10
$\begingroup$

I would like to understand better why LSTM can remember information for a longer time period than vanilla/simple recurrent neural network (SRNN) by redoing an experiment from the paper Learning Long-Term Dependencies with Gradient Descent is Difficult by Bengio et al. 1994.

See Fig 1. and 2 on that paper. The task is simple, given a sequence, if it starts with a high value (e.g. 1), then the output label is 1; if it starts with a low value (e.g. -1), then the output label is 0. The middle is noise. This task is called information latching since the model needs to remember the starting value while going through the middle noise in order to output a correct label. It used a single neuron RNN to build a model that exhibit such behavior. The Figure 2(b) shows the results, and the success frequency of training such a model decreases dramatically as the sequence length increases. There was no result for LSTM since it was wasn't invented yet in 1994.

So, I become curious and would like to see that if LSTM would indeed perform better for such a task. Similarly, I constructed a single neuron RNN for both vanilla and LSTM cell to model information latching. Surprisingly, I found LSTM performs worse, and I don't know why. Could anyone help me explain or if there is anything wrong with my code, please?

Here is my result:

enter image description here

Here is my code:

import matplotlib.pyplot as plt
import numpy as np    
from keras.models import Model
from keras.layers import Input, LSTM, Dense, SimpleRNN


N = 10000
num_repeats = 30
num_epochs = 5
# sequence length options
lens = [2, 5, 8, 10, 15, 20, 25, 30] + np.arange(30, 210, 10).tolist()

res = {}
for (RNN_CELL, key) in zip([SimpleRNN, LSTM], ['srnn', 'lstm']):
    res[key] = {}
    print(key, end=': ')
    for seq_len in lens:
        print(seq_len, end=',')
        xs = np.zeros((N, seq_len))
        ys = np.zeros(N)

        # construct input data
        positive_indexes = np.arange(N // 2)
        negative_indexes = np.arange(N // 2, N)

        xs[positive_indexes, 0] = 1
        ys[positive_indexes] = 1

        xs[negative_indexes, 0] = -1
        ys[negative_indexes] = 0

        noise = np.random.normal(loc=0, scale=0.1, size=(N, seq_len))

        train_xs = (xs + noise).reshape(N, seq_len, 1)
        train_ys = ys

        # repeat each experiments multiple times
        hists = []
        for i in range(num_repeats):
            inputs = Input(shape=(None, 1), name='input')

            rnn = RNN_CELL(1, input_shape=(None, 1), name='rnn')(inputs)
            out = Dense(2, activation='softmax', name='output')(rnn)
            model = Model(inputs, out)
            model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
            hist = model.fit(train_xs, train_ys, epochs=num_epochs, shuffle=True, validation_split=0.2, batch_size=16, verbose=0)
            hists.append(hist.history['val_acc'][-1])
        res[key][seq_len] = hists
    print()


fig = plt.figure()
ax = fig.add_subplot(111)
ax.plot(pd.DataFrame.from_dict(res['lstm']).mean(), label='lstm')
ax.plot(pd.DataFrame.from_dict(res['srnn']).mean(), label='srnn')
ax.legend()

I also have the result shown in notebook, which would be convenient if you'd like to replicate the results. It took over a day to run the experiment on my machine using CPU only. It could be faster on a GPU-enabled machine.

Update 2018-04-18:

I tried to reproduce a figure on the landscape of RNN inspired by Figure 6 in On the difficulty of training Recurrent Neural Networks. I find it interesting seeing the formation of cliff in the loss landscape as the number of recurrence/time steps/sequence length increases, which could be related related to explaining the difficult of training long sequences observed here. More details is available here.

enter image description here

Update 2018-04-19

Extending @shimao's experiment. It seems that LSTM and GRU are just not so good at latching on information. But switching to a different task, which I call bit relay, (@shimao's task 2), GRU performs better while SRNN and LSTM are equally bad.

Now, I tend to think the performance of a cell-type could be task-specific.

