We’ve made improvements to our knowledge transfer algorithms and neural network architecture, so all your Custom Models have been automatically upgraded with greater accuracy. And, we’re now allowing you to build Custom Models using knowledge from more of our base models (not just our General model) with a new feature called ‘Base Workflows.’
Last year, we launched our Custom Training product which empowered our customers to create their own visual recognition models without having any background in machine learning. Since then, customers have built thousands of models for their businesses and passion projects, and we’ve been learning how to make Custom Training better from each and every one of you. That means we’ve been working hard on finding ways to improve Custom Training such that users are able to get more accurate prediction results with fewer labeled images.
So, without further ado, we’re excited to announce two new features that will make you love Custom Training even more.
Clarifai uses convolutional neural networks (CNN) for visual recognition. Using CNN, we have built a large collection of base models to classify objects in various domains. Building a base model requires millions of labeled inputs, hundreds of training hours on high-performance machines, and many experiments conducted by machine learning researchers. In the end, a base model is produced that is accurate enough to be trusted by our users.
Building custom models using the methodology defined above is not pragmatic for a non-AI company due to the data, time, and resource requirements. This is why we decided to use knowledge transfer from our base models to build the Custom Training product. Our knowledge transfer methodology uses the knowledge that was developed during the base model creation process and builds additional knowledge from all the feedback received from usage. This knowledge base is what allows our users to create an accurate custom model with just ten labeled images.
In our initial launch of the Custom Training product, we were using knowledge built from our initial base models. By combining the cumulative knowledge from all of our other sources, we have improved our Custom Training product to deliver higher accuracy models. The changes have been applied automatically to the platform, so starting right now, users should see a 19% increase in accuracy, or an average of 0.2 increase in both Precision and Recall for their custom models. Go ahead, check out your prediction results in your custom models!
Since we released Custom Training, we’ve been using our knowledge base from our most comprehensive public model – General. Our General model is still the most comprehensive general recognition model in the market, covering 11,000+ concepts trained over millions of labeled images. Therefore, knowledge transfer from this model generally works well for most custom models that are built on top of it.
However, we’ve found that sometimes users want to build a specific Custom Model that has a closer match to the knowledge base from another base model. For example, a user might want to build a Calorie Recognition Model from our Food model. We are now allowing our users to modify the Base Workflow* field to select a model, which is a closer representation of the model that they would like to build. By selecting a Base Workflow that is visually more similar to a user’s desired model, the accuracy of the custom model should increase, and the required number of labeled inputs should also go down!
*A Workflow is a new set of features for Clarifai that empower users to be more flexible with our API and feature set. Read more about Workflows here!