Volant AI
Published in

Volant AI

Selling in our AI Model Marketplace

An AI model marketplace, like the one we are building at Volant AI, is fundamentally a marketplace. But a Deep Learning Model is not a traditional good. So, how can we protect sellers and give them an advantage?

Who we are

As we discussed in the previous article, we are building a hybrid AI model marketplace with Model-Selection-as-a-Service. It’s ultimately our goal to empower people to build with AI by utilizing high performing, crowdsourced models from our marketplace. We also see our mission deeply coupled with helping folks with data science expertise monetize their skills and invest in themselves and their future. If you want a bit more information, please take a look at this article.

What we are selling

Before we dive deep into the focus of this article, which is how we aim to protect seller rights, it is probably prudent to detail what exactly we are selling. It’s an AI model, right? Well…, yes. But, there’s a bit more.

We are selling a docker image that essentially serves a model as a REST API endpoint. It’s a generic image because we want to capitalize on the abstractions that blackbox-ing an API with Docker can provide. Here’s an embarrassingly simplified view of the process:

Super simplified view of the delivery of a model from B2C/B2B

We tell the seller of the AI model to upload a model asset (since we are accepting Keras models, this resembles a .h5 binary file), add some metadata and description and publish it to our marketplace. When a buyer obtains a model after payment, they obtain the generic docker image, supply some model identification information and are then authorized to run the model as an inference endpoint.

Our solution is inherently On-Prem with intention*. This allows any potential buyer to retain full ownership of their model usage so they can run it in their Virtual Private Cloud (VPC), locally for testing or wrap it under OAuth and REST APIs for external usage. There’s no additional cost with obtaining the docker image either.

*We’ll also offer models from 3rd party vendors with their own existing solutions, but our platform is not primarily focused on those type of sellers.

Protecting Seller rights

Now that it is clear what we are selling, it’s important to acknowledge that though a model is a type of “good”, it is not the same as one that you would find on Amazon. It’s not exactly a SaaS good either, because the fundamental asset is a static model asset file shrouded within a docker image. When a model is trained, there are a myriad of factors that affect its creation, and ultimately, its performance. These would generally fall under:

  1. Hyperparameters: factors affecting how a model is trained (tolerance, type of optimizations, etc.)
  2. Architecture: the structure of the network (number of layers, types of neurons, connections)
  3. Weights: activation coefficients, calculated through training + backprop

By some accounts, there may be overlap between the architecture and hyperparameters (the number of neurons in a hidden layer may be treated as a hyperparam), but we’ll maintain this distinction for now. Ultimately, the direct defensibility of a model comes from two things: 1) architecture and 2) weights. The caveat, is of course, that this is in the environment of our marketplace.

The distinguishing characteristics of models in our marketplace

We see better weights as implicitly representing better data. An architecture with trained weights (i.e. changed values post-backprop) should perform better than one with random weights. There should be statistically significant improvement of the model after training. In an effort to be more data-centric, we’ll even encourage making architectures public on our marketplace so sellers can be encouraged to build with better data and smarter techniques.

An AI model architecture snippet

Competitive advantage should be gained with better weights (and also a better architecture). The weights of your network are your most secret of sauces. It’ll never be shared because they will decide how successfully your models perform on the marketplace. As a seller, your goals should be to choose quality datasets, train smartly and evangelize your model’s quality on our marketplace.

The Future

We will continue learning as we continue on our two-fold goal of 1) empowering people to build with AI and 2) helping data scientists monetize their skills. If you are interested in a selling a model you have built in the past, then you should join our waitlist as we have started onboarding sellers into our marketplace.

Learn More.

As always, thank you for reading and feel free to share your thoughts in the comments. Please check out our page and give us a follow on our LinkedIn and Twitter.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store