Double Deep | Devpost

Double Deep: AI, for AI.
Who we are. Watch our live demo for more information!
Landing Page.

Inspiration

A.I. is like an iceberg. It can yield great results, yet demands a mounatin of computation and fine tuning to produce satisfying results.

The problem is because some AI algorithms, notably neural networks, are black boxes. Little is known on why they perform so well. In addition, methods such as transfer learning that aim to use pretraiend neural networks to solve are limited, because a heavily trained model is refitted on a completely different data problem.

Instead of relying on trial and error to solve data problems, we developed Double Deep, a web-based algorithm that can effectively and accurately craft a customized model that fits each unique problem.

What it does

AI, for AI. DoubleDeep crafts a custom neural network model for a specific problem. And this is possible with only a click of a button.

How we built it

We used Python's Scikit-Learn machine learning package to build a model that can takes as input quantitative values of the complexity of the problem and predicts the optimal architecture of a neural network. UsingKeras (with a Tensorflow backend), we recorded the architecture of a specific neural network and its associated loss. Having recorded several such neural networks, we trained our Scikit-Learn model to recognize patterns in the neural network topology.

Challenges we ran into

This project consisted of discovering a completely new method to optimize AI. We had to map out an algorithm that can measure the performance of a neural network and suggest ways to optimize its structure. We did not know if such an approach would yield significant results as no similar approach has ever been attempted. After gathering enough training data, the results positively showed that a Double Deep generated neural network had a lower loss than a baseline model.

We also encountered problems correctly rendering a website using Flask and then deploying it to Azure. We learned different techniques using Flask, Git and Azure to convert our bugs into successes.

Accomplishments that we're proud of

We are proud, as a Sec. V student and a CÉGEP 2 programmer, having taken no actual university Computer Science courses, to have developed a novel method for an AI to build itself an AI that can be swiftly customized for a specific problem.

What we learned

We learned to not be afraid to dream big and dare new things that have never been done before.

We learned to deploy a Python app into production on a web hosting service.

A big thank you to Microsoft for providing us with access to Azure Cloud for web hosting, as well as the dedication and support to the ImplementAI organizers!

What's next for Double Deep

Extending the possible outputs of neural network architectures! Currently, Double Deep specializes in feed forward neural networks best suited for quantitative classification and regression problems. We aim to include future tests with more varied AI models ranging from image recognition to time series, which would allow Double Deep to craft potentially undiscovered cutting-edge models to solve data problems.