CNN – STEP BY STEP NUMERICAL EXAMPLE

We will cover:

1. Input image

2. Convolution

3. ReLU

4. Max Pooling

5. Flattening

6. Fully Connected Layer

7. Softmax Output

Step 1: Input Image

Assume a 4×4 grayscale image:

1 2 3 0
4 5 6 1
𝑋=[ ]
7 8 9 2
0 1 2 3

Step 2: Convolution Operation

Filter (Kernel): 2×2

1 0
𝐾=[ ]
0 −1

• Stride = 1

• Padding = 0

Output size:

(4 − 2 + 1) × (4 − 2 + 1) = 3 × 3

Convolution Calculation
Position (1,1):
1 2 1 0
[ ]⊙[ ] = (1 × 1) + (2 × 0) + (4 × 0) + (5 × −1) = 1 − 5 = −4
4 5 0 −1

Position (1,2):
2 3
[ ] → (2 × 1) + (3 × 0) + (5 × 0) + (6 × −1) = −4
5 6

Continue similarly.

Feature Map
−4 −4 3
𝑍 = [−4 −4 4]
7 7 6

Step 3: ReLU Activation

𝑅𝑒𝐿𝑈(𝑥) = max⁡(0, 𝑥)
0 0 3
𝐴 = [0 0 4]
7 7 6

Step 4: Max Pooling (2×2, stride = 2)

Take maximum value from each 2×2 region.

0 0 3 4
[ ] → 0[ ]→7
0 0 7 7

Pooled Output

𝑃 = [0 7]

Step 5: Flattening

Convert matrix → vector:

𝐹 = [0, 7]
Step 6: Fully Connected Layer

Assume 2 neurons, weights and bias:

0.5 −0.5
𝑊=[ ] , 𝑏 = [1, − 1]
1.0 0.5

Computation

Neuron 1:

𝑦1 = (0 × 0.5) + (7 × 1.0) + 1 = 8

Neuron 2:

𝑦2 = (0 × −0.5) + (7 × 0.5) − 1 = 2.5

Step 7: Softmax (Classification)

𝑒 𝑦𝑖
𝑆𝑜𝑓𝑡𝑚𝑎𝑥(𝑦𝑖 ) =
∑𝑒 𝑦
𝑒 8 = 2981, 𝑒 2.5 = 12.18
2981
𝑃(𝑐𝑙𝑎𝑠𝑠1) = ≈ 0.996
2993
12.18
𝑃(𝑐𝑙𝑎𝑠𝑠2) = ≈ 0.004
2993

Final Output

Class Probability

Class 1 99.6%

Class 2 0.4%

Predicted Class = Class 1

Summary

• Convolution → Feature extraction

• ReLU → Non-linearity
 • Pooling → Dimensionality reduction

• Flatten + FC → Decision making

• Softmax → Probability output

CNN OUTPUT SIZE – STEP BY STEP PROCESS

We will understand output size for:

1. Convolution layer

2. Pooling layer

3. General formula

4. Complete CNN flow example

1. Output Size of CONVOLUTION Layer

Formula (Very Important)

𝑁 − 𝐹 + 2𝑃
Output Size = +1
𝑆

Where:

• 𝑁= Input size

• 𝐹= Filter (kernel) size

• 𝑃= Padding

• 𝑆= Stride

Example 1: Simple Convolution

Input image: 5 × 5
Filter: 3 × 3
Padding: 0
Stride: 1
5 − 3 + 2(0)
+1 =3
1
Output = 3 × 3

Example 2: With Padding

Input: 5 × 5
Filter: 3 × 3
Padding: 1
Stride: 1
5 − 3 + 2(1)
+1 =5
1

Output = 5 × 5
This is called “same convolution”

Example 3: With Stride

Input: 7 × 7
Filter: 3 × 3
Padding: 0
Stride: 2
7−3+0
+1=3
2

Output = 3 × 3

2. Effect of NUMBER OF FILTERS (Depth)

Very important concept

If:

• Input = H × W × D

• Number of filters = K

Then:

• Output depth = K

Example
Input = 32 × 32 × 3
Filters = 10 (size 3×3×3)

Output = 30 × 30 × 10

Each filter creates one feature map

3. Output Size of POOLING Layer

Formula
𝑁−𝐹
+1
𝑆

(No padding usually)

Pooling Example

Input: 4 × 4
Pooling: 2 × 2
Stride: 2
4−2
+1=2
2

Output = 2 × 2

Depth remains SAME

4. Full CNN Output Size Flow

Given CNN Architecture

Input Image: 32 × 32 × 3

Conv Layer

• Filter = 5 × 5

• Stride = 1

• Padding = 0

• Filters = 6
 32 − 5 + 0
+ 1 = 28
1

Output = 28 × 28 × 6

Max Pooling

• Pool = 2 × 2

• Stride = 2
28 − 2
+ 1 = 14
2

Output = 14 × 14 × 6

Second Conv Layer

• Filter = 3 × 3

• Stride = 1

• Padding = 0

• Filters = 16
14 − 3
+ 1 = 12
1

Output = 12 × 12 × 16

Flatten Layer

12 × 12 × 16 = 2304 neurons

Convolution:

(𝑁 − 𝐹 + 2𝑃)/𝑆 + 1

Pooling:

(𝑁 − 𝐹)/𝑆 + 1
Depth:

• Conv → Number of filters

• Pool → Same as input

The output size of a convolution layer depends on input size, filter size, padding, and stride. It is
calculated using the formula
(𝑁 − 𝐹 + 2𝑃)/𝑆 + 1.
Pooling layers reduce spatial dimensions but keep depth unchanged.

STEP 4: MAX POOLING (CNN)

What is Max Pooling?

Max Pooling:

• Reduces height & width

• Keeps important (maximum) features

• Makes CNN faster & robust to small shifts

It works independently on each feature map

Inputs to Max Pooling

You must know:

• Input feature map size

• Pooling window size (F × F)

• Stride (S)

(No weights, no learning here )

1. Output Size Formula (Very Important)

𝑁−𝐹
Output size = +1
𝑆
Where:

• 𝑁= Input size

• 𝐹= Pool size

• 𝑆= Stride

2. SIMPLE NUMERICAL EXAMPLE

Input Feature Map (4 × 4)

1 3 2 4
5 6 1 2
𝑋=[ ]
4 2 8 0
1 3 2 6

Pooling Parameters

• Pool size = 2 × 2

• Stride = 2

3. Calculate Output Size

4−2
+1=2
2

Output = 2 × 2

4. Perform MAX POOLING (Step by Step)

Region 1 (Top-Left)
1 3
[ ] ⇒ max⁡ = 6
5 6

Region 2 (Top-Right)
2 4
[ ] ⇒ max⁡ = 4
1 2

Region 3 (Bottom-Left)
 4 2
[ ] ⇒ max⁡ = 4
1 3

Region 4 (Bottom-Right)
8 0
[ ] ⇒ max⁡ = 8
2 6

5. Final Max Pooled Output

6 4
𝑃=[ ]
4 8

6. Key Observations (Concept Check )

Property Result

Height & Width Reduced

Depth SAME

Parameters NONE

Training NOT learned

Pooling is fixed, no weights, no bias.

7. Max Pooling with STRIDE = 1

Input = 4 × 4

Pool = 2 × 2
Stride = 1
4−2
+1=3
1

Output = 3 × 3

(Overlapping pooling)

8. Multiple Feature Maps Case

Input = 6 × 6 × 3
Apply pooling separately on each channel:

Output = 3 × 3 × 3

Depth unchanged