity.

  • Network structure: A JASTAP network is an artificial neural network whose basic elements are JASTAP neurons.Each neuron is connected to its environment through one or more synaptic inputs and a single axon output.
  • Synapses: PSP waveform parameters t1 and t2 control potential decay, with values up to 5 ms and 15 ms used for more moderate decay.The model uses time-and-space summation and simulation time discretization during processing.
  • Plasticity: Plastic changes can modify synaptic weights, but online learning during information processing is not used in this work.The paper identifies this capability as a possible future mechanism rather than part of the present experiments.
  • Neuron dynamics: The instantaneous membrane potential is the sum of PSPs limited by a nonlinear function, while threshold θ determines firing.Membrane potential is constrained to the ⟨-1, 1⟩ range and threshold to ⟨0, 1⟩.
  • Neuron dynamics: Spike frequency is constrained by minimum and maximum inter-spike intervals, with standard values Imin = 1 ms and Imax = 10 ms.The refractory constraint prevents firing sooner even when membrane potential exceeds threshold.

C. Tasks Where JASTAP Has Already Succeeded

Earlier JASTAP studies established that the model can process temporal spike patterns, distinguish spike rates, and make decisions about distributional features with few neurons.

  • Temporal processing: JASTAP networks can handle information encoded in temporal patterns, including noisy temporal information.
  • Spike-rate processing: JASTAP can recognize and distinguish different spike rates, which contrasts with viewing standard MLP weights as mean spiking rates.
  • Distributional features: JASTAP models can decide about Gamma-distribution features such as mean rate and coefficient of variation with very few neuron units.Prior work also examined evolved networks and described their low-level decision mechanisms.

III. FEASANNT

FeaSANNT addresses the joint search for neural-network features and parameters using evolutionary optimization. This motivation is tied to large, noisy, multimodal search spaces where gradient descent can converge to local minima.

  • FeaSANNT definition: FeaSANNT simultaneously solves feature selection and parameter learning for artificial neural-network classifier systems.
  • Optimization problem: ANN training often involves a large set of parameters, especially connection weights.
  • JASTAP context: The section includes figures showing PSP waveforms and a limiting nonlinear function relevant to JASTAP neuron processing.
  • Optimization problem: Gradient-descent training is prone to sub-optimal convergence to local minima, while evolutionary algorithms target noisy, multimodal, deceptive search spaces.

B. Feature Selection in an Embedded Approach

The paper extends embedded evolutionary feature selection to JASTAP spiking neural networks, jointly searching input masks and network parameters. This approach is motivated by potential reductions in network size, improved noise handling, and more analyzable temporal processing.

  • Feature Selection in an Embedded Approach: Feature selection searches the discrete space of attribute subsets, whose interactions make optimal selection difficult and computationally expensive.The search space is described as noisy, complex, nondifferentiable, multimodal, and deceptive.
  • Feature Selection in an Embedded Approach: FeaSANNT uses genetic algorithms to learn input masks while evolving classifier parameters in an embedded feature-selection procedure.The extension targets JASTAP spiking neural networks and jointly evolves features, thresholds, weights, and related parameters.
  • Feature Selection in an Embedded Approach: Earlier FeaSANNT experiments jointly evolved input vectors and ANN weights, achieving computational savings and substantial feature reduction across benchmark problems.The prior work used multi-layer perceptron classifiers and reported accurate, robust learning results on six real-world numerical datasets.
  • Feature Selection in an Embedded Approach: JASTAP is expected to require fewer neurons than MLPs for some pattern-recognition tasks.Evidence cited in the paper includes reduced unit counts on a simple XOR problem using a spike-response model.
  • Feature Selection in an Embedded Approach: JASTAP is investigated for extracting information from background spiking noise, especially biologically relevant Gamma noise.The paper also highlights temporal coding and smaller networks as potentially enabling inspection of processing at synapse level.

B. Encoding Variables into Spike–Time Patterns

The method encodes numeric inputs as repeated spike-time intervals and decodes decisions from designated output neurons. Evolution uses a composite fitness function because simple accuracy ratios poorly distinguish silent, indiscriminate, and selective networks.

  • Encoding Variables into Spike–Time Patterns: The temporal encoding choice remains an open issue because the relative importance of temporal and rate coding in information processing is unresolved.This uncertainty is presented as a broader neurophysiological problem rather than a result established by these experiments.
  • Encoding Variables into Spike–Time Patterns: Input data are converted into repeated interspike intervals whose lengths depend linearly on feature values within a 5–15 ms range.Noise experiments replace each interval with a value drawn from a Gamma distribution, considered biologically relevant in the paper.
  • Encoding Variables into Spike–Time Patterns: Output neurons define the decision: firing by any designated output neuron produces a task-dependent hotspot classification and stops the simulation.Subsequent network activity therefore does not affect the final decision.
  • Encoding Variables into Spike–Time Patterns: Fitness-function design affected both evolutionary iteration counts and learning effectiveness in recurrent JASTAP networks.The experiments used [c1, c2, c3] = [1, 30, 7].
  • Encoding Variables into Spike–Time Patterns: Simple success-ratio fitness functions performed poorly because they could not distinguish silent networks from networks that fired incorrectly, and local minima were common.The authors attribute this to individuals responding identically to all inputs.
  • Encoding Variables into Spike–Time Patterns: The composite fitness criteria reward heterogeneous class coverage, class selectivity, and multiclass responsiveness.The final function combines criteria for balanced correct responses, minimum class-wise fitness, and correct responses across class pairs.

