Mastering Image Classification: A Journey with Cakes

One model saw a cake and called it a soap dispenser. This is the story of building a better image classifier with transfer learning.

#1about 5 minutes

Building an image classification game inspired by "Is It Cake?"

The project goal is to build a game that can distinguish between real objects and hyper-realistic cakes using machine learning.

#2about 2 minutes

Sourcing cake and not-cake images for training data

Cake images were scraped from bakers' websites using Playwright, while non-cake images were sourced from the Unsplash API.

#3about 1 minute

Identifying limitations and biases in the image dataset

The collected dataset had issues like non-cake items (logos, biscuits) in the cake set and ambiguous images in the non-cake set.

#4about 3 minutes

First attempt using the MobileNet classification model

The pre-trained MobileNet model struggled, often misclassifying cakes as candles or bakeries due to its training on the ImageNet database.

#5about 2 minutes

Testing the Coco-SSD object detection model

The Coco-SSD model performed slightly better by identifying cake as a class, but still made significant errors like classifying a cake as a person.

#6about 6 minutes

Building a custom convolutional neural network that failed

A custom sequential model was built using convolutional layers, but it failed during training and classified every image as "not cake".

#7about 2 minutes

Achieving better results with transfer learning

Transfer learning leverages a pre-trained model (MobileNet) for feature extraction and adds a small, custom classification head for the specific task.

#8about 3 minutes

Playing the "Is It (F)ake?" game with the audience

A live demonstration of the game built with the transfer learning model shows the audience outperforming all the machine learning models.

#9about 1 minute

Key takeaways from the machine learning journey

The project highlights the value of experimentation and persistence, showing that even failed attempts provide valuable learning experiences in machine learning.