The machines are watching and learning. From our news to our photos, machine learning is starting to influence our lives in subtle ways that will fundamentally change our relationship with information. To leverage and master this new technology, it is important to grasp the need for a critical review of these systems. Many will start to create experiences that will wrap our users in warm blankets, but those blankets may get too familiar.
It all starts with a model that takes input (preferences, behaviour, images, text, etc…) and renders a confidence at the likelihood of an outcome. Is that image Dan? Is the user interested in this content? Is the user likely to purchase? Many of these models are quite terrible when initially launched, but over time new data is collected, fed back into the training of the models, and they improve. Dan is definitely Dan when he is identified in pictures, the user interacts with a quantifiable percentage of more content, and users purchase at a more frequent rate. What an amazing technology.
News illustrates a very interesting use of machine learning. As we interact with content we’re training the model to become familiar with our preferences. As an example, I was following the Robbie Williams and Jimmy Page feud. That preference was recognised and more stories started to appear from a variety of sources. Then it was too much. Repetitive articles, articles showing up well after the story was ‘over’. Although it found a preference, the machine over indexed. Searching for G-Sync monitors began an influx of related articles to the topic, none too relevant besides trying to create buyer’s remorse after I already purchased one. Interest in the video game Anthem had alerts on my screen about crashes on PS4… But I play on a PC. Getting these types of signals, negative ones that should help train the model, are a real challenge.
How do you combat this? First, don’t just let the machines rule the product. Pairing this powerful technology with human guidance will be critical. Just because we gave the machine more data doesn’t mean it will work in the way we expect it to. Deploying a couple of models to audiences at scale and observing results should allow for a more educated plan around which model to continue to invest in. Need to go backwards? By all means, do!
Now, it would be unfair to say that the model is totally losing on the news front. The regular flow of articles around my favourite music and technical topics I’m interested in is spot on. New items on a regular basis, content from sources I appreciate. This is the warm blanket. It drives engagement from me on a regular basis. However, it is important that the model takes chances every so often and introduces content that I will most likely be interested in, but doesn’t match the core categories it has identified for me. This is where the combination of a personal model with a model generated from a larger audience could prove valuable. In other words, combining multiple models will result in higher probability outcomes. Tesla, as an example, is processing data for their self-driving features by piping that data through multiple models and specialised hardware.
The big players are going to continue to democratise the technology, allowing more product owners to experiment and include the tech in a reasonable amount of time with limited need for specialised data scientists. Microsoft, Amazon, Google, and Apple are all part of the puzzle. Apple in particular is where we feel the most opportunity exists. With iOS reaching one billion devices, the scale is unquestionable. Furthermore, Apple is building specific hardware optimised to process machine learning models into these solutions. They are also building some of the most user-friendly tools to train models, and we anticipate WWDC 2019 will reveal further advances in the space.
In closing, as products are created and roadmaps are planned, machine learning will be a default on the list, not a ‘nice to have’ feature. Mastering the deployment, training, and human refinement will separate the successful achievement of business goals from the failure to realise them.
Ben Reubenstein is CEO of POSSIBLE Mobile