Write your post content here in normal
markdown. An example post is shown below for reference.
Recurrent Neural Networks and their variations are very likely to overfit the training data. This is due to the large network formed by unfolding each cell of the RNN, and relatively small number of parameters (since they are shared over each time step) and training data. Thus, the perplexities obtained on the test data are often quite larger than expected. Several attempts have been made to minimize this problem using varied regularization techniques. This paper tackles this issue by proposing a model that combines several of such existing methods.
Merity et al’s model is a modification of the standard LSTM in which DropConnect is applied to the hidden weights in the recurrent connections of the LSTM for regularization. The dropout mask for each weight is preserved and the same mask is used across all time steps, thereby adding negligible computation overhead. Apart from this, several other techniques have been incorporated :
- Variational dropout : The same dropout mask is used for a particular recurrent connection in both the forward and backward pass for all time steps. Each input of a mini-batch has a separate dropout mask, which ensures that the regularizing effect due to it isn’t identical across different inputs.
- Embedding dropout : Dropout with dropout probability is applied to word embedding vectors, which results in new word vectors which are identically zero for the dropped words. The remaining word vectors are scaled by as compensation.
- AR and TAR : AR (Activation Regularization) and TAR (Temporal Activation Regularization) are modifications of regularization, wherein the standard technique is applied to dropped output activations and dropped change in output activations respectively. Mathematically, the additional terms in the cost function are (here and are scaling constants and is the dropout mask) :
- Weight tying : In this method, the parameters for word embeddings and the final output layer are shared.
- Variable backpropagation steps : A random number of BPTT steps are taken instead of a fixed number, whose mean is very close to the original fixed value (). The BPTT step-size () is drawn from the following distribution (here is the Gaussian distribution, is a number close to 0.95 and is the desired variance) :
- Independent sizes of word embeddings and hidden layer : The sizes of the hidden layer and word embeddings are kept independent of each other.
The paper also introduces a new optimization algorithm, namely Non-monotonically Triggered Averaged Stochastic Gradient Descent or NT-ASGD, which can be programmatically described as follows :
They also combined their AWD-LSTM (ASGD Weight Dropped LSTM) with a neural cache model to obtain further reduction in perplexities. A neural cache model stores previous states in memory, and predicts the output obtained by a convex combination of the output using stored states and the AWD-LSTM.
Merity et al’s model used a 3-layer weight dropped LSTM with dropout probability
0.5 for PTB corpus and
0.65 for WikiText-2, combined with several of the above regularization techniques. The different hyperparameters (as referred to in the discussion above) are as follows : hidden layer size () =
1150, embedding size () =
400, number of epochs =
5, learning rate =
30, Gradients clipped at
1, dropout probabilities for input, hidden outputs, final output and embeddings as
Word embedding weights were initialized from and all other hidden weights from . Mini-batch size of
40 was used for PTB and
80 for WT-2.
- 3-layer AWD-LSTM with weight tying attained 57.3 PPL on PTB
- 3-layer AWD-LSTM with weight tying and a continuous cache pointer attained 52.8 PPL on PTB