Distribution & Delivery Global Viewpoint – August 2020

At first, it might be tempting to think of ML as a form of programming or statistical analysis, but in essence it’s much more interesting than that and falls under the heading “data-driven learning”, that is, the data is the program.

Fundamentally, and very simply, the act of teaching an ML engine is to present it with as many pre-classified datasets as possible so it can learn to recognize certain patterns. The more datasets we present then the better the ML engine becomes at recognizing patterns it’s never seen before. These patterns can be anything from human faces to lions, or English spoken words, or predicting packet congestion in an IP stream.

To me, this is incredibly exciting as anybody who has programmed a computer knows they will spend a great deal of their time analyzing data to find patterns and then creating hundreds, if not thousands of if-then-else statements in the hope of catching and modelling the data. But with ML, we’re presenting the data to the engine and leaving it to work out the patterns for itself.

Clearly, there are applications that will benefit directly from video and audio object recognition. For example, using object recognition would help with automatic metadata generation when ingesting library material to an archive, or detecting spoken words to create subtitles.

So, what does this mean for broadcasting?

It’s my view, that due to the importance of data a new type of engineer will emerge in addition to the broadcast and IT engineer, we will soon be seeing the data-driven-learning engineer. They already exist in some companies in the guise of data scientists as there are vendors in the broadcast arena who are conducting ML and data-driven research, but as ML becomes more mainstream, data-driven-learning engineers will soon sit alongside broadcast and IT engineers in broadcast facilities, and dare I say, even take over from traditional software developers.

Also, anybody specifying products will need to be much more aware of how vendors are designing their ML solutions and more specifically, the type and quality of datasets they’re using. The more diverse and expansive the training-dataset, the better chance the ML engine has of both detecting a new pattern and providing a more accurate result. The whole point of ML is that it can detect patterns that it has not specifically been trained on (that is, not seen before). To achieve this, the training has to be immense. Just like a human.

Broadcast television seems to have gone through an incredible journey over the past twenty years. We’ve gone from analogue, to digital, to HD 4K 8K, to MPEG compression, to satellite, to OTT, and now vendors are actively developing ML systems that will teach themselves and further improve our broadcast facilities. But it’s my view, that as engineers, we must dig deep beneath the hype and really understand what ML is really about, and I truly believe that training-datasets will be a differentiating factor between vendors.