Why RNNs process sequential data in TensorFlow - Experiment to Prove It

Experiment - Why RNNs process sequential data

Problem:We want to understand why Recurrent Neural Networks (RNNs) are good at handling sequential data like sentences or time series. Currently, a simple feedforward neural network is used on sequence data, but it cannot remember previous steps, so it performs poorly.

Current Metrics:Training accuracy: 60%, Validation accuracy: 58%, Loss: 0.9

Issue:The model does not capture the order or context in sequences, leading to low accuracy and poor understanding of sequential patterns.

Your Task

Replace the feedforward network with an RNN to improve understanding of sequence order and increase validation accuracy to above 75%.

Use TensorFlow and Keras only.

Keep the dataset and preprocessing the same.

Train for no more than 10 epochs.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, SimpleRNN, Dense
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras.datasets import imdb

# Load dataset (IMDB movie reviews as sequences of word indexes)
max_features = 10000  # number of words to consider
maxlen = 100  # cut texts after this number of words

(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)

# Pad sequences to the same length
x_train = pad_sequences(x_train, maxlen=maxlen)
x_test = pad_sequences(x_test, maxlen=maxlen)

# Build RNN model
model = Sequential([
    Embedding(max_features, 32, input_length=maxlen),
    SimpleRNN(32),
    Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

# Train model
history = model.fit(x_train, y_train, epochs=10, batch_size=64, validation_split=0.2)

# Evaluate on test data
test_loss, test_acc = model.evaluate(x_test, y_test)

print(f'Test accuracy: {test_acc:.2f}')

Replaced feedforward layers with an Embedding layer followed by a SimpleRNN layer.

Used an RNN to process sequences step-by-step, capturing order and context.

Kept dataset and preprocessing unchanged to isolate effect of RNN.

Results Interpretation

Before: Training accuracy 60%, Validation accuracy 58%, Loss 0.9

After: Training accuracy 85%, Validation accuracy 78%, Loss 0.4

RNNs process data step-by-step and remember previous information, which helps them understand sequences better than simple feedforward networks. This leads to higher accuracy on tasks involving sequential data.

Bonus Experiment

Try replacing SimpleRNN with LSTM layers to see if the model learns even better.

💡 Hint

LSTM layers have special gates that help remember information longer and avoid forgetting important details in sequences.

Practice

1. Why are Recurrent Neural Networks (RNNs) especially good at processing sequential data like sentences or time series?

easy

A. Because they remember information from previous steps while processing new data

B. Because they process all data points independently without any memory

C. Because they use convolution to detect patterns in images

D. Because they only work with fixed-size input data

Solution

1. Step 1: Understand RNN memory feature

   RNNs keep track of past information by passing hidden states through time steps.

2. Step 2: Connect memory to sequential data

   This memory helps RNNs understand order and context in sequences like sentences or time series.

3. Final Answer:

   Because they remember information from previous steps while processing new data -> Option A

4. Quick Check:

   RNN memory = good for sequences [OK]

Hint: RNNs remember past inputs to handle sequences well [OK]

Common Mistakes:

• Thinking RNNs treat each input independently
• Confusing RNNs with convolutional networks
• Assuming RNNs only work with fixed-size data

2. Which of the following is the correct way to define a simple RNN layer in TensorFlow using Keras?

easy

A. tf.keras.layers.MaxPooling2D(pool_size=2)

B. tf.keras.layers.Dense(units=50)

C. tf.keras.layers.SimpleRNN(units=50)

D. tf.keras.layers.Conv2D(filters=50)

Solution

1. Step 1: Identify RNN layer syntax in TensorFlow

   The SimpleRNN layer is created with tf.keras.layers.SimpleRNN(units=...).

2. Step 2: Compare options to correct syntax

   Options A, B, and C are for other layer types (MaxPooling2D, Dense, Conv2D), not RNN.

3. Final Answer:

   tf.keras.layers.SimpleRNN(units=50) -> Option C

4. Quick Check:

   SimpleRNN layer syntax = tf.keras.layers.SimpleRNN(units=50) [OK]

Hint: SimpleRNN layer uses tf.keras.layers.SimpleRNN [OK]

Common Mistakes:

• Using Dense layer instead of RNN
• Confusing convolution layers with RNN
• Missing required units parameter

3. Consider this TensorFlow code snippet:

import tensorflow as tf
model = tf.keras.Sequential([
  tf.keras.layers.SimpleRNN(3, input_shape=(4, 2))
])
input_data = tf.constant([[[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]]])
output = model(input_data)
print(output.numpy())

What is the shape of the output printed?

medium

A. (1, 3)

B. (4, 3)

C. (1, 4, 3)

D. (4, 2)

Solution

1. Step 1: Understand input and RNN output shapes

   Input shape is (batch=1, time=4, features=2). SimpleRNN with 3 units outputs (batch, units) by default.

2. Step 2: Confirm output shape without return_sequences

   Since return_sequences is False by default, output shape is (1, 3), one output per batch.

3. Final Answer:

   (1, 3) -> Option A

4. Quick Check:

   Output shape = (batch, units) = (1, 3) [OK]

Hint: Default SimpleRNN outputs last step shape (batch, units) [OK]

Common Mistakes:

• Assuming output shape includes time steps
• Confusing input shape with output shape
• Expecting 3D output without return_sequences=True

4. You wrote this code to create an RNN model:

model = tf.keras.Sequential([
  tf.keras.layers.SimpleRNN(5, input_shape=(None, 10))
])
model.compile(optimizer='adam', loss='mse')
model.fit(x_train, y_train, epochs=3)

But you get an error about input shape mismatch. What is the likely cause?

medium

A. SimpleRNN requires input_shape to be (10, None) instead

B. The optimizer 'adam' is not supported for RNNs

C. Loss 'mse' cannot be used with RNN models

D. x_train does not have 3 dimensions matching (batch, time, features)

Solution

1. Step 1: Check expected input shape for SimpleRNN

   SimpleRNN expects input shape (batch, time steps, features). The input_shape=(None, 10) means variable time steps, 10 features.

2. Step 2: Identify common input shape mismatch

   If x_train is missing the time dimension or has wrong shape, TensorFlow raises an error.

3. Final Answer:

   x_train does not have 3 dimensions matching (batch, time, features) -> Option D

4. Quick Check:

   Input shape mismatch = x_train does not have 3 dimensions matching (batch, time, features) [OK]

Hint: Check x_train shape matches (batch, time, features) [OK]

Common Mistakes:

• Blaming optimizer or loss for input shape errors
• Swapping time and feature dimensions in input_shape
• Using fixed time steps when variable allowed

5. You want to build an RNN in TensorFlow to predict the next word in a sentence. Which approach best uses RNNs' ability to process sequential data?

hard

A. Feed the entire sentence as one vector without sequence order

B. Feed the sentence word by word to the RNN, keeping track of hidden states

C. Use a convolutional layer to scan the sentence for word patterns

D. Shuffle the words randomly before feeding to the RNN

Solution

1. Step 1: Recall RNNs process sequences stepwise

   RNNs read inputs one step at a time, remembering past info via hidden states.

2. Step 2: Apply this to next word prediction

   Feeding words in order lets the RNN learn context to predict the next word accurately.

3. Final Answer:

   Feed the sentence word by word to the RNN, keeping track of hidden states -> Option B

4. Quick Check:

   Sequential input + memory = Feed the sentence word by word to the RNN, keeping track of hidden states [OK]

Hint: Feed words in order to keep context for prediction [OK]

Common Mistakes:

• Ignoring word order by feeding whole sentence as one vector
• Using convolution instead of RNN for sequence tasks
• Randomizing word order before input