Introduction
Introduction
Traditional neural networks excel at processing single, static inputs—like classifying a photograph. But they falter when faced with sequences where context and order matter, such as understanding speech, analyzing video frames, or predicting stock prices over time.
Recurrent Neural Networks (RNNs) solve this fundamental limitation by maintaining memory across sequences, allowing them to understand not just what they’re seeing now, but how it connects to what came before.
What can RNNs do?
Process Sequences
one to one
This is what we are doing using CNNs
For example: image classification
one to many
We has one input, and generate a sequence of output
For example: image captioning
many to one
We has a sequence of input which order matters, and we’ll give an output
For example: video classification
many to many (1)
We have many input and the networks give many output, the number of input and output may not be the same
For example: Google Translate (the input and output may have different number of words)
many to many (2)
Similar to the previous section, but we give an output for every input
For example: Per-frame video classification
Non-Sequential Data
Introduction
Sometimes we can process non-sequential data in a sequential way
Example: Digit Recognition
Instead of feeding the network with the entire image, we let the network decide which part of the image it wants to see, the process might be like follows
- The network asks to see a small fragment of the image
- The network process that part of image
- It figures out what should I see next to understand the entire image, then request us to give it see that part of data (Step 1)
Example: Image Generation
The network decide where to write and what to write, doing this process many times and stacking all the result will give us the complete image
What is RNNs structure looks like?
Introduction
Sequence to Sequence (seq2seq)
Long Short Term Memory (LSTM)
Problem with Vanilla RNNs Gradient Flow
LSTM
Multilayer RNNs
Every RNN layer has its own weights