Data vectorization converts inputs and targets into tensors that a neural network can process, typically floating-point tensors and sometimes integer or string tensors. It is one part of preprocessing, alongside normalization and handling missing values. section 6.2.1
The conversion depends on the data:
- Text represented as sequences of word integers can be converted into floating-point tensors with multi-hot encoding. section 6.2.1
- Some datasets, such as digit images and house-price features, may already be vectorized, so this step can be skipped. section 6.2.1
- Vectorized array operations are a related but distinct idea: built-in operations process tensor entries efficiently, unlike element-by-element loops. section 2.3.1
z <- x + y #1 z <- pmax(z, 0.) #2
Use built-in array operations to add tensors and apply ReLU without writing nested loops. section 2.3.1
Vectorize data as part of the preparation workflow before developing a model. Domain-specific techniques for text and images require choices beyond these shared basics. section 20.1.5 section 6.2.1
