The AI Stimulus: Optimal is the New Functional
- Thoughtfully apply AI
- Design for adaptive experiences
- Move to agile machine learning
When software can infer the best possible answer from all the available evidence and continue to get better at it over time—which is what humans do reasonably well—then we call it artificial intelligence.
Now, algorithms that run on teraflops of compute power can sift through exponentially more data than humans to arrive at outcomes that are near perfect. Add learning-at-scale to an algorithm’s repertoire and it converges on an optimal state for some of the toughest problems. Theoretically, any system in this optimal state can become autonomous whether in network operations, industrial control, customer service or medical diagnostics.
There are a few dynamics that are speeding-up the adoption of AI. First, the cost of computationally complex operations is getting cheaper due to the proliferation of high-performance servers in the cloud. Second, bespoke algorithms are being democratized through full stack development tools and pre-trained algorithms in machine and deep learning. And third, cloud platforms are simplifying data acquisition, management and training methods.
Today, there are an estimated 1,800-plus AI companies that have captured over $16 billion in funding. Forrester predicts a 300% increase in AI investments in 2017 over 2016. And, IDC believes AI will be a $47 billion market by 2020.
While it’s undeniable that AI will find its way into just about every product, we believe that AI is a means to an end. To achieve revenue and margin growth, operational efficiency and market differentiation, the key question is: What is the problem you are trying to solve and how can AI play a role?
Thoughtfully apply AI
While AI appears to have all the answers, the benefits are determined by the business case. Domain knowledge brings the necessary focus to understand the customer value proposition. For example, an industrial-equipment maker will benefit from better predictions about the onset of failure so it can schedule maintenance in advance. Likewise, better predictions will help a mobile operator cut the time it takes to resolve network congestion and isolate instability. Or, they help a software developer speed up the release in the backlog of features by rooting out priority bugs much earlier.
There are a number of drivers to bring maching learning into products. Each will guide the way to a domain-specific use case that can have a positive impact on cost and revenue. Here are four:
- Better predictions. Companies can no longer passively wait for a major crisis, like an energy outage or a product malfunction, which can damage a brand’s reputation, nor can they rely on generic, impersonal user experiences. Making predictions that get better over time is one thing machine-learning products do very well.
- Faster Time to Outcome. Algorithms can quickly sift through massive volumes of data that humans or traditional computational models could never match. AI improves the performance of products by seeing patterns and making inferences at speeds that cannot be achieved by humans alone.
- Better human-centered interactions: AI is not about taking people out of the equation. Instead, it’s about designing the human-machine interface so it’s more natural, effortless and deeply personal.
- Learn and improve over time. The AI dividend is the learning feedback loop as a product or service gets better at what it does, such as raising the confidence of a prediction, more tailored personalization or more accurate sensing.
Without data, machines can’t learn
Data is the dynamic wireframe for machine-learning products. So, it’s critical to identify and collect real-world data and observations to train the algorithms that will drive decision making, actions and ultimately experiences. For example, Tesla asked permission from its customers to collect short video clips from cars to improve its Autopilot self-driving feature. By synthesizing this real-world customer data, a better algorithm can be identified, or developed and turned into an inference model—to make the correct turns on the road.
Typically, thousands of labeled examples are needed to train a model such as recognizing a car, a sign or a sidewalk using a machine-vision model. Disney Research has been working on a different way to recognize objects even if models have never seen a labeled example as part of the training. Disney’s approach comes in handy when the model is trying to figure out what’s happening in, say, a movie or surveillance video when an object in the frame is unknown and difficult to categorize.
To compete in the future algorithm economy, companies need to learn to fluently engage with data. A data-driven organization needs a product management process that is fine tuned for machine-learning products. The starting point is letting the development team research and experiment with trainable and pre-trained AI services. Innovation will then start to happen as the team gains experience with data acquisition, feature selection and algorithm development. There are numerous algorithm frameworks to choose from, such as Google TensorFlow, Microsoft AzureML, Caffe MXNet and others. And data integration and management tools help filter the mass of data they ingest by clearly defining and classifying the data, so they process only the comprehensible data.