Join Volant AI: The App Store of AI
Yes, the “App Store of AI” sounds like a nifty catchphrase. It’s more than just an AI model marketplace.
We’re trying to keep this article as a quick read. More details to come.
Check us out at https://volantai.org
First, some trends
It’s no secret that people are engaging with Artificial Intelligence/Machine Learning/Deep Learning more than ever over the past decade.
Thanks to the power of the internet and MOOCs, learning is much more democratized. Anyone can learn how to whip up a simple MNIST digit recognition model after going through one of the hundreds, if not thousands of tutorials online. A very popular platform to apply the data science and AI/ML skills is Kaggle¹ where folks can compete to create models that perform well with datasets. It has more than 1.5 million active monthly users engaging in its hands-on competitions.
There is a funny post I saw the other day that captures this overall sentiment:
Whether you agree with this (lighthearted) comment or not, there is no denying the the growth of a “shadow workforce” of passionate AI enthusiasts. This, coupled with a growing AI software market projected to be 125 billion dollars by 2025² illustrates a positive trend of engagement with AI by many. We’re hoping to leverage them to deliver high-performing AI/ML models faster and cheaper than ever.
Why we are doing this
There’s so much we want to say on this, but let’s keep it high-level for now.
The Problem: Cost + Expertise
We’re going to dive deep into in future articles, but the problem(s) in AI model development boils down to the following:
- Cost: Developing a deep learning model can be very time consuming and costly money-wise. This includes cost of training (compute) and timescale.
- Expertise: There is a lot of benefit to having a model developed by someone with domain expertise.
There are certainly necessary use-cases (particularly enterprise) wherein training a model for several months and spending thousands of dollars in people-hours and development is necessary. This really should not be compromised if an AI-driven product is the backbone of the company. However, for situations where there is higher margin for error — like in startup or SMB — it is important to not burn through money for computer or labor hours that could be used elsewhere.
Typically, assuming there is not a similar use-case model in your repository, this process involves reading up on State-of-The-Art (SOTA) papers, building or github-diving these models and tweaking them to work for you. Figuring out whether the model that you are developing actually works involves training the model, validating its performance with a validation set, tweaking the model’s hyperparameters and re-training, re-validating, etc.
Beyond the time cycle that training a AI model takes (massive neural nets with millions of parameters can take days), the cost of the actual infrastructure is unavoidable. Graphics Processing Units (GPUs) are the ideal weapon to train with and using these can incur drastic invoices, especially through a cloud service provider like AWS, Azure or GCP. Running multiple cycles of the train-val cycle can cost thousands of dollars in compute and cognitive time labor. It is generally a bad idea to “productionize” a model without any sort of tuning, though. This process is cumbersome and certainly necessary for high-stakes, enterprise-level contexts, but it’s not ideal for smaller businesses.
Domain expertise can be very difficult and expensive to come across. Many larger enterprises can afford to have their own AI shops that are poised to interact with cross-team engineers and Product Managers. They have success because they are able to leverage the expertise of different experts to make the model work. SMBs may not be able to do this. Hiring an AI engineer/Data scientist costs at least $100K if not more depending on expertise and location. What if you could have your own personal data scientist but would use it (and pay for it) when you need to?
Our solution: The Hybrid AI Model Marketplace + MSaaS
The Hybrid AI model marketplace refers to a marketplace that has both models that are not only domain-specific but also canonical models. We will have a taskboard where entities looking for specific models pertinent to their domain can write out some quick specs (think data + some description) and the best AI developers in the world can build a model that satisfies these requirements. The marketplace will also have models that are canonical; they will be reminiscent of typical use cases and problems (digit recognizer, sentiment analysis, image segmentation, etc.) seen throughout AI. We expect that buyers can find value in these models as well.
With our ecosystem, we’ll crowdsource the best quality models for you and your use-case(s). The models are pre-trained and we’ll display the performance in order to encourage transparent decision making. If you have an AI shop and are especially picky about performance, this marketplace will be a good starting point for hyper-parameter tuning. If you don’t have an AI shop or find the model satisfies your standards, use it out-of-the-box and integrate it easily with your stack.
The other aspect of our platform is Model-Selection-as-a-Service (MSaaS). This will essentially consist of a slew of APIs and tools that will help aid transparent decision making. It stems from the belief that though there are many factors that may affect your decision to buy a model from a marketplace, the most fundamentally important one really should be the performance on any associated datasets. Here’s a sneak peek of part of a subset of the features of MSaaS:
However, that isn’t all that MSaaS is. It also includes an algorithm that was developed during a research project where the expected validation performance of training an algorithm over a certain amount of time-steps was achieved in a little more than half of the time due to better data serving to the model. We’ll have a future article out about this that will condense the results of paper (keep an eye out for “Machine Teaching for Optimal Performance of Active One-Shot Learning”) and give more information about MSaaS.
Our mission
We’re not building just an AI model marketplace, we’re building a means of empowering and revolutionizing the general relationship with AI. Though we are focusing on enhancing businesses and products of tech companies and independent developers, we envision a future where anyone: the mom-and-pops flower shop, the bakery down the street or your local plumber can utilize and integrate AI to help them make better business decisions.
This isn’t going to be easy. We’ll be driving it, but we need like-minded, inspired people who are going to help us build this ecosystem and change mentalities. Whether you’re a developer, business or someone who truly believes in democratizing AI, we need you.
Join us.
Check our site out: https://volantai.org
Sign up for our waitlist. Get updates and keep in touch with us: https://volantai.org/waitlist
Please comment and let us know what you think of this article! Share with your friends and family! Our hope is to spark conversations.
Socials (follow us!)
linkedin: https://linkedin.com/company/volant-ai
facebook: https://www.facebook.com/VolantAIMarketplace
twitter: https://twitter.com/ai_volant
