Deep Learning + Model Analysis · Public Kaggle Experiments

MNIST CNN Experiments

Controlled Deep Learning Comparison

A controlled PyTorch comparison of CNN architectures for handwritten digit classification, focused on the effect of Batch Normalization on accuracy, loss behavior, runtime, and prediction quality.

RoleDeveloper

Year2025

StatusPublic Kaggle Experiments

MNIST CNN Experiments controlled Batch Normalization comparison cover

OVERVIEW

A controlled CNN experiment centered on Batch Normalization

This project combines two MNIST notebooks into one focused deep-learning case study. The first establishes a compact CNN baseline, while the second performs a controlled comparison between closely matched CNN architectures with and without Batch Normalization.

The comparison keeps the task and overall training setup aligned so the effect of Batch Normalization can be examined through accuracy, loss behavior, runtime, confusion matrices, and prediction examples.

EXPERIMENT STRUCTURE

98.84%Compact baseline CNN test accuracy

99.19%CNN without BatchNorm

99.37%CNN with BatchNorm

+0.18 ppAccuracy gain in the controlled comparison

CONTROLLED COMPARISON

Same classification task, one deliberate architectural change

01CNN without Batch Normalization

Convolutional architecture trained without BatchNorm layers to establish the comparison reference.

99.19% Test Accuracy

02CNN with Batch Normalization

BatchNorm layers are introduced while preserving the same task and comparable training conditions.

99.37% Test Accuracy

RESULT AT A GLANCE

The corresponding no-BatchNorm training time in the notebook is 69.54 seconds, so the accuracy improvement came with only a small runtime increase in this experiment.

BASELINE CONTEXT

A separate compact CNN experiment anchors the progression

The earlier MNIST notebook provides additional baseline context with a compact PyTorch CNN that achieved 98.84% test accuracy. It also includes feature-map inspection, confusion-matrix analysis, standard test predictions, and custom handwritten digit examples.

WHAT THIS PROJECT DEMONSTRATES

Controlled architecture comparison
Batch Normalization as an isolated experimental factor
Training-loss and runtime analysis
Confusion-matrix and prediction inspection
Feature-map visualization in a compact CNN

PROJECT GALLERY

Evidence from both completed MNIST notebooks

These figures are taken directly from the uploaded notebooks and show the data, model comparison, training behavior, feature maps, evaluation, and prediction examples.

Sample handwritten digit images from the MNIST training set — MNIST · Sample training digits

Accuracy comparison between CNN models without and with batch normalization — Controlled comparison · Test accuracy

Training loss comparison for CNN models without and with batch normalization — Training behavior · Loss comparison

Confusion matrix for the CNN with batch normalization — Best model · Confusion matrix

Example feature maps produced by convolutional blocks in the baseline MNIST CNN — Baseline CNN · Feature-map inspection

Sample MNIST predictions produced by the CNN with batch normalization — Best model · Sample test predictions

Predictions on custom handwritten digit images from the baseline MNIST CNN experiment — Baseline experiment · Custom handwritten digit predictions

INTERPRETATION

The value is in the comparison, not just the final percentage

MNIST accuracy is already high for compact CNNs, so the portfolio value of this project is the controlled experimental design. The notebooks make it possible to compare a standard CNN, a no-BatchNorm reference, and a BatchNorm variant while inspecting both quantitative and visual evidence.

TECH STACK

PythonPyTorchCNNBatch NormalizationMNIST

PUBLIC NOTEBOOKS

Explore both MNIST experiments on Kaggle.

The two public notebooks provide the compact CNN baseline and the focused Batch Normalization comparison that are combined in this portfolio case study.

BatchNorm Comparison Baseline CNN Notebook Back to top