Is it (F)ake?! Image Classification with TensorFlow.js

Can JavaScript tell a real object from a hyper-realistic cake? This talk charts the journey from failed models to a successful cake detector using TensorFlow.js.

#1about 3 minutes

Using JavaScript and ML to solve a baking show challenge

The speaker introduces the goal of using machine learning to identify hyper-realistic cakes from the TV show "Is it Cake?".

#2about 2 minutes

Collecting and balancing the cake vs not-cake dataset

Images of cakes and non-cakes are collected using Playwright and the Unsplash API to create a balanced binary classification dataset.

#3about 5 minutes

Evaluating pre-trained models for image classification and object detection

Pre-existing models like MobileNet and Coco-SSD are tested on the dataset, but they produce inaccurate and strange classifications.

#4about 6 minutes

Building a custom convolutional neural network from scratch

A custom convolutional neural network is built using TensorFlow.js sequential models and convolution layers, but it fails to accurately classify images.

#5about 5 minutes

Applying transfer learning to improve model accuracy

Transfer learning is used by combining a pre-trained MobileNet feature vector model with a custom classification head, significantly improving results.

#6about 4 minutes

Playing an interactive game to compare human and model performance

An interactive web game allows the audience to test their cake-spotting skills against the various machine learning models.

#7about 1 minute

Key takeaways and resources for getting started with TensorFlow.js

The talk concludes by summarizing the journey from using pre-existing models to applying transfer learning and provides resources for further learning.