PaperlayerView exact revision on arXiv

Source-linked AI summary

A Deep Evolutionary Approach to Bioinspired Classifier Optimisation for Brain-Machine Interaction

Jordan J. Bird, Diego R. Faria, Luis J. Manso, Anikó Ekárt, Christopher D. Buckingham

arXiv:1908.04784v1cs.NEcs.AIcs.LG

TL;DR

Inferring attentional and emotional states from complex EEG signals is a challenge in brain-machine interaction. The paper combines evolutionary feature selection and classifier optimization with LSTM comparisons; boosted LSTMs lead on the two state datasets, while a boosted DEvo MLP performs best on digit classification.

  • Problem

    Inferring attentional and emotional states from EEG is challenging because the signals are complex, nonlinear, and nonstationary.

    Source: §1 · p. 1

  • Method

    The study uses evolutionary search to select EEG features and optimize MLP topology, and compares these models with manually tuned LSTMs and boosted variants.

    Source: §4.1 · p. 8, §4.2 · p. 9, §4.3 · p. 9

  • Results

    AdaBoosted LSTMs achieved 84.44% accuracy on the mental state dataset and 97.06% on the mental emotional state dataset; the Adaptive Boosted DEvo MLP was best by far on MindBigData digits.

    Source: §4.4 · p. 11

  • Takeaways & Limitations

    The DEvo approach produced a resource-light model that closely matched a resource-heavy deep-learning model while computing in approximately 10% of the time, except in one case where it far outperformed competitor models.

    Source: §5 · p. 12

    TP9 features may reflect stress-related Sternocleidomastoid muscle activity, so concentration-related classification may inadvertently classify EMG rather than EEG.

    Source: §4.1 · p. 8

Abstract

from arXiv · show

This study suggests a new approach to EEG data classification by exploring the idea of using evolutionary computation to both select useful discriminative EEG features and optimise the topology of Artificial Neural Networks. An evolutionary algorithm is applied to select the most informative features from an initial set of 2550 EEG statistical features. Optimisation of a Multilayer Perceptron (MLP) is performed with an evolutionary approach before classification to estimate the best hyperparameters of the network. Deep learning and tuning with Long Short-Term Memory (LSTM) are also explored, and Adaptive Boosting of the two types of models is tested for each problem. Three experiments are provided for comparison using different classifiers: one for attention state classification, one for emotional sentiment classification, and a third experiment in which the goal is to guess the number a subject is thinking of. The obtained results show that an Adaptive Boosted LSTM can achieve an accuracy of 84.44%, 97.06%, and 9.94% on the attentional, emotional, and number datasets, respectively. An evolutionary-optimised MLP achieves results close to the Adaptive Boosted LSTM for the two first experiments and significantly higher for the number-guessing experiment with an Adaptive Boosted DEvo MLP reaching 31.35%, while being significantly quicker to train and classify. In particular, the accuracy of the nonboosted DEvo MLP was of 79.81%, 96.11%, and 27.07% in the same benchmarks. Two datasets for the experiments were gathered using a Muse EEG headband with four electrodes corresponding to TP9, AF7, AF8, and TP10 locations of the international EEG placement standard. The EEG MindBigData digits dataset was gathered from the TP9, FP1, FP2, and TP10 locations.

1. Introduction

Brain-machine interaction depends on classifying complex, nonlinear, and nonstationary EEG signals to infer attention and emotional states. The study proposes evolutionary optimisation of bioinspired classifiers for attention, emotion, and imagined-digit classification.

1. Introduction

  • Commercial dry EEG electrodes make brain electrical activity accessible outside laboratory setups, supporting its use as a source of user control.The introduction situates this opportunity in applications including IoT, HRI, and HCI.Source: §1 · p. 1
  • The study targets EEG-based inference of attention, emotion, and imagined digits, with the resulting states considered as possible control inputs for human-robot interaction.The authors identify the complexity, nonlinearity, and nonstationarity of EEG signals as a challenge for inferring attentional and emotional states.Source: §1 · pp. 1–2
  • The proposed framework combines evolutionary optimisation with bioinspired classification of complex brainwave signals.The stated contributions include a new evolutionary hyperheuristic intended to prevent convergence on local minima in the EEG feature space.Source: §1 · p. 2
  • The approach seeks accuracies close to resource-intensive deep learning through optimisation inspired by natural processes, and reports exceeding it in one case.The comparison is stated as a contribution; the specific experimental outcomes are reported later.Source: §1 · p. 2

2. Background

The background introduces EEG and statistical feature-based classification, then outlines evolutionary optimisation, neural-network classifiers, LSTMs, and boosting. It motivates these methods by EEG’s temporal structure and its use in brainwave classification.

2.1. Electroencephalography and Machine Learning with EEG

  • EEG records electrical activity produced by the brain, and statistical wave features are commonly used as inputs to machine-learning classifiers for mental and emotional states.EEG signals are measured in microvolts over time; skull placement is less invasive but reads signals through the skull bone.Source: §2 · pp. 2–3

2.2. Evolutionary Algorithms

  • Evolutionary algorithms search for solutions through populations, fitness-based selection, and reproduction, and have been used to choose neural-network parameters.The background also describes Denser and EvoDeep as related evolutionary approaches to network optimisation.Source: §2.2 · p. 3

2.3. Multilayer Perceptron

  • An MLP maps inputs through neuron layers to class or real-valued outputs, while backpropagation uses loss gradients to refine its weights.For classification, the background describes entropy as a measure of randomness or predictability; positive information gain denotes lower error and a better model.Source: §2.3 · pp. 3–4

2.4. Long Short-Term Memory

  • LSTMs use prior states and unit memory, making them effective for time-dependent data such as brainwaves.The background notes their use in speech recognition and brainwave classification.Source: §2.4 · p. 4, §2.4.3 · p. 5

2.5. Adaptive Boosting

  • AdaBoost combines weighted predictions from multiple model instances trained on random data subsets.The background describes this as a way to mitigate cases where selected parameters are less effective.Source: §2.5 · p. 5

3. Method

The DEvo approach applies biological inspiration to both EEG data preprocessing and classifier optimisation. It combines evolutionary feature selection and neural-network optimisation with classification, aiming to retain accuracy while reducing processing time.

3. Method

  • DEvo combines evolutionary optimisation of data preprocessing with optimisation of a neural network before classification.The stated goal is to retain high accuracies while reducing processing time.Source: §3 · p. 5
  • The three experiments classify attention, emotional state, and the digit imagined by a subject, using two prior Muse datasets and MindBigData.MindBigData supplies a single subject’s brainwaves; 15 experiments per digit were extracted, yielding 300 seconds per digit class.Source: §3.1 · pp. 5–6
  • The attention and emotional datasets are preliminary and are being developed toward broader population representativeness.The paper describes the MindBigData experiment as classifying one subject’s brainwaves and identifies gathering data from more subjects as future work.Source: §3.1 · p. 6
  • Because EEG is nonlinear and nonstationary, the method extracts temporal statistics from overlapping one-second windows with 0.5-second overlap, producing 2147 scalars per measure.The feature set is described as too large for real-time use and as a source of high dimensionality, motivating attribute selection.Source: §3.2 · pp. 6–7
  • Evolutionary optimisation selects discriminative attributes and searches MLP topologies, using classification accuracy as fitness and testing up to three hidden layers with 100 neurons each.The attribute-selection population starts with 20 subsets and runs for 20 generations; MLP evaluation uses 10-fold cross-validation.Source: §3.3 · pp. 7–8
  • The selected attributes feed an MLP and an LSTM; LSTM topology is manually tuned, while AdaBoost is applied to both model types.The LSTM search tests one hidden layer with 25–125 units in steps of 25; evolutionary tuning was avoided as computationally expensive.Source: §3.3 · p. 8

4. Results and Discussion

Across the three datasets, boosted LSTM performed best on the two state-classification tasks, while boosted DEvo MLP performed best on MindBigData. The results also show dataset-specific feature selection, evolutionary MLP search, and manually tuned LSTM topologies.

4. Results and Discussion

  • 99 mental-state, 500 emotional-state, and 40 MindBigData attributes were selected from 2550 dimensions; the authors suggest emotional classification uses more informative statistical attributes.The original datasets were discarded, and subsequent experiments used only the selected datasets.Source: §4.1 · p. 8
  • MindBigData’s selected attributes had weak information gain, including 0.128 and 0.125 for the leading FP1 and FP2 covariance features, and the authors describe the dataset as extremely difficult to classify.By comparison, the best single-attribute OneR accuracies were 17.13% for MindBigData, 49.27% for mental state, and 85.27% for emotional state.Source: §4.1 · p. 8
  • On average, MLP search made no further improvement after generation 2; after speciation was added, increasing gradients in Figures 7–9 indicated constant improvement and enabled discovery of more complex, better solutions.Source: §4.2 · pp. 8–9
  • Manual LSTM tuning selected 100 units for mental state, 25 for emotional state, and 100 for MindBigData, although the latter result was extremely low for a uniform 10-class task.Evolutionary LSTM tuning was not attempted because individual network training took multiple hours on the available GPU.Source: §4.3 · p. 9
  • Mental-state training took only 11% longer on average (30.26s) than emotional-state training despite having approximately five times as many attributes.Figure 11 reports approximate training times; 10-fold cross-validation makes actual time around ten times the displayed value.Source: §4.4 · p. 11, Figure 11 · p. 11
  • The authors identify limited computation as the reason for manual LSTM topology tuning and note that all three experiments used the Muse headband’s four electrodes.They propose evolutionary LSTM optimisation and higher-resolution EEG setups as future directions.Source: §4.4 · pp. 11–12
  • 84.44% and 97.06% accuracy were achieved by AdaBoosted LSTM on mental-state and emotional-state classification, while Adaptive Boosted DEvo MLP was best on MindBigData.Boosting the LSTM reduced its accuracy on MindBigData; the overall comparison is presented in Table 8 and Figure 12.Source: §4.4 · p. 11, Table 8 · p. 12, Figure 12 · p. 12

5. Conclusion

The conclusion presents DEvo as a resource-light EEG classification approach, with boosted LSTMs strongest on attention and emotion and a boosted DEvo MLP strongest on digit guessing. It highlights faster training and the need for broader subject data in future work.

5. Conclusion

  • 30% accuracy made the Adaptive Boosted DEvo MLP the most accurate model for the MindBigData digits dataset; LSTMs produced no meaningful or useful results.The conclusion also describes the DEvo MLP results as useful and says the approach saved time.Source: §5 · p. 12
  • Approximately 10% of the time was needed to compute while closely matching a resource-heavy deep-learning model, except in one case where DEvo far outperformed competitors.The conclusion characterizes the resulting model as resource-light.Source: §5 · p. 12
  • The models were trained on attributes selected with a bioinspired evolutionary algorithm.Source: §5 · p. 12
  • Future work should use larger datasets from more diverse subjects to better model thought patterns across a global population.Source: §5 · p. 12
Loading 1908.04784v1…