12 RNN

Critique of CNN

CNN only generates one answer according to one input, it cannot resolve one-to-many and many-to-one sequence-related problems

Problems to be solved:

1. Sequence input, one output (videos)
2. One input (CNN), sequence output (photo to text)
3. Sequence input, sequence output (translation)
4. Sequence input, sequence output about the original sequence ( per-frame video classification )

[!TIP] RNNs for non-sequential tasks Actually, RNNs can perform sequential processing on non-sequential data RNNs take multiple glimpses at the image, and the position RNN is focused on depends on the information extracted from previous glimpses When generating images, RNNs can output a series of color painting over time, building up the entire image

RNNs

The whole pipeline of NLP with Encoder——Decoder Models

flowchart LD
  A[Input Text] -->|Tokenizer} B(Input Number Sequence)
  B -->|Encoder by RNNs| C(Input Vector Sequence)
  C --> |Decoder by RNNs| D(Output Vector Sequence)
  D -->|