Improving Minimal Gated Unit for Sequential Data

TL;DR

This paper introduces an improved Minimal Gated Unit for sequential data processing by integrating Chrono Initializer, demonstrating enhanced speed and accuracy in tasks like machine translation and speech recognition.

Contribution

The paper proposes using Chrono Initializer with the Minimal Gated Unit to improve its performance on sequential data tasks.

Findings

Confirmed effectiveness of the proposed method

Improved processing speed and accuracy

Validated on adding and copy tasks

Abstract

In order to obtain a model which can process sequential data related to machine translation and speech recognition faster and more accurately, we propose adopting Chrono Initializer as the initialization method of Minimal Gated Unit. We evaluated the method with two tasks: adding task and copy task. As a result of the experiment, the effectiveness of the proposed method was confirmed.

Tables

Table 1. TABLE I: Parameters for experiment

	iteration	batch size	hidden size	learning rate
adding	5,000	50	128	0.001
copy	5,000	128	128	0.001

Equations

$${\mathbf{f}}_t$$

$$\widetilde{{\mathbf{h}}_t}$$

$${\mathbf{h}}_t$$

$${\mathbf{b}}^{(f)}$$

$${\mathbf{b}}^{(i)}$$

Full text

I INTRODUCTION

Recently, Long Short-Term Memory (LSTM)[1] has been successful in various fields. For example, machine translation and speech recognition are well known. Moreover, there are researches to find a simpler and more compact model than LSTM such as Gated Recurrent Unit (GRU)[2] and Minimal Gated Unit (MGU)[3]. The advantages of these models are the clarity of structure, the reduction of the number of parameters, and the shortening of learning time. On the other hand, they have a problem that the accuracy of the model becomes worse because the number of parameters are reduced.

In this paper, in order to obtain a model which can process sequential data faster and more accurately, we change the initialization method of MGU which has a relatively small number of parameters.

II RELATED WORKS

II-A Minimal Gated Unit

Minimal Gated Unit[3] is a derivative of GRU[2]. In detail, Reset gate and Input gate in GRU are integrated to one gate.

MGU is expressed by the following equations, where the weight for hidden state is $\left(\bm{W}^{(f)}_{h},\bm{W}_{h}\right)$, the weight for inputs is $\left(\bm{W}^{(f)}_{x},\bm{W}_{x}\right)$, the bias is $\left(\bm{b}^{(f)},\bm{b}\right)$.

[TABLE]

II-B Chrono Initializer

Chrono Initializer[4] was proposed as a gate bias initialization method for LSTM. This initializer makes LSTM better capture long-term dependence of data.

Chrono Initializer is expressed by the following equations, where $U$ is the uniform distribution and $T_{max}$ is the expected range of long-term dependencies to be captured.

[TABLE]

III PROPOSED METHOD AND COMPARISON

We propose adopting Chrono Initializer[4] as the initialization method of MGU[3]. By using Chrono Initializer, it is expected that long-term dependency can be easily grasped by MGU and the performance will be improved. The equations of our proposed method are expressed by adding Eq. (4) to Eq. (1), (2), and (3). We compare the proposed method with the method whose gate bias value is set to 1. For convenience, we call this method MGU (Const.).

IV EXPERIMENTS AND RESULTS

IV-A Methods

We evaluate the performance of the model with two tasks: adding task and copy task[5].

Adding task is a task that adds real numbers. The input consists of two rows. The first row is a random real-value column, and the second row is a mask of 0 and 1 (1 appears only twice). The output is the value obtained by adding the real-value in the column where the mask value is 1. In this task, we set the length of column to 50 and 250.

Copy task is a task to copy a sequence. When T is given, the length of input and output is T+20 and these components are integers from 0 to 9. The input consists of 10 random values (from 1 to 8), T-1 dummies (0), 1 signal (9), and 10 dummies in order. The output consists of T+10 dummies and first 10 random values of the input in order. In this task, we set T to 50 and 200.

IV-B Results

Table I represents the parameters of our experiment. We experimented three times for each previous method and our proposed method. The solid lines represent the average of the values, and the colored ranges represent the degree of variation.

Fig. 2 and 3 are the results of the adding task. The vertical axis represents Mean Squared Error (MSE), and the horizontal axis represents the number of iterations. From Fig. 2, MSE of the proposed method sharply decreases from around the iteration number 800, and converges to 0 around the iteration number 1,500. On the other hand, the MGU (Const.) converges to 0 around the iteration number 2,700. From Fig. 3, MSE of the proposed method sharply decreases from around 1,700 iterations and converges to 0. In the MGU (Const.), MSE never falls below 0.17.

Fig. 4 and 5 are the results of the copy task. The vertical axis represents Softmax Cross Entropy, and the horizontal axis represents the number of iterations. In both figures, the Softmax Cross Entropy of the MGU (Const.) is smaller than our proposed one up to iteration number 1,000. After exceeding the number of iterations, the value of the proposed method falls below the value of the MGU (Const.).

From the above four figures, our proposed method performs better than the MGU (Const.) in these tasks.

V CONCLUSION

We adopted Chrono Initializer as the initialization method of MGU in order to improve the performance. As a result of our experiment, it turned out that our proposed method performed better than MGU (Const.) in the two tasks. We showed that Chrono Initializer is effective to improve performance of MGU. We would like to conduct experiments that apply the proposed method to machine translation and speech recognition.