Task 1: information latching (1 unit; 10 repeats; 10 epochs)

enter image description here

Task 2: bit relay (8 unit; 10 repeats; 10 epochs)

enter image description here

Error bars are standard deviations.

Then, an intriguing question is why LSTM doesn't work on information latching. Given the simplicity of the task, it should be able to work, shouldn't it? Could be related to the landscape (e.g. Cliffs) with respect to its gradients.

$\endgroup$
12
  • $\begingroup$ It looks like you're using a single-unit RNN in both cases. Have you tried making it bigger,say 10 or 100? $\endgroup$
    – Alex R.
    Apr 16, 2018 at 20:39
  • 1
    $\begingroup$ No, the idea from the original paper is to design the simplest system so we could reason effectively. What's the motivation of using multiple cells? The task is so simple, and you could see that when the sequence is short, both LSTM and SRNN works well. $\endgroup$
    – zyxue
    Apr 16, 2018 at 20:44
  • $\begingroup$ I haven't read this paper closely (or recently) but are you sure that you're exactly reproducing the experimental conditions? It seems plausible that there are any number of small but important differences between your setup and the paper that could account for the error. Or, stated another way, this is basically a debugging problem where there is only one, relatively uninformative, unit test. $\endgroup$
    – Sycorax
    Apr 16, 2018 at 21:49
  • $\begingroup$ @Sycorax, the paper is only about vanilla RNN as LSTM wasn't invented yet in 1994. My results on vanilla RNN is consistent with theirs, it's why LSTM performs worse that puzzles me. $\endgroup$
    – zyxue
    Apr 16, 2018 at 21:57
  • $\begingroup$ This paper papers.nips.cc/paper/… may be of interest -- it addresses the long-term dependency issues raised by Bengio in the paper you cite, so replicating the experiment used by Schmidhuber might be a more productive path $\endgroup$
    – Sycorax
    Apr 16, 2018 at 22:10

2 Answers 2

3
+25
$\begingroup$

There is a bug in your code, since the first half of your constructed examples are positive and the rest are negative, but keras does not shuffle before splitting the data into train and val, which means all of the val set is negative, and the train set is biased towards positive, which is why you got strange results such as 0 accuracy (worse than chance).

In addition, I tweaked some parameters (such as the learning rate, number of epochs, and batch size) to make sure training always converged.

Finally, I ran only for 5 and 100 time steps to save on computation.

enter image description here

Curiously, the LSTM doesn't train properly, although the GRU almost does as well as the RNN.

I tried on a slightly more difficult task: in positive sequences, the sign of the first element and an element halfway through the sequence is the same (both +1 or both -1), in negative sequences, the signs are different. I was hoping that the additional memory cell in the LSTM would benefit here

enter image description here

It ended up working better than RNN, but only marginally, and the GRU wins out for some reason.

I don't have a complete answer to why the RNN does better than the LSTM on the simple task. I think it must be that we haven't found the right hyperparameters to properly train the LSTM, in addition to the fact that the problem is easy for a simple RNN. Possibly, a model with so few parameters is also more prone to getting stuck in local minimum.

The modified code

$\endgroup$
1
  • $\begingroup$ I extended your experiment and obtained consistent results. I think the performance is likely to be task specific. $\endgroup$
    – zyxue
    Apr 19, 2018 at 18:40
0
$\begingroup$

I don't know if it will a difference, but I'd use:

out = Dense (1, activation='sigmoid', ...

and

model.compile(loss='binary_crossentropy', ...

for a binary classification problem.

$\endgroup$
2
  • 1
    $\begingroup$ I believe mathematically the two are the same. In particular, sigmoid is equivalent to softmax where there are only two classes. binary_crossentropy is after all crossentropy for two classes. This problem is really puzzling me, and I am open to increasing the bounty if anyone could provide a nice explanation. $\endgroup$
    – zyxue
    Apr 19, 2018 at 3:07
  • $\begingroup$ Are you using the same parameters for each version? LSTM is more complex and might require more data, more epochs, a different learning rate, etc.@zyxue $\endgroup$
    – Wayne
    Apr 19, 2018 at 11:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.