Which layer is defined as not trainable?

The hashing layer is defined as not trainable because it performs a fixed transformation of inputs into a hashed representation. This transformation is typically a predetermined mapping that does not involve any learnable parameters or adjustments based on the training data. As a result, the hashing layer serves more as a preprocessing step that facilitates the model in managing high-dimensional input space by reducing dimensionality or encoding categorical variables, without the need for adaptive learning.

In contrast, the pooling layer, dense layer, and dropout layer typically involve parameters or mechanisms for adjusting based on the training process. The pooling layer reduces the spatial dimensions of the input while maintaining the most important features, often using operations like max or average pooling. The dense layer consists of neurons with weights that are specifically trained during the learning process to establish relationships between inputs and outputs. The dropout layer helps prevent overfitting by randomly setting a fraction of its inputs to zero during training, but it does so based on the results of the training process. Thus, the hashing layer stands out as the non-trainable layer among these options.