Sequence classification is a common and important application of recurrent neural networks. The objective is to assign a single categorical label to an entire input sequence. Think of tasks like determining the sentiment (positive or negative) of a movie review, identifying the topic of a news article, or classifying the intent behind a user's query.

Unlike sequence prediction where we might predict the next element, or sequence-to-sequence tasks where we generate an output sequence, classification requires summarizing the information across the entire input sequence into a single decision. RNNs, LSTMs, and GRUs are well-suited for this because their hidden state acts as a evolving summary of the sequence processed so far.

Using Recurrent Layers for Classification

The fundamental idea is to process the input sequence step-by-step using a recurrent layer (SimpleRNN, LSTM, or GRU). As the network processes each element, it updates its hidden state, incorporating information from the current input and the previous state. By the time the network reaches the end of the sequence, the final hidden state (or states, in the case of bidirectional RNNs) should ideally capture a meaningful representation of the entire sequence's content, relevant to the classification task.

This final representation is then typically fed into one or more standard feedforward layers (often called Dense layers or Fully Connected layers) to perform the final classification.

Common Architectural Patterns

There are a couple of primary ways to use the output of the recurrent layer for classification:

Using the Final Hidden State: This is the most frequent approach. The RNN processes the sequence, and only the hidden state from the very last time step is used as the input to the subsequent classification layer(s). This final state is assumed to encapsulate the necessary information from the entire sequence. Framework APIs often have a parameter (like return_sequences=False in Keras) that controls whether the layer outputs the state only at the last time step or the hidden states for all time steps. For this pattern, you typically want the output only from the last step of the final recurrent layer in your stack.

A common architecture where the final hidden state from the recurrent layer is passed to a Dense layer for classification.
Using Pooled Hidden States: Instead of relying solely on the final hidden state, you can use the hidden states from all time steps. The return_sequences=True parameter (or equivalent) would be set on the last recurrent layer. These states are then aggregated using a pooling operation before being passed to the classification layer. Common pooling strategies include:
- Max Pooling: Takes the maximum value across the time dimension for each feature in the hidden state. This might capture the most salient features detected anywhere in the sequence.
- Average Pooling: Takes the average value across the time dimension for each feature. This provides a summary of the features present throughout the sequence.
Pooling can sometimes be beneficial if important information for classification might appear at any point in the sequence, not just towards the end. However, using the final hidden state is often simpler and performs well, especially with LSTMs and GRUs which are designed to maintain relevant information over long sequences.

Implementation Considerations

Input Preparation: As detailed in Chapter 8, your input sequences (e.g., text) need to be converted into numerical format (integer encoding), potentially passed through an Embedding layer, and padded to ensure all sequences in a batch have the same length. Masking should typically be used to inform the recurrent layers to ignore these padded steps. The input shape for the recurrent layer will usually be (batch_size, time_steps, feature_dimension).
Output Layer: The final Dense layer needs an appropriate activation function based on the nature of the classification task:
- Binary Classification: Use a single output unit with a sigmoid activation function. The corresponding loss function is typically BinaryCrossentropy.
- Multi-class Classification (Single Label): Use $N$ output units (where $N$ is the number of classes) with a softmax activation function. The typical loss function is CategoricalCrossentropy.
Return Sequences: Remember to configure the return_sequences parameter of your recurrent layers correctly based on whether you are using the final hidden state (False for the last layer) or pooling (True for the last layer). If stacking recurrent layers, intermediate layers must have return_sequences=True to pass the full sequence of hidden states to the next layer.
Bidirectional RNNs: For many classification tasks, especially in NLP, using Bidirectional LSTMs or GRUs (Chapter 7) can improve performance. By processing the sequence in both forward and backward directions, the model can form representations based on the context from both sides of any given element. The final forward hidden state and the final backward hidden state are typically concatenated (or sometimes averaged/summed) before being passed to the classification layer.

Examples in Practice

Sentiment Analysis: Given a sequence of words representing a product review, classify it as 'positive', 'negative', or perhaps 'neutral'. The RNN reads the review word by word, and the final state summarizes the overall sentiment expressed.
Topic Classification: Given a sequence of words from a document, classify it into one of predefined categories like 'Technology', 'Sports', 'Finance', etc. The RNN processes the document text to capture its main theme.
Intent Recognition: Given a sequence of words representing a user's command to a voice assistant (e.g., "What's the weather like tomorrow?"), classify the user's intent (e.g., 'GetWeather', 'PlayMusic', 'SetTimer').

Sequence classification is a powerful technique where the ability of RNNs to process ordered data and maintain state allows them to effectively summarize sequential information for categorization. By understanding how to structure the model architecture, particularly how to utilize the recurrent layer's outputs, you can build effective classifiers for a wide range of sequence-based problems.