Skip to content
Arjit Kulkarni
03Applied ML

CIFAR-10 Transfer Learning

How much is a pretrained backbone actually worth?
Role
Solo — experiment design, training, analysis
Context
Personal research project
Stack
PythonPyTorchtorchvisionNumPy
Source
arjitkulkarni/CIFAR10_TransferLearning
01  The problem

A controlled PyTorch study measuring ImageNet-pretrained ResNet18 fine-tuning against a 330K-parameter CNN trained from scratch, under deliberate label scarcity.

“Transfer learning helps” is a claim everyone repeats and few quantify. The question worth answering is how much it helps, and specifically how the gap moves as labels get scarce — which only means something if the comparison is genuinely controlled.
02  Approach
  1. 01

    Held everything constant except the thing under test: the same stratified subsets, the same preprocessing, the same evaluation protocol for both models.

  2. 02

    Compared an ImageNet-pretrained ResNet18, fine-tuned, against a 330K-parameter CNN trained from scratch.

  3. 03

    Ran the comparison at two label budgets — 5% and 20% of CIFAR-10 — to measure how the advantage scales with data.

  4. 04

    Made the whole thing reproducible: fixed seeds, scripted dataset preprocessing, experiment tracking and JSON metrics written out per run.

03  Outcome

  • At 5% of labels: 89.2% test accuracy fine-tuned, against 50.9% from scratch.

  • At 20% of labels: 94.5% against 74.7%.

  • A quantified answer — the pretrained backbone is worth most exactly where data is scarcest.

Fine-tuned @ 5% labels
0.0%Fine-tuned @ 5% labels
From scratch @ 5% labels
0.0%From scratch @ 5% labels
Scratch CNN parameters
0KScratch CNN parameters

Concepts applied

  • Controlled experiments
  • Transfer learning
  • Reproducible pipelines
  • Experiment tracking
  • Evaluation design

Next project

04NyayMitra
All work