Back to Blog

Introducing Masterful AI

Image of Tom Rikert
Tom Rikert

Today we’re thrilled to introduce Masterful AI - a smarter, more automated way to build machine learning models.  We believe that we’re at a unique point in history when we’re beginning to entrust machines with decisions that impact billions of people’s lives and livelihoods - from analyzing medical images, to driving cars, to running a manufacturing line.  Yet the process to build the models powering these advances is surprisingly primitive.  Machine learning may look like the space age on the outside, but it’s really the stone age on the inside.  Behind the scenes, there are armies of people manually labeling training data.  These armies then hand the data over to developers who manually run countless experiments to build a production-ready model.  Our mission at Masterful is to bring the power and efficiency of modern software development to ML.  The AutoML platform we're announcing today supports this mission by reducing labeling and shortening the time to achieve a great model.  We do this in two ways:  1) we use unlabeled, augmented, and synthetic data to improve your model and 2) we automatically test and tune your training loop.

street segmentation 2

While modern software development has a mature “build and test” workflow, creating production-ready ML models still demands months of frustrating experimentation with data labeling and model training.  Yet, if you don't build an accurate model, there are dire consequences.  Self-driving cars kill people.  A tumor in a medical scan is mis-diagnosed.  Biased content recommendations increase polarization.  Fire hazards near powerlines are overlooked. 

We’ve experienced first-hand how hard it can be to take models out of the lab and make them perform in the real world.  Our founding team has AI research roots at MIT, Stanford, and Google, and has shipped products to millions of users at Microsoft and YouTube.  We've built our experience into the Masterful AutoML platform, which forms the engine of a high-performance Deep Learning pipeline.  The platform we're announcing today removes the major inefficiencies and endless experimentation from ML development - enabling any engineer to build more accurate models, faster.  

architecture_semi

 

We’ve started by focusing on the most archaic and error-prone aspects of ML development: getting accurately labeled training data.  While some labeled data is necessary, Masterful’s goal is to reduce the amount of labeled data required to train a model by 10x.  This eliminates the time delays, cost, and potential for bias inherent in human labeling.  We do this by using small amounts of labeled data, combined with unlabeled, augmented, and synthetic data that can be managed by code instead of people.

Next, we're radically accelerating model development by getting rid of the manual experimentation required to test different datasets and training algorithms.  Instead, Masterful automatically tests and evaluates a suite of training algorithms and hyperparameters.  It then produces an optimal training policy, as well as a visualization of what was chosen and why.  

Integrating Masterful only takes a few lines of code with our simple API.  Unlike other AutoML platforms, Masterful is transparent.  You retain full control over your data, model, and development environment.  You can train with your model, your training data, run on-prem or in your private cloud, and measure with your preferred metrics. 

 

Masterful_API_nopadding

We’ve already delivered results for our customers across a range of industries using computer vision models.  For example, Masterful has improved quality analysis models for manufacturing, models that detect sensitive documents on corporate networks, and models analyzing aerial imagery for insurance underwriting. On average, Masterful can reduce model error by at least 40%, and shorten time to deployment by up to 50%.  

Interesting in seeing the results we can deliver for you?  To get started, you can get a free report here showing what Masterful can do for your model.  We’d also love to hear from you - your ideas, feedback, and stories of the ML applications you’re able to bring into production with Masterful. Share with us at learn@masterfulai.com or on Twitter at @masterful_ai.  Join us in building the future of ML.


Related Posts

Four buckets of deep learning innovation: architecture, data, regularization, and optimization

Image of Yaoshiang Ho
Yaoshiang Ho

Intro

It’s hard to stay current and maintain competency in deep learning. It’s a young and fast...

Read more

Avoiding Three Types of Geospatial Data Leakage

Image of Jack Lynch
Jack Lynch

 

Read more