Bamikole Olaleye Akinsehinde

It is true that deep learning models, especially Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have the ability to automatically learn features from data. However, feature engineering can still be a valuable approach in some cases to improve the performance of deep learning models. Feature engineering involves transforming the raw data into a more informative representation that can help the model learn more effectively. For example, in computer vision, hand-engineered features such as SIFT and HOG can be used to represent images and provide additional information to the model beyond the raw pixels.

Whether or not additional feature engineering is needed for a specific problem depends on the complexity of the data and the type of problem being solved. For example, if the data is relatively simple and the model is able to learn effective features from the raw data, additional feature engineering may not be necessary. On the other hand, if the data is complex and the model is struggling to learn effective features, additional feature engineering may be necessary to improve the performance of the model.

In terms of recommendations for deep learning algorithms, here is a rough guide:

• Image Recognition: CNNs are the most commonly used type of deep learning algorithm for image recognition, as they are well-suited to handling image data. ResNet, Inception, and VGG are popular architectures for this task.
• Natural Language Processing (NLP): RNNs and Transformer models are commonly used for NLP tasks, such as sentiment analysis and language translation. Bidirectional LSTM, GRU, and BERT are popular architectures for this task.
• Predictive modeling: Deep feedforward neural networks (also known as multilayer perceptrons) can be used for a wide range of predictive modeling tasks, such as regression and classification. The architecture of the network can be adjusted to handle different types of data and different levels of complexity.

It's important to note that these are rough guidelines and the best algorithm for a specific problem will depend on the characteristics of the data and the problem being solved. In practice, it is often necessary to experiment with different algorithms and architectures to find the best solution.

Feature Engineering (FE) plays a vital role in Deep learning despite deep learning algorithms is very efficient in learning more complex patterns, but implementation of FE could still be valuable depending on the complexity of data in achieving better performance with deep learning. Other machine learning methods likewise heavily rely on the use of FE.

Deep learning models gain improved model accuracy on unseen data by transforming raw data into features that more accurately depict the core problem to the predictive models.

Implementation of FE to data before training deep learning models would improve implementation speed and lower the processing power needed for deep learning models.

By simplifying the data complexity through FE, this implies we could use a lesser number of hidden layers (neurons) for deep learning instead of using so many of them to achieve better accuracy and efficiency.

Reducing model complexity is an excellent step to better understanding and interpreting created models and its result (performance). Also, reducing the number of hidden layers is a great starting to preventing overfitting in deep learning models.

Different deep learning algorithms are effective at solving a particular task and proficient implementation of these algorithms require their thorough understanding.

I'll be providing a lot of information soon about the use of the deep learning algorithms listed below, including implementation of FE when necessary for greater performance in addressing their suite of examples. Code examples would be provided.

(1) Autoencoders

(2) Convolutional Neural Networks (CNNs),

(3) Deep Belief Networks (DBNs),

(4) Generative Adversarial Networks (GANs),

(5) Long Short Term Memory Networks (LSTMs)- These include:

(i) Bidirectional LSTM (BiLSTM),

(ii) Gated Recurrent LSTM (GRU-LSTM),

(iii) Multi Timeframe LSTM (MTS-LSTM),

(iv) Stacked LSTM,

(v) Standard LSTM (LSTM),

(6) Multilayer Perceptrons (MLPs),

(7) Radial Basis Function Networks (RBFNs),

(8) Recurrent Neural Networks (RNNs),

(9) Restricted Boltzmann Machines (RBMs),

(10) Self Organizing Maps (SOMs)

Dear readers, if you find this information helpful, please recommend this discussion and follow me to get alert on ResearchGate and Medium - https://medium.com/@bamikole-akinsehinde. Do me well to connect with me on LinkedIn - https://www.linkedin.com/in/bamikole/

Thanks.

Yes, definitely. Feature engineering is the main step of extracting features that are used to classify the data. But in deep learning feature engineering is done automatically in deep NN. Complex patterns depend on how complex the data is. As the layers increases in NN it becomes useful to identify more complex patterns in the data.

In deep NN, as the layers increase, it takes more minute details and avoids over fitting when we use dropout layer, regularization and data augmentation, cross-validation when the dataset is very large.

Layers such as Convolutional layer (more than one), ReLU layer, Max Pooling Layer, Dropout layer, Fully Connected (feature extraction) layer, SoftMax layer are included in deep NN.