The conversation is past the point of whether or not artificial intelligence (AI) is impacting or is going to impact our everyday lives. That said, today the discussion still oftentimes revolves around expectations on reaching artificial general intelligence (AGI), including superintelligence, humanoids and alike. As hinted by Gartner, AI has already been close to the peak of the hype cycle for a good number of years. So, the question really is: how likely is it that we are actually going to deliver on the expectations on AI?
This post summarizes lessons learned in more than 100 real-world AI projects that I’ve been a part of. To answer the above question, we first need to understand what AI is and how it interrelates with human intelligence. The future of AI relies heavily on the co-existence and co-operation of the machine and the human. Hence the fundamental impact AI is having on most industries is happening through the interaction we have with existing technology.
Today’s AI is narrow and for specific problems
To start with, I think there is a fundamental issue with the definition of an emerging technology like AI. The way we define it is bound not to materialize. It is more a philosophy, an emerging “next thing” we can’t really do yet. However, once we do, it will actually stop being called “AI”.
To give an example, the most active discussion around AI today concerns assistants like Siri, Alexa, chatbots, self-driving cars, humanoid robots etc. These general assistants or systems with full autonomy rarely cater all needs and fulfills all expectations. Yet, already now we are every single day using AI-driven products, such as Google, Amazon, Netflix, Spotify and Facebook, among many others. The challenge is thereby not too high expectations, but wrong expectations.
Organisations shouldn’t be most concerned with or try to reach AGI, fully autonomous AI, or similar. In most cases, we simply can’t build that. What we can do is weak AI for specific and narrowly-defined problems. Yet, through this AI is nevertheless impacting most business processes, and changing organisations, the way we work and the society at large. And this significantly impacts the way value is created in organisations.
AI means machines that learn from examples
One of the most central technologies behind the uprise of AI is machine learning (ML), a field concerned with machines that learn without being explicitly programmed. Now, ML is frequently divided into three types of learning:
- supervised learning implies learning from examples (or labels),
- unsupervised learning implies learning with no examples, and
- reinforcement learning from rewards when taking actions in a given environment.
While media has extensively covered advances in unsupervised and reinforcement learning, we easily forget that most of this is either research or applied to problems that are unrealistic in most real-world settings. As an example, Google DeepMind’s AlphaGo is a remarkable breakthrough in applying reinforcement learning to the game Go with superhuman performance, and also applying the same approach to other games such as chess, Atari and more.
Wouldn’t it be impressive to see unsupervised learning to on its own start to comprehend problems and solve one after another? Unfortunately that is rarely what we are able to do in the real world. What we can see being applied successfully today is supervised learning, the task of teaching a machine with large volumes of labeled data. This is nothing else than a mapping function between ‘A’ and ‘B’, inputs and outputs. In the real world, we can identify cats in images, give personalized product recommendations, identify defaulting creditors, translate sentences to different languages, among many other examples.
Now, in the real world, there is typically one significant challenge: where to get good enough large-volume labeled data? This is where many thriving new fields in ML come to help, such as semi-supervised learning, active learning, transfer learning and reinforced learning. Accordingly, these thriving fields of ML should be a focus of research today as they will be crucial for the further development of (mostly supervised) learning algorithms.
Closed-loop data for model-driven business
In 2011, Marc Andreessen wrote a widely-cited piece in Wall Street Journal on software eating traditional industries. Looking at some of the most valuable companies today, his conclusions were not far from correct, such as Netflix, Amazon, Spotify, Skype, Pixar, Google and many more. This highlights the impact of the software revolution on several traditional industries.
What is going to happen to industries during the AI revolution? On top of the known dynamics of platforms, ecosystems and network effects, AI will leverage the business models of software to significantly further accelerate their value creation. And most of this relies on having access to closed-loop data by collecting from every end-user interaction inputs and outputs or eventually predictions and outcomes.
Companies that put learning models, based on closed-loop data, at the center of their business operations will thrive through virtuous cycles. Think of Netflix, Amazon or Spotify, for whom every single user interaction improves their models, which improves their products, which again attracts more users, and so forth. Your product, be it an internal efficiency tool or customer-interfacing app, accumulates competitive advantage from every interaction. Accordingly, building AI is not a one-off effort or a side-function outsourced to IT. Companies should rather focus on closed-loop AI at the core of their operations to enable model-driven business.