A. IRIS Data Set

The evaluation uses the IRIS benchmark, which contains three flower classes described by four numeric petal and sepal measurements. Prior FeaSANNT results selected two features with 94.7 average accuracy, while a four-feature MLP achieved 96.2% accuracy.

  • IRIS Data Set: The IRIS dataset contains 150 flower examples across three classes, with four numeric features measuring petal and sepal dimensions.Only one class is linearly separable from the other two because some points overlap.
  • IRIS Data Set: 94.7 average accuracy was obtained when FeaSANNT selected only two features.This result is reported for the earlier FeaSANNT algorithm.
  • IRIS Data Set: 96.2% accuracy was achieved by an MLP using all four features with backpropagation learning.The figure is reported as a comparison with the two-feature FeaSANNT result.

B. Set–Up

The setup encodes each IRIS feature into a temporal spike pattern and represents evolving individuals with network-parameter chromosomes plus a feature-selection mask. Evolution retains the better half of parents and recombined offspring.

  • Set–Up: Individuals contain chromosome values subject to evolution together with a binary mask specifying selected features.Gray binary coding represents chromosome network values.
  • Set–Up: The first IRIS–setosa example maps four measurements to [7.2, 11.3, 5.7, 5.4] ms for four input neurons.The encoded values are repeated over 300 ms, with ±1 ms Gamma noise applied to every interspike interval.
  • Set–Up: The setup references datasets and evolution parameters through Tables I–III.The supplied table labels identify datasets, parameter boundaries, and evolutionary-algorithm settings.
  • Set–Up: JASTAP evolution leaves PSP parameters t1 and t2, along with Imin and Imax, unevolved according to prior results.The paper states that these parameters define bounding firing periods and are excluded from evolution.
  • Set–Up: After parents and recombined offspring are evaluated on the same examples, the best half of the combined chromosomes becomes the next population.This is an elitism-like selection procedure.

C. Results

The Iris experiment evaluates JASTAP classification with a seven-neuron architecture and tests robustness to independently perturbed inter-spike intervals. Classification quality remains stable through 10% noise, although learning becomes more difficult at that level and accuracy declines beyond it.

  • Iris data set: The Iris network uses seven neurons: four input neurons and three output neurons, with no hidden layer.Input-layer interconnections enable mutual information exchange and recurrence.
  • Handling the noise: 10% noise does not reduce classification quality, but learning becomes more difficult at that level.Noise was applied independently to each inter-spike interval, including intervals within the same input train.
  • Handling the noise: Noise beyond 10% compromises classifier accuracy.The reported robustness therefore has a boundary at noise levels above 10%.

VI. CONCLUSIONS

The paper extends evolutionary feature selection and parameter adjustment to the biologically realistic JASTAP model for classification. Preliminary Iris results indicate smaller networks and noise tolerance without degraded accuracy, while computational cost and possible overlearning remain concerns.

  • Conclusions: The paper aims to replace the basic perceptron with the biologically realistic JASTAP model for classification tasks.The conclusions identify this replacement as the paper’s main goal.
  • Conclusions: FeaSTAP applies an evolutionary procedure to simultaneous feature selection and JASTAP parameter adjustment.FeaSTAP is identified as JASTAP incorporated into FeaSANNT.
  • Conclusions: Preliminary Iris results use smaller neural networks without compromising accuracy and tolerate artificially inserted training noise without degrading classification accuracy.The reported feature-selection reduction was limited, with three or four mask elements retained.
  • Conclusions: The observed 100% Iris precision may reflect a statistical abnormality because the dataset is described as not fully separable.The authors recommend caution in future experiments to avoid opportunistic overlearning strategies.
  • Conclusions: The major limitation is the computational time required by the JASTAP learning model.Future remedies include faster time simulation, restricting learned parameters, and parallelizing FeaSANNT.
  • Conclusions: JASTAP’s small neuron count could help make networks more modular.The paper presents this as a prospective use rather than an established result.
  • Conclusions: JASTAP’s plastic changes and temporal coding are proposed as useful for online adaptation and classification problems involving context.The authors connect temporal coding with discovering statistical regularities over longer input-pattern times.
Loading 2604.26654v1…