DATA: Sequences

• Examples:
  • Stock market data
  • Text (Book, dialogues corpus)
  • Music
  • Handwriting/Art strokes etc.

Problem?

• Can't use previous models on sequences.
• No context information.
• No time-dependency.
• Sequences are temporal in nature.

Any Solutions?

• Convert Temporal data to Spatial.
• Markov model: model the data as probabilistic model using markov chains.

 

 

 

Problem in mc models

Markov Chains (MC) models have fixed-sized context (order of the chain)

 

Fixed-size context == no context preservation for long term contextual information.

Memory (context) issue

Solution?

enter recurrent architectures 

recurrent neural network

Recurrent nodes

• Contain Feedback loops.
• Analogous to sequential digital circuits.
• Propagate information through time.

unroll in time

Information is passed via hidden states through time.

Working of rnn

issues with rnn

and improvements

vanishing gradients

• Occurs due to backpropagation through time.
• As sequence length increases, the gradient value diminishes or explodes!
• Happens due to consecutive differential matrices in chain rule!

fixing vanishing gradients

• Better Initialization of W to Identity.
• Application of ReLu as activation.

context (long-term) loss

Although theoretically possible, RNNs are very bad at capturing long-term dependencies in data.

Can be improved by increasing hidden state size, but there's a limit to space complexity!

 

SOLUTION?

LSTMs

lstm

long short-term memory

lstm node

• Specifically designed to overcome the long-term dependency problem.
• Has shown tremendous performance improvements!

Notation

core improvements

• The Context vector along with the hidden state.
• Use of Addition gate and Forget gate!

lstm forget gate

lstm addition gate

lstm context update

lstm output state