All that said, this is obviously nothing that comes with ease. To summarise, focusing on these three ingredients will position you at the forefront in the AI race.
- Closed-loop data. Do not only focus on large volumes of isolated data dumps, but focus on closing the loop between predictions and outcomes. You might want to start with simple predictions, recommendations or other signals, but make sure to also collect feedback. Interestingly, contrary to the software era, the new innovative startups are not necessarily best positioned to do this. For many incumbents, this is an opportunity to create models with their data and close the loop with existing processes or customers. This implies thinking about your existing business with an AI mindset. A good example is the traditional explosive manufacturing company Orica that transitioned into a scalable service by building a digital service for explosion optimisation by collecting feedback from every explosion.
- Build end-to-end capabilities. To jump on the bandwagon of AI, companies will need the people, processes and technologies required to create end-to-end machine learning. Similarly to when software became prevalent, organisations now need to develop, validate, deliver and monitor machine learning models. This doesn’t obviously mean that every company in the world should build their own AI team and tools. Certain infrastructure tooling is already becoming a commodity, while large ML teams have rarely been recruited and organically grown into large corporations (e.g. Google’s DeepMind acquisition). Oftentimes it might be worth partnering up, of which our AI & water-sector case is a great example.
- Acknowledge ethical responsibilities. More widespread use of ML and AI is going to raise significant issues around ethical and compliance challenges caused by the use and abuse of companies’ complete data assets. Their impact on human behavior is and should be a major concern not only in public discourse and for regulators, but also for companies themselves. Already now, you can observe markets pricing the abuse of data. As a prominent example, Facebook lost more than $100 billion of their market capitalisation in the aftermath of the data leakages related to Cambridge Analytica. Along these lines, Silo.AI has outlined their privacy and ethics policies in this video.
Human-in-the-loop solutions will take us there
AI is rarely about fully autonomous systems, ex-machina humanoids or even just end-to-end automated processes. I see it as “AI by people, with people, for people”, implying human experts building models, humans creating data sets and humans augmenting machines, respectively. These are often also known as human-in-the-loop solutions. Building out a solution is most often an iterative process that involves humans in the loop until one day, perhaps, machines are capable of taking over. The underlying reasoning behind the need for human involvement relates to three traits:
- Collect training data. To get started we desperately need training data, especially labels that enrich input data with certain meta data. There is no better way to access more data than getting started with modeling. This might relate to simply rule-based recommendations in labor-intensive processes, or even making use of ML strategies to collect human feedback for improved models. Thus, we need simple and quick ways to deploy models, and collect more data.
- Augment rather than automate. AI technologies are rarely mature enough for full automation, but instead they are helping humans. Especially when getting started with AI, we need solutions to augment and support human intelligence and decision-making. This points to the importance of solutions specifically designed for human-machine co-operation.
- People & processes. Beyond technological efforts, change also involves people and processes. That said, there is not necessarily a need to fully change existing processes. While ML models may help experts in their work, human experts have also the knowledge to help ML models. Thus, involve and find ways to help your human experts in their every day work, in order to assist both the human and the machine.
To exemplify, obvious cases of products collecting closed-loop data with a human-in-the-loop AI solution are Amazon, Netflix and Spotify recommendations as well as Google search. Likewise, many labor intensive service processes around documents and images or video could benefit from human-in-the-loop AI, such as automated document readers for contracts, investment documents, news or recruiting and automated image or video processing for visual quality control in production lines or applications to surveillance and security. Likewise, many standard machine learning tasks also require and benefit from interaction with human experts, such as fraud detection, predictive maintenance and water-quality management.
Along these lines, the four key messages in meeting expectations on and creating value with AI are as follows:
- The only AI that exists today is weak AI for narrowly defined problems.
- With only a few exceptions, real-world AI relies on supervised learning.
- Software is eating the world, yet model-driven business accelerates it.
- Human-in-the-loop AI allows solutions to improve with human intervention.