Håkan Silfvernagel - Machine learning in the browser with TensorFlowjs

#1about 3 minutes

Understanding the fundamentals of machine learning

Machine learning is defined as pattern recognition in historical data, with supervised learning being a common approach for tasks like prediction and clustering.

#2about 2 minutes

Exploring the TensorFlow library and tensor data structures

TensorFlow is an open-source library that uses tensors, which are multi-dimensional arrays like scalars, vectors, or matrices, to perform computations.

#3about 5 minutes

Loading and visualizing car data with TensorFlow.js

A JSON dataset of car information is loaded and visualized as a scatter plot to identify the negative correlation between horsepower and miles per gallon.

#4about 10 minutes

Building and training a simple sequential model

A sequential model is defined, compiled with an optimizer and loss function, and then trained on normalized and shuffled car data to predict MPG.

#5about 6 minutes

Improving model predictions with additional layers

The initial linear model is improved by adding more dense layers to the neural network, which better captures the non-linear relationship in the data.

#6about 1 minute

Converting and using pre-trained Keras models

Existing models, such as a Keras H5 file, can be converted into the TensorFlow.js layers format using the command-line converter for use in the browser.

#7about 2 minutes

The benefits of running machine learning in the browser

Running machine learning on the client-side eliminates server roundtrips, enhances data privacy, and provides easy access to device sensors like cameras and microphones.

#8about 4 minutes

Building an image classifier with a pre-trained model

A web application is built to classify images by loading a pre-trained MobileNet model that has been converted for TensorFlow.js.

#9about 1 minute

Real-world applications of TensorFlow.js in production

Companies like Uber, Airbnb, and Google's Magenta project use TensorFlow.js for visual debugging, client-side document detection, and music composition.

#10about 2 minutes