Backpropogating an LSTM: A Numerical Example
Let’s do this…
We all know LSTM’s are super powerful; So, we should know how they work and how to use them.
Syntactic notes
- Above ⨀ is the element-wise product or Hadamard product.
- Inner products will be represented as ⋅
- Outer products will be respresented as ⨂
- σ represents the sigmoid function
The forward components
The gates are defined as:
Which leads to:
Note for simplicity we define:
The backward components
Given:
- ΔT the output difference as computed by any subsequent layers (i.e. the rest of your network), and;
- Δout the output difference as computed by the next time-step LSTM (the equation for t-1 is below).
Find:
The final updates to the internal parameters is computed as:
Putting this all together we can begin…
The Example
Let us begin by defining out internal weights:
And now input data:
* Mohamed Challal pointed out to me that a label of 1.25 makes no sense since the outputs are a product of a tanh and sigmoid. Mohamed is completely correct!
I’m using a sequence length of two here to demonstrate the unrolling over time of RNNs.
Forward @ t=0
From here, we can pass forward our state and output and begin the next time-step.
Forward @ t=1
And since we’re done our sequence we have everything we need to begin backpropogating.
Backward @ t=1
First we’ll need to compute the difference in output from the expected (label).
Note for this we’ll be using L2 Loss:
The derivate w.r.t. x is:
So,
Now we can pass back our Δout and continue on computing…
Backward @ t=0
And we’re done the backward step!
Now we’ll need to update our internal parameters according to whatever solving algorithm you’ve chosen. I’m going to use a simple Stochastic Gradient Descent (SGD) update with learning rate: λ=0.1λ0.1.
We’ll need to compute how much our weights are going to change by:
And updating out parameters based on the SGD update function:
And that completes one iteration of solving an LSTM cell!
Errata and Frequently Asked Questions:
- Q: in `d state_t` did you mean to use `tanh²(state_{t-1})`?
A: No. - Q: you compute `d x` but never use it. Why?
A: you would use it if there were LSTMs stacked beneath, or any trainable component leading into the LSTM. Since `x` is the input data in my example, we don’t really care about that particular gradient. - Q: under Backwards @ t=0: you use `delta out_{-1} = U^T d gates_1`, but it should use `gates_0`.
A: Nice catch!
Of course, this whole process is sequential in nature and a small error will render all subsequent calculations useless, so if you catch something email me at hello@aidangomez.ca
Please feel free to share with the machine learning enthusiasts in your life!