First neural network in TensorFlow - Model Pipeline Trace

Model Pipeline - First neural network

This pipeline shows how a simple neural network learns to classify handwritten digits from images. It starts with raw image data, processes it, trains a small neural network, and then makes predictions.

Data Flow - 5 Stages

1Load Data

70000 rows x 28 x 28 pixels→Load MNIST handwritten digits dataset→70000 rows x 28 x 28 pixels

Image of digit '5' as 28x28 grayscale pixels

↓

2Preprocessing

70000 rows x 28 x 28 pixels→Normalize pixel values from 0-255 to 0-1→70000 rows x 28 x 28 pixels

Pixel value 0 becomes 0.0, 255 becomes 1.0

↓

3Flatten Images

70000 rows x 28 x 28 pixels→Convert 2D images to 1D vectors→70000 rows x 784 columns

28x28 image becomes a list of 784 pixel values

↓

4Train/Test Split

70000 rows x 784 columns→Split data into training (60000) and test (10000) sets→Training: 60000 rows x 784 columns, Test: 10000 rows x 784 columns

Training set image vector and label for digit '3'

↓

5Model Training

60000 rows x 784 columns→Train neural network with one hidden layer→Trained model with weights and biases

Model learns to map input vectors to digit labels

Training Trace - Epoch by Epoch

Loss
0.5 |****
0.4 |***
0.3 |**
0.2 |*
0.1 |
    +---------
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.45	0.87	Model starts learning, accuracy improves quickly
2	0.30	0.92	Loss decreases, accuracy increases as model fits data
3	0.25	0.94	Model continues to improve with more training
4	0.22	0.95	Loss decreases steadily, accuracy nears 95%
5	0.20	0.96	Training converges with high accuracy

Prediction Trace - 4 Layers

Layer 1: Input Layer

Layer 2: Hidden Layer (Dense + ReLU)

Layer 3: Output Layer (Dense + Softmax)

Layer 4: Prediction

Model Quiz - 3 Questions

What does the 'Flatten Images' step do to the data?

A Converts 2D images into 1D vectors

B Splits data into training and test sets

C Normalizes pixel values between 0 and 1

D Applies activation functions to neurons

Key Insight

A simple neural network learns by adjusting weights to reduce error (loss) and improve accuracy. Flattening images and normalizing pixels prepare data for the model. Activation functions like ReLU and softmax help the network learn complex patterns and output probabilities.