Smart interaction design is the proper way to solve the learning problem in AI

The Challenge is Human-Computer-Interaction, not Algorithms

Artificial intelligence (AI) is one of the most hyped terms in the 21st century, and yet one of the most misunderstood.

Very often, when talking about AI, we like to automatically couple it with other terms such as Machine Learning, Deep Learning, and Neural Networks. This makes it sound like over 90% of AI is this kind of statistical algorithm that only PhDs can understand.

Understanding human interaction gives data it’s meaning

While automated learning and classification algorithms are vital to the development of an artificially intelligent system, they only serve as enablers of true intelligence. These algorithms are necessary to develop intelligence, but not sufficient.

What this means is that the improvement of machine learning algorithms over the years has made it possible for machines to perceive the world almost like humans do. However, in order for a machine to think and act like a human, we must look elsewhere for the answer.

Over the thousands of years of evolution, humans developed unique ways of interpreting and thinking about the world. In order for “AI” (thus all the technologies that hide behind this term) to have a significant impact on our society, it must understand not only how to act like a human, but also how to think like us.

Unfortunately, while the information revolution has enabled us to collect petabytes of data on how we actin a certain situation, not much data has been collected on how we think. This makes it impossible to properly train an AI system.

Gathering meaningful data through smart interaction design

We need to start working on large-scale interaction systems that enable machines to rapidly communicate and collaborate with humans. Machines need to start learning how we conceptualize the world.

What this means for researchers and companies is that the a big part of the future of “AI” lies in design, in a systems ability to interact with and learn from humans, and in understanding human contexts — not just in more powerful GPUs and algorithms. A deep empathy toward the needs and challenges of the end users who will be interacting with these systems will be needed.

In this article, I will take an in-depth dive into the current state of “AI” (in particular CV and state of the art neural networks) and demonstrate where these technologies today fall short when compared to human intelligence.

I will also discuss why learning by interaction has been one of the best ways, if not the only way, for both AIs and humans alike to learn about our world. Finally some suggestions will be offered on how we can create these interaction systems that will ultimately enable true artificial intelligence.

Current AI Closely Mimics the Human Thinking Process, But Falls Short in Learning on Its Own

Most of the AI systems on the market mimic the human information processing model in psychology. To oversimplify, an intelligent system processes information in three very distinctive stages: reception, interpretation, and learning.

First stage: Reception

Receptionis the process in which some receptors (e.g. eyes, ears of the human body) receives signals from the environment, and send those signals to a processing agent (the brain) in formats that are interpretable by the processing system (electromagnetic signals).

In AI, examples of these receptors include the cameras on Tesla’s semi-self-driving cars, a LIDAR in a Waymo car or a far-field microphone for Alexa.

Second Stage: Interpretation

Subsequently comes the interpretation process, in which the processing agent (the brain) performs three operations to the data sent by the receptors:

1. It identifies several objects of relevance from the data
2. Then, it goes into a library of references (the human memory), searches (with to us unknown “algorithms”) for references that will help it identify the objects, and then identifies them (recognizes that this shape is an apple).
3. Based on the current state of the entire system (evaluation of basic needs, like hunger, tiredness), the processing agent (the brain) determines the importance of each piece of information it receives, and present to the users only the information that passes a certain threshold (in human cognition we call this attention). All of us experience this cognitive bias frequently. For example, when you are hungry, you are likely noticing more food around and develop sudden superpowers by smelling the garlic from that Italian place that’s a mile away.

In “AI”, interpretation usually happens in a large GPU information processing systems on the cloud using state of the art machine learning algorithms.

With recent developments in machine learning and game-playing algorithms (especially with deep neural networks and recently computer vision), “AI” systems can identify objects based on a body of reference exceptionally well, enabling amazing innovations such as self-driving cars to develop.

However, we cannot stop here, since the library of reference used by the processing agent is limited, especially in the beginning of its life cycle (a baby might not know what garlic even is).

Third stage: Learning

Here is why learning has to occur to continuously expand this library of references for the system to reach its full potential.

If you look at current AI research, this is where the real challenges lie. Neural networks are really good at classifying a situation into categories and optimizing based on the parameters provided. However, the data it needs to create these categories or parameters from scratch without help from human developers is sheer unimaginable and thus this is extremly hard.

This is because an “AI” sees the world as multiple, purely mathematical matrices, and does not have the intrinsic ability to empathize with human experiences unless we teach it to.

During the training of these classification models, they are only given the outcomes of each specific situation (labeling), instead of the entire thinking process and rationale that led to that specific outcome, which makes comprehension a lot harder.

For example, a system might be able to programmed recognize the image of a baby, but it will not understand why recognizing the image of a baby is needed in the first place since that information was never given to it by the engineers who created it. Here it’s important to point out, that we are not trying to develop a general artificial intelligence, but to optimise current, spezialized “AI” systems.

