Presentation: ML in the Browser: Interactive Experiences with Tensorflow.js

At Fast Forward Labs, we like to see ourselves as the link between academia and industry. Research Engineer have two main tasks. In the first part of our work, we research tools and technologies from academia and the goal is to focus on tech that makes sense for the industry within the next six months to two years. For each selected topic, we conduct in depth research and produce an accessible report for our clients. In addition to that, we build prototypes that communicate the ideas behind these technologies and provides insights on their use in practice. The second half of our jobs entails working with clients, to build and prototype machine learning solutions to their specific problems.

My research interests are at the intersection of human computer interaction and applied machine learning, and one of the sub-topics in this area focuses on how machine learning can make user interactions more interactive and engaging. ML in the browser, provides opportunities to craft rich experiences such as predicting the user’s next action, message auto-complete or even gesture-based interaction. Beyond interaction improvement, browser deployment enables other benefits such as privacy, ease of distribution and improved latency. By performing predictions on user data locally in the browser, you could lay claim to strong privacy because user data is never sent to remote servers.  Another really exciting benefit is the fact that the distribution of the model becomes a lot easier. For most ML developers, packaging and distributing an ML application can be a challenging process which involves the installation of drivers, libraries and other system specific dependencies. However, If you go ahead and do all of this in the browser, all of that installation and distribution hassle goes away. With Tensorflow.js, it is as easy as including a link to the Tensorflow.js library, and loading your ML model files where appropriate. In addition, there are also cases where, deploying a model in the browser can improve the latency of your application. For these situations, it can be faster to perform computations in the browser as opposed to the round trip of sending user data to a remote server which returns results. Of course, the caveat here is that your model needs to be relatively small for this to work well in practice.

It's an intermediate level talk, and the goal is to first get users excited about the possibilities of ML in the browser, and to show an end-to-end use case of how they can develop a model and get that deployed in a web application. This includes steps such as data collection, training a model, converting it into a format we can deploy in the browser and looking at runtime statistics. The talk will also address questions such as - Should you train a model in the browser? Do you train it offline? And if you train it offline, what are the challenges associated with converting it to a format where you can run it in the browser?

This is one of the things I will cover in this talk. With Tensorflow.js there are a couple of different flows that a developer can follow. Flow number one, you can collect user data and train your model from scratch in the browser. The user can even specify hyper parameters of the model such as the architecture,  the number of layers, building blocks, etc. In practice, this approach is suitable for small models which are trained on relatively small datasets. In the second flow, the model is trained offline and loaded for inference in the browser. This offline training approach allows you to take advantage of fast compute resources (GPUs, TPUs etc), train on large data sets, use complex models and then convert the trained model into a format that can be loaded by a JavaScript application. My talk and use case example will focus on this approach. The third flow follows a similar pattern where the model is trained offline, loaded in the browser but can now be retrained or finetuned in the browser using additional user data.

I am talking to an ML engineer, preferably one who is comfortable deploying machine learning models in backend applications in Python but has limited experience with browser use cases. The second target is the software engineer, perhaps even the front-end software engineer who wants to integrate ML models into front-end applications built with Tools like React.

I am hoping audience members will walk away with an understanding of use cases for ML in the browser, how to implement these use cases with the Tensorflow.js library, and best practices around deployment these models.