An alternative for one-hot encoding in neural network models

One-hot encoding offers input data, while also implementing changes the advantage of considering the in the forward and backpropagation contributions of input data instance procedures in order to achieve the property belonging to each category, to the output of having model weight changes, that result separately, by encoding the categories in from the neural network learning process for mutually orthogonal vectors. However onehot certain data instances of some feature encoding is impractical for features that category, only affect the forward pass take categories from a set of large cardinality.