Mark Kurtz

Mark Kurtz

Senior Full Stack Engineer

At Burst we often receive questions about how our artificial intelligence team (which we affectionately refer to as our Smarts Team) competes or compares with the big tech companies. These are more than fair questions, especially given the current hype in the industry, and almost every week a new story breaks about how Google, Facebook, or Amazon’s new AI tech will soon be replacing jobs.

With shadows that big, it is hard to see how small companies ever have any hope of seeing a ray of light again in terms of making a significant impact. For us, it’s tough to even see the shadow because we are sunbathing on the other side of the AI globe sipping proprietary data and surrounded by domain expertise and actual business needs. These principles guide us to state of the art AI solutions that enable our customers to look at 30 highly rated media instead of sifting through 3,000 photos/videos to find the gold. As a result of these enhanced workflows and efficiencies, the best media is more easily presented and helping our customers drive engagement becomes the focus.

Dig through the hype

Given all of the successes in AI, it is important to keep track of a couple failures in the field to keep us grounded. For example, chatbots were promised to disrupt industries over two years ago, but instead of this coming to fruition, Facebook disbanded its chatbot project and chatbots are only now starting to return in more constrained settings. Similarly, self-driving cars, from Tesla to Google, are well behind the pace the creators envisioned.

Why the delays? AI can be very hard, and true general AI is still decades away. In the future, it might resemble the current state of the art neural networks or hidden Markov models, but in the end, true general AI will be very different from what we know and use today.

The lack of general AI is what allows the Burst algorithms to be proprietary and is why we do not consider the big tech companies competitors. It takes a lot of effort to build AI solutions, even when the conditions are right, and the algorithms do not work in our domain without our data or unique knowledge of the industry. The business needs and specific use cases we encounter give us a competitive advantage in making headway on these problems that translate to immediate value for our customers.

Dive into Burst

Overall, our process on the Smarts Team has been to identify customer needs and use that information to frame our thinking through state of the art research papers and brainstorming sessions. We leverage millions of data points that we have collected over the years, apply specific domain knowledge, and are able to be agile with frequent product releases that provide value to our customers.

One of our recently released products is Unsupervised Media Curation — an algorithm to help our customers quickly discover the best media from all of the submissions they have received. This algorithm is born out of a simple assumption that has held up well to analysis — approvable media (media that our customers deem worthy to be featured on their broadcast, digital and social channels) looks similar and rejectable media looks similar, allowing us to score new uploads based on previously made media decisions. For many of our customers that are acquiring thousands of photos/videos around big storms, breaking news, and events, quickly being able to identify the best content from this new, sortable column in our product is integral for workflow efficiencies and driving engagement. Overall, we have seen great results with upwards of 80% accuracy for the top-ranked and bottom-ranked media.

Unsupervised Media Curation

Two new products that we are excited to release are the Image Quality Network and Video Quality Network. These are deep neural networks that we have trained based on media ranked 1 through 10 by humans. The models react to not only low-level features such as noise, color balance, and exposure levels, but also very high-level features such as foreground, sharpness, composition, etc. We have also seen great results here with the model’s score consistently being less than 0.5 points away from how humans rated the media. The initial business application for this algorithm will be to automatically determine the “best” approved media to present in our new gallery product. This will enable our clients to build more engagement with consumers via media views, click-throughs, and revenue opportunities.

Image Quality Network media ranking

Creating our own path

To answer the initial question, we don’t try to compete with big tech companies at Burst. Instead, we carve out new pathways that enable our customers to succeed in creating workflow efficiencies, driving UGC engagement, and creating new revenue opportunities. Our unique approach, data sets, and proprietary algorithms set our clients apart with state of the art UGC solutions. Overall, it is a very exciting time to be in the AI field in general, and certainly for Burst, so please check back for updates soon and thank you for reading!

Before you go, here is a quick glimpse of what’s to come from our Smarts Team:

  • Media Searching
    • Powered by an object detection deep learning network that we have tweaked for our dataset and serve at a fraction of the cost of comparable implementations
  • Deep Clustering
    • Powered by a deep learning network and clustering algorithm
    • Used to find similar media both for Unsupervised Media Curation as well as displaying similar media to the user
  • Context Relevance Network
    • Powered by word embeddings and object detection
    • Used to tell if a media belongs within a given domain and rank the media accordingly. Ex: A picture of a cat most likely should not be in a sports domain
  • Benchmark Quality Network
    • Powered by a deep learning network
    • Creates ideal representations of what customers have approved so new media that matches can be quickly surfaced