The Weak AI is Going Strong
Businesses today just want to join the AI race without giving any thought to what can be the best “AI-Market” fit to achieve success. There is a lot spoken about how AI will take over the world. We are seeing a lot of research and progress in this area of technology. While we are still in its earliest era and definitely far away from Singularity, we have started realizing the real power of AI in certain fields. Yes, even in its lowest levels, AI is taking small bites of human dependent tasks. Before I dive into the topic, I would like to touch upon a few things to provide context. In the philosophy of AI, there are three, generally known, levels of AI
Three Levels of AI
- Artificial Narrow Intelligence (ANI) or Weak AI — This is DeepMind’s AlphaGo that can play Go or Uber’s Autonomous Vehicle that can take you from point A to B. These machines try to accomplish specific objectives they are trained for. They are generally accepted only when they can perform the task on par or better than humans.
- Artificial General Intelligence (AGI) or Strong AI — This is the form of AI at which it reaches the level of human intelligence. These machines can think abstractly, show wisdom, and synthesize human emotions just like humans can.
- Artificial Super Intelligence (ASI) or Super AI — This is “Skynet”! Basically, this level of intelligence can surpass the collective intelligence of many or all humans. Well, if you think about this, once you reach AGI, reaching this level might not take long.
With that in mind, this article will be addressing ANI, the lowest level. All AI products in the market today fall under this category. But, with in ANI, we see different forms depending on what percentage of the objective the machine contributes to and how much autonomy it has.
Intelligence Augmentation
Machines are only as good as the data they are fed. That task ultimately lies in our hands. Setting an objective and feeding relevant data is pretty much deterministic to the outcome of the machine. You feed bad data, you receive bad outcomes. The majority market use cases operational today use this form of AI. This helps us perform complex tasks faster. We are always in the loop during the learning process. We are also responsible to make the insights, actionable. In this form, the real value is in processing vast amounts of information and performing complex calculations that would take long time for us to interpret and analyze. Hence, this is sometimes referred to as “cognitive automation”. Machines act as a catalyst in the whole process but humans are required at the beginning and at the end of the process. In fact, there has been research that this format of human-machine co-operation is more effective than complete autonomy. The most well know research of this form is from J. C. R. Licklider in his Man-Computer Symbiosis. He saw machines as extensions rather than adversaries to humans. By keeping control over complex situations, we have the ability to perform non-linear operations using these machines.
The nature of the process is open loop where humans have to close the loop by providing feedback. These machines usually have a well defined user interface to enable human-machine interaction. The ultimate goal of the user interface is to reduce human-machine friction and the control pretty much lies with us. And because, we have good control, there is a high level of transparency and trust with the system. This is particularly important in medical applications where there are huge implications. This form of intelligence has proven real business value and is the most widely used AI.
Semi Autonomous
In this form, there are multiple tasks to achieve the objective and the machine performs some of those tasks in an automated fashion leaving the rest to us. A well known example is the Tesla Autopilot which is a driver assist feature to stay on track and keep driving until the driver changes the course. This is a closed loop form where the machine is programmed to gather feedback on its own and reinforce to improve itself. The human still has some control over the process by setting the objectives and course correct by providing external stimulus to the machines. These machines are programmed to accept user feedback. The user interface design principles are similar to Intelligence Augmentation machines but sometimes a little less transparent and more complex as to how the machines work. This system is also designed keeping the human in the design process. And just like the previous form of ANI, human can help the computer detect novel patterns and behaviors.
Although, we do not have the control over the machine’s tasks, the control and success of the outcome still lies with us. The value of such an AI is to offload complex, error prone tasks that machines can perform in sub-seconds but humans would take days or months. In other words, in every process, there are tasks that take longer, tasks that could be parallelized, and tasks that need large amount of resources. Machines take over them. With such a reduction of processing time, it unleashes a plethora of new possibilities that is otherwise difficult.
One of the challenges these systems face is the accountability aspect. As we do not have control over the machine dedicated tasks, we accept certain amount of risk. It also makes the machine vulnerable to any kind of misdirection or deception. Recent example of such a scenario being again the Tesla Autopilot that had a fatal collision as the machine mistook the side of a trailer truck to sky.
Autonomous
Self-driving cars! The most common and well know example of this form. We have seen a lot of focus on this area during the near past and probably will continue in the distant future. The motivation of having such systems is to bring error rate close to zero. The objective is still set by us, but we do not have much control over the process or the outcome. The systems can be very complex giving only the scientists who built them any control. This form of ANI is built focussing on weak humans to off load all cognitive processing that will be required towards achieving the objectives.
Other well know examples we see in this form are Chat Bots and Assistants. The capability of one machine replacing thousands of customer support human workers is the driving force for all businesses to adopt AI in this area. But, as we talked about the vulnerabilities of semi-autonomous, these machines also have seen very public and embarrassing failures. Best example, Microsoft’s Tay, that recently was fooled into making racist comments on a public forum.
Conclusion
We see that a lot needs to be worked on the semi-autonomous and autonomous area. But, we definitely see a promising future for Intelligence Augmentation form of ANI that has provided value in many industries and is seeing a wide spread adoption. With that, we will have to wait and see how these things unfold. Companies have to be smart in selecting the form of AI they want to embrace rather than just participating in the rat race. Merely adopting advanced technologies may not provide the best value to the business nor its customers. A thorough evaluation of their strategy and customers needs can bring out the best fit.
Hope this helps! Thank you for liking my post.
^\_(**/)_/^
