CNNs & Time Series | Fakhruddin Khambaty's Learning Hub

What Is a CNN?

👶 In Simple Terms

A Convolutional Neural Network (CNN) is built for grid-like data (images, or 1D sequences). It uses convolution layers that slide small filters over the input to detect edges, textures, and patterns. Then pooling reduces size; after several such blocks, the result is flattened and passed to dense layers for classification. So: convolution → pooling → … → flatten → dense → output.

MNIST: Handwritten Digits

MNIST is a classic dataset: 28×28 grayscale images of digits 0–9. We build a small CNN in Keras to classify them.

import tensorflow as tf
from tensorflow.keras import layers, models

# Load MNIST
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
x_train = x_train.reshape(-1, 28, 28, 1).astype('float32') / 255.0
x_test = x_test.reshape(-1, 28, 28, 1).astype('float32') / 255.0

# Simple CNN: Conv -> Pool -> Conv -> Pool -> Flatten -> Dense -> Output
model = models.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),
    layers.MaxPooling2D((2, 2)),
    layers.Conv2D(64, (3, 3), activation='relu'),
    layers.MaxPooling2D((2, 2)),
    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
model.fit(x_train, y_train, epochs=5, validation_data=(x_test, y_test))

Time Series with Deep Learning

Time series = data ordered in time (e.g. daily sales, temperature). We can use:

1D CNNs: Sliding filters over the time axis to capture local patterns.
RNNs / LSTM: Recurrent units that remember past values; good for long sequences.
Transformers: Attention over time steps (used in many modern models).

Typical steps: load sequence data, create windows (e.g. last 30 days → next day), normalize, build a model (e.g. LSTM or 1D CNN), train and predict. Libraries like TensorFlow/Keras and PyTorch support RNN/LSTM and 1D Conv layers for time series.

🚫 Common Mistakes: CNNs & Time Series

Not normalizing image or sequence data — Scale inputs (e.g. 0–1 for images, z-score for time series) or training can be slow or unstable.
Data leakage in time series — Never use future data to predict the past; split by time (train on past, test on future) and build windows only from past values.
Using 2D conv layers for 1D sequences — Use Conv1D for time series; Conv2D is for images (height × width).

💭 Short reflection

In one sentence: why do we use Conv2D and pooling for images but LSTM or 1D CNN for time series?

✅ CORE (Must know)

CNN for images: Conv2D (spatial filters), MaxPooling, Flatten, Dense; good for local patterns.
Time series: 1D CNN (sliding over time) or RNN/LSTM (memory over sequence).
Create windows: past N steps → next value; normalize; train and predict.

📚 NON-CORE (Good to know)

Transformers for sequences; attention; data augmentation for images.

Previous: Neural Networks Next: NLP Basics