Sequence-to-Sequence Data Generation via Encoded Perturbations

The patent segments the training data generation process into three independent components: an encoder that transforms input sequences into encoded representations, a perturbation module that modifies encoded representations by adding noise or transformations, and a decoder that reconstructs sequences from modified encodings. This segmentation allows each component to be optimized independently and enables diverse data generation by varying perturbation strategies without limiting the model to only reproducing training data pairs.

The patent introduces encoded representations as an intermediary between the input sequences and output sequences. Instead of directly mapping input to output pairs, the system transforms inputs into encoded form, applies perturbations to these encodings, and then decodes to generate outputs. This intermediary representation layer enables the generation of diverse sequences that go beyond the original training data pairs while maintaining semantic coherence.

The patent makes the encoding process dynamic by introducing perturbation to the encoded representations. Instead of using fixed encodings, the system dynamically modifies encoded representations through added noise or transformations, allowing the same input sequence to generate multiple varied outputs while maintaining the underlying semantic structure processed by the RNN model.

The patent changes parameters of the encoded representations by applying perturbations such as noise addition or transformations to the encoded form. This parameter modification at the encoding level allows the system to generate diverse sequences while leveraging the strong sequential processing capabilities of the RNN, as the perturbations are applied to the encoded form rather than the raw input sequences.