In a way, current “AI” systems are like a super intelligent newborn baby — while you can show it all the knowledge in the world, it cannot understand how the world really works unless it actually gets out into the world to learn from experience.

Because it lacks the ability to create its own contexts, most of the commands we are asking Siri and Alexa are actually manually programmed by the engineers at Apple and Amazon. It’s also why Amazon spends so much effort to create an open eco-system around Alexa, to encourage companies to program skills on its Alexa platform. Precisely because they are so human-dependent, the current systems such as Alexa cannot really develop new context and learn like humans. Therefore, it is really not accurate to call them “artificial intelligence” (I do know, that hypocritically I used the term continously in this article).

Smart interaction design is the easy way to solve the learning problem

So how can we create an AI system that can empathize with the human world? The answer is rather simple — we teach it to interact with humans and ask questions.

As we concluded earlier, the biggest challenge of training any “AI” system a form of human interaction atmis the lack of detailed, interaction-level data about the human thought process.

AI’s should focus on data collection, not just data consumption

In order to collect this level of data, we need to ramp up it’s role not only as a data consumption(using data generated elsewhere) agent, but also data collection agent (generating its own datasets).

This means we need to design AIs in such a way that they can interact with humans to understand not only what the human wants, but why they want it (context). To oversimplify: Just like an apprentice to a master craftsman, it needs to learn on the job.

In order to have this level of interaction with humans, we must first shift our perception of AI. Currently, we perceive AI as this omnipotent black box that can solve all the problems automatically without any requirement of human input.

When inexperienced users try an AI system for the first time (think Siri or Alexa) you can experience this dissonance quite often. But even in general, when one things of “AI” , you’d think that you can an AI a general command such as “Alexa, do my work for me” and expect Alexa to run our businesses for us as we stay in bed and watch Netflix.

Let’s start thinking of “AI” as a consultant

When a human consultant interacts with clients, he will never pull the client aside and tell them, “Hey we understand your business needs perfectly. We can do everything to improve your business for you, you just need to sit and watch.”

Instead, they will spend hours and hours sitting down with the clients, asking carefully crafted questions to better understand the needs of their clients, and ultimately work withthe client to create a solution that is tailored to their specific needs.

The success of a consulting project really hinges on a consultant’s ability to draw out the needs of the client and to deliver the most value for their client based on their constraints.

If this is the standard that we hold ourselves to when interacting with one of the more intelligent segments of human society, we should not expect AI to automatically understand all our needs and provide the perfect solution without interacting with us.

Creating a more interactive AI

In order to give AIs the ability to ask intelligent questions the way consultants do, we must place a lot less emphasis on creating the most powerful machine learning algorithm. Rather, we ought to focus on designing systems that enable maximum interaction between AI and human userswhile achieving the task the AI is designed for.

More practically, it means that AI product managers should focus not only on hiring engineers that are good at algorithm design, but also on recruiting human-centered designers and engineers with an understanding of ML. They are the ones who can talk with the end users of particular AI products and facilitate interaction. The task of this duo, engineering and design, is to identify the best ways for AI to work withusers to improve both the systems own intelligence and thus lives of human users.

AI should serve people, not replace them

Essentially, creating interactive AIs demands that AI product managers build AIs with the goal of understanding and servingpeople, instead of replacingpeople.

Essential for that is also, that development of AI must also be much more transparent than it is right now. As design firm IDEOpoints out in “A Message to Companies That Collect Our Data: Don’t Forget About Us,” businesses today must design for transparency and user control if they want to build trust with customers.

At the moment, many AI companies refuse to demystify what is actually happening under the hood to their users. One would think that is because they fear competitors stealing their technology secrets, but it rather because they fear that if the consumer saw how much manual IF THEN is still under the hood of some “AI” systems, users will lose trust in the company. Apart from that, as previously discussed, the nature of ML systems is, that they are somewhat of a black box, so transparency often is impossible.

To achieve the highest level of intelligence possible for AI technologies, users must be intensively involved in the design iteration process. Therefore, while transparency might harm the early adoption of an AI product in the short term, the long-term benefit of more transparency to the entire AI system is limitless.

Final thoughts

The changes in our daily lifes the technologies hidden under the term AI will bring us are immense. Like it or not, AI will be the mayor force in changes to our workforce in the next few decades.

However, to make AIs truly useful for our society, they need to understand not only what we humans do, but also why we do it, and this learning requires AIs to jump out of the black box and interact with its users.

As ML, CV and other dominant AI technologies move from the academic stage into a production stage, AI will be less and less of a coding issue and more of a design problem in which human-centered designers who empathize with end users will play a critical role.

Ultimately, I can see a future world in which these systems and humans exist in harmony, with each party playing its unique role in human society. Only then will the AI revolution result in prosperity for humanity.