Table 1 Advantages and disadvantages of existing research works.

⁴⁴

Using the MNIST dataset the hybrid model combines SVM for binary classification and CNN for automatic feature extraction to improve handwritten digit recognition.

On the MNIST dataset the CNN-SVM hybrid model uses the advantages of SVM’s classification capabilities and CNN’s feature extraction to achieve excellent recognition accuracy (99.28%).

Because it combines CNN and SVM the hybrid model might need a significant amount of processing power to train.

⁴⁵

Using global optimization and genetic operations such as selection, crossover, and mutation the method combines genetic algorithms with CNNs to optimize initial weights for the classification of liver CT tumor pictures.

Compared to conventional CNN and SVM approaches combining genetic algorithms with CNN improves medical-aided diagnosis and increases classification accuracy for liver CT images.

Because genetic algorithms are used in this method for initial weight optimization there may be an increase in computational complexity and training time.

⁴⁶

Use evolutionary techniques for joint optimization of a committee of CNNs and hyperparameter optimization.

Reduces the need for human tuning beats the state-of-the-art on MNIST and improves performance with a CNN committee.

Needs a lot of resources is sensitive to the initial settings and has problems with bigger datasets.

⁴⁷

Combines training three-layer CNN with GA for global search and optimal weight initialization.

Improves training time and accuracy by using GA to optimally initialize network weights.

Increased computational complexity due to the GA optimization process.

⁴⁸

Use GA to optimize hyperparameters and combine it with SAE, CNN, and GA for the prediction of anemia.

98.50% prediction accuracy for anemia is attained using GA-assisted hyperparameter optimization.

Complexities in choosing appropriate hyperparameters and higher processing demands.

⁴⁹

Selects trainable layers for transfer CNN models using the GA optimizing according to accuracy and the number of trainable layers.

By using GA to optimize trainable layers 97% classification accuracy is attained for datasets about cats and dogs.

Computationally demanding and needs to converge across several generations.

⁵⁰

Used evolutionary algorithms and Bayesian optimization to investigate hyperparameter search techniques concentrating on CIFAR-10 datasets and investigating the hybridization of genetic algorithms with local search techniques.

Potential advancement for network construction and training optimization through the hybridization of evolutionary algorithms with local search techniques.

On CIFAR-10 datasets no discernible gain in performance over state-of-the-art approaches.

⁵¹

Enhanced evolutionary algorithms with elements from nature for hyperparameter optimization and included significance sampling a Monte Carlo-based technique for reducing variance.

Improved hyperparameter solution space exploration resulting in improved model performance.

Added complexity and computational expense as a result of more improvements inspired by nature.

⁵²

Efficiently explored and optimized CNN topologies and hyperparameters for image classification using a genetic approach called fast-CNN.

CNN architectures can be designed and optimized more quickly than with conventional techniques.

Possibly not as accurate as the best manually optimized models.

⁵³

Analyzed using the Gradient-Descent Algorithm and the GA with a particular encoding technique for layer connectivity, filter dimensions, and fully connected layer nodes.

Automated design of CNN architectures without the need for data preprocessing or post-processing allowing for efficient exploration of network configurations.

It could take a lot of processing power to train CNNs from scratch at every stage of evolution.

⁵⁴

Framework using genetic algorithms to optimize and choose features from CNN models that have already been trained for various detection tasks.

Minimizes human labor and optimizes the procedure for various tasks by automating the selection of helpful features from CNN models that have already been trained.

It could take a lot of computer power to assess and choose features from several trained models.