Fifty years of computer intelligence development and we are essentially nearing the processing level of a three year old, though in the past five years we've achieved more advancement in this area than we have in the past fifty years. This is all due to the rise of activity in machine learning and what is fundamentally changing how we deal with data and think about insights.
We've moved from creating rule based systems to systems that can learn from data. Machine learning, deep learning, SVM, neural networks, all provide new opportunities. Similarly, the abundance of readily available tools such as R and Python libraries to accomplish processing and analysis not possible just a few short years ago.
For most people, it is a difficult concept to get their mind around though a number of recent developments have started to emerge that highlight the power of this approach. Namely, Tesla and it's efforts around autonomous vehicles. The idea of cars driving themselves is something most individuals can grasp and many may be either fearful or impressed with the concept. Not surprisingly, much of the discussion around artificial intelligence (AI) has been around automation or building autonomous systems that can perform tasks that humans do today. The primary difference with this concept from that of automation in the industrial revolution is back to the core difference of these 'machines' being able to learn as they automate processes.
These learning systems typically work by learning from numerous examples, much like the puppy example in the image above. The more examples, the more attempts, the rate of success increases. This is much how the human brain learns, thus the effort is to build systems that mimic how the human brain works. AI systems can learn either by using techniques to discover patterns in data, or through a process called reinforcement learning. Reinforcement learning is based on the concept on how humans learn, through vast examples of trial and error and establishing a reward system when there is success.
One of the core challenges as a result of these techniques is to have a sufficient number or recurring number of features or examples for the machines to process, find patterns and learn from. Furthermore, understanding the breadth of options to explore becomes increasingly important for reinforcement learning.
Humans have an incredible ability to learn quickly, though are relatively inefficient in being able to transfer knowledge from one entity to another. One of the key benefits of a learning system, is that the intelligence can easily be replicated and applied to other system entities. For example, autonomous cars require the processing of millions of miles of driving, though the true benefit is the accumulated learning across many vehicles and not a single vehicle. Similarly, voice recognition systems such as Siri depend on the collective knowledge to improve it's accuracy.
The true success of AI may indeed depend on the ability to foster the availability of data features and examples for processing. I believe this is where the human assisted element plays an important role. As companies and industries start to adopt these technologies, considerable effort is required to navigate the data elements available for processing and building a reliable data pipeline, even before any machine learning can occur.
One of the richest sources of data for processing is from humans themselves. Increasingly it will become critical for systems to be able to monitor and track human behaviors, their trials, their errors and their successes. How we process data, conduct certain tasks, make our own decisions on what to do next, may provide a rich set of data for learning from.
One area of application where this may become quite fruitful is in digital advertising. Most ad technology of the past decade has been predicated on a process of throwing shit at the wall and seeing what sticks. As the industry has started to mature, standards evolving, emergence of new channels, formats and walled gardens, the utilization of learning systems AI is prime for the next wave of development; and within this realm exists experienced and nuanced ad traffickers and media planners which can clumsily navigate multiple DSPs, Facebook campaigns, search campaigns, and interpret objectives, results and success. The increased complexity of the systems involved, scale of data, and response times required to be efficient and successful, the human powered ad technology is endangered, though I would argue that the human assisted AI for ad technology will still be viable and valuable for quite some time.
Still the biggest problem in most industries and applications is "what to do next". This is where machine learning and AI holds it's greatest promise.