Building Your First Neural Network with TensorFlow, Classifying Handwritten Digits

2026-07-18 · Source: Deep Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Novice, long

Summary

This article details building a multi-layer neural network using TensorFlow and Keras to classify handwritten digits from the MNIST dataset. It explains how layers of neurons abstract features, starting from raw pixels to complex patterns like "eye-likeness" or "face-likeness." The practical guide covers loading the 70,000-image MNIST dataset, flattening 28x28 pixel images into 784-value vectors, and normalizing pixel values to a 0-1 range. It then demonstrates constructing a simple single-layer model, compiling it with `adam` optimizer and `sparse_categorical_crossentropy` loss, training for 5 epochs, and evaluating its accuracy. The article further shows how adding a hidden layer of 100 neurons with `relu` activation significantly improves performance, illustrating the power of network depth. A `keras.layers.Flatten` shortcut is also introduced for streamlined data preparation.

Key takeaway

For AI Engineers building image classification systems, understanding multi-layer neural network architecture is critical. You should begin by preprocessing image data through flattening and normalization, then incrementally add hidden layers to your Keras `Sequential` model to enhance feature abstraction and improve accuracy. Always evaluate performance on unseen test data and use confusion matrices to diagnose specific classification challenges.

Key insights

Multi-layer neural networks learn hierarchical feature representations, combining simple patterns into complex concepts for robust classification.

Principles

• Neurons specialize in detecting specific patterns through learned weights.
• Deeper layers combine outputs of previous layers for increased abstraction.
• Input normalization (0-255 to 0-1) enhances network training stability.

Method

Load MNIST data, flatten 28x28 images to 784-pixel vectors, normalize 0-1. Construct `keras.Sequential` model with `Dense` layers, compile using `adam` and `sparse_categorical_crossentropy`, then train and evaluate.

In practice

• Use `np.argmax` on prediction outputs to get the highest confidence digit.
• Visualize model errors with `tf.math.confusion_matrix` and `seaborn.heatmap`.
• Integrate `keras.layers.Flatten(input_shape=(28, 28))` directly into the model.

Topics

• TensorFlow
• Keras
• Neural Networks
• MNIST Dataset
• Image Classification
• Deep Learning
• Confusion Matrix

Best for: AI Student, Machine Learning Engineer, AI Engineer

Related on AIssential

• Python for Data Science & AI · Blog 18 of 20 — CNNs for Image Classification — Convolutional Neural Networks efficiently classify images by learning hierarchical features through local, reusable filters and weight sharing.
• TensorFlow From Basics to Advanced — A Beginner-Friendly Guide With Python Examples — TensorFlow simplifies deep learning development by providing tools for model building, training, and deployment, from basic tensors to complex architectures.
• KVNN: Learnable Multi-Kernel Volterra Neural Networks — kVNNs use learnable multi-kernel representations for efficient higher-order feature learning in deep networks.
• The Math Behind Neural Networks, Part 0 — Where Deep Learning Comes From — A neural network is a machine that learns its own features to approximate complex, unwritten rules.
• The Beginner's Guide to Neural Networks — Neural network learning hinges on weights and biases, adjusted via gradient descent and stabilized by batch processing.
• Hybrid ANN-SNN Pipeline with Local Plasticity — A hybrid ANN-SNN pipeline combines pretrained ANN embeddings and local SNN plasticity for high accuracy, bypassing end-to-end gradient training.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Deep Learning on Medium.