indico.io: machine learning as a service

With the growing excitement around machine learning technologies, I have connected with Gideon Wulfsohn who works with indico to understand better what they do.

0) A couple of sentences about indico

• indico is helping individuals, small to medium sized teams and businesses translate their community’s pictures, documents and conversations into insightful feedback in minutes. Built with real life data and tailored to what you need, our pre-trained models balance accuracy and speed, allowing you to use powerful machine learning in realistic settings.

1) What is the indicio business model and user license?

• The model is to build API endpoints that allow developers to rapidly prototype and deploy solutions within a predictive application. We love chatting with our users and are always looking to improve what we offer to better fit their needs.
• Staying up to date with the latest research papers is a huge part of our development process, thus ensuring all models are tuned to industry standards.
• We also have a private cloud offering for enterprise.

2) Which underlying deep learning toolkits do you use?

• As Python continues to gain steam within the Data Science Community, Theano has stepped up as our de facto numerical computing library. Theano is designed to evaluate mathematical expressions fast with dynamic C code generation.
• We have developed abstraction layers for interfacing with Theano for building models using convnets and RNNs. We have open sourced the RNN abstraction layer under the name Passage

3) Do you support GPU computation?

• Thanks to the Theano, all of our models are runnable on either a CPU or a GPU.

4) What tricks do you use when building models using deep learning?

• We are always looking for ways to speed up computational time. Here are a couple tricks:
• Remove all redundant computation in sliding window classification strategies
• http://arxiv.org/abs/1412.4526
• Filter compression
• http://arxiv.org/abs/1404.0736
• Training a smaller model to emulate a bigger model
• http://arxiv.org/abs/1412.6550

5) What is your target user. Do I have to be a deep learning expert? programmer? business analyst?

• The target user is anyone with a bit of programing/problem-solving chops looking to answer a burning question. This often comes in the form of developers, entrepreneurs, hackathon goers, and small-medium sized businesses.

6) Which programming languages do you support?

• Java, Javascript, Objective C, PHP, Python, R, Mashape, and Ruby

7) What is the typical dataset size where you find deep learning to be effective on. How many images?

• As a rule of thumb, 100,000 examples is a good starting point for training a model from scratch.