Module 6 - Sequences and Time Series

Most real-world data has an order. Log files, sensor readings, financial prices, user sessions, speech, text - the value of each data point depends on what came before it. Standard feedforward networks throw that order away. This module covers the architectures that don't.

What You'll Learn

Lessons in This Module

#	Lesson	Core Concept
01	RNNs and Vanishing Gradients	Hidden state, BPTT, why gradients vanish over long sequences
02	LSTM and GRU Deep Dive	Forget/input/output gates, cell state, GRU simplification
03	Seq2Seq and Encoder-Decoder	Context vector, Bahdanau attention, teacher forcing
04	Time Series Forecasting Patterns	Decomposition, walk-forward validation, deep forecasting
05	Temporal Convolutional Networks	Causal + dilated convolutions, receptive field, WaveNet
06	Anomaly Detection in Sequences	Point vs contextual anomalies, LSTM autoencoder, thresholds

Key Concepts at a Glance

The core problem: sequences have temporal dependencies - the model needs memory.

Three architectural answers:

• RNNs/LSTMs - recurrent connections carry state forward through time
• TCNs - dilated causal convolutions capture long-range dependencies without recurrence
• Transformers - attention over all positions simultaneously (covered in Module 9)

When each wins in production:

Scenario	Best Choice
Streaming / low latency	TCN or LSTM
Long sequences, parallelism	TCN
Variable-length translation	Seq2Seq + attention
Demand forecasting	Temporal Fusion Transformer
Real-time anomaly detection	LSTM autoencoder
Edge device deployment	GRU (fewer params than LSTM)