Doubt as a Strength: AI’s Journey towards Intelligent Learning

“Dubito, ergo cogito, ergo sum,” AI ponders

Questioning leads to thinking, and thinking defines existence—a step forward for AI

It’s the beginning of a new week, and Paul opens an email from his boss, Heather. Something feels off as Heather requests an urgent purchase of $200 gift certificates from a nearby store, followed by an immediate response with the codes. Doubt creeps into Paul’s mind. After a closer look, he discovers a typo in Heather’s email address, revealing the fraudulent nature of the message—a doubt confirmed by the company’s IT department soon after.

Paul’s ability to doubt saved him from falling victim to a common phishing attack, highlighting one of the fundamental faculties of human thought. When we contemplate human and machine perception, such as reading and comprehending an email, we envision a seemingly mechanical process: pixels on a screen translating into a semantic interpretation, like “My boss wants me to buy $200 in gift certificates.”

In this feedforward process, doubt finds little room to manifest itself. The combination of letters on the screen straightforwardly leads to a specific meaning. However, Paul’s doubt regarding the email unveils a different truth. Paul possessed a mental model or expectation of how Heather’s emails should appear and resonate with him, allowing him to spot the impostor. If Heather regularly asked Paul to purchase gift certificates every Monday, the email would not have aroused suspicion in Paul’s mind; it would have been classified as normal.

Doubt is selective and can be learned.

Deep Learning: The Pitfalls of Overconfidence

But what about machines? Specifically, what about AI and its latest triumph—deep learning?

Building upon theoretical work from