CNN Explorer

1. Data & Preprocessing

The Nature of Images

To a computer, an image is just a grid of numbers (a tensor). Each pixel has 3 values (Red, Green, Blue) ranging from 0 to 255.

Why Preprocess?

Raw pixel values [0, 255] are often too large and uncentered for neural networks. Normalization subtracts the per-channel mean and divides by the per-channel standard deviation (CIFAR-10 stats: mean ≈ 0.49, std ≈ 0.25 per channel), producing values roughly in the range −2.1 to +2.5. This zero-centres the data and speeds up gradient descent.

Augmentation

We artificially expand our dataset by transforming existing images (flipping, cropping). This forces the model to learn invariant features, improving generalization.

1. Input Data

2. Preprocessing

Normalize

Mean

Std

3. Augmentations

Transformation Pipeline

Hover input...

Hover output...

Original (32×32)

→

Transformed Tensor

Channel Statistics

Ch	Min	Max	Mean

Input Data

Preprocessing

Normalize

Augmentations

ConvNet Architecture

Layer Parameters Layer 0

Kernel Size

Filters

Stride

Padding

Pooling

Activation

Input Image: Speed:

Calculation Engine

Output(x,y) = Σ (Input × Weight) + Bias

Hover/Drag input to see math...

Bias: 0.1
Result: 0

Playback

Layer Parameters

Kernel: Filters: Stride: Padding:

Navigation

The Classifier Head

Intuitive Explanation

The Fully Connected (FC) head takes everything the CNN has detected and considers it all at once.

1. Flattening / Pooling

Removes spatial structure, turning feature maps into a single list of evidence.

2. Global Reasoning

Each neuron asks: "How much do these combined features support a specific idea?" Weights decide which features matter and how strongly they vote.

3. Logits & Decisions

Logits are raw class evidence; the highest one wins after normalization. Bigger heads can model more complex combinations but are easier to overfit.

In short: the FC head is a global decision-maker that turns extracted features into class scores.

Classification Head (GAP → Dense → Softmax)

Params: 0

Configuration Modern Head

Input Feature Map

Architecture

Hidden Units: 12

Hyperparameters

Dropout: 0.38

Weights

Math: Dense Layer

a^[l] = g( W^[l] a^[l-1] + b^[l] )

Hover over a node or connection to see what it does.

GAP: 256 feature map averages fed as input. Hidden neurons combine inputs using learned weights. Output logits are class confidence scores. Connections carry weighted signals between layers.

Predict

Architecture

Hidden Units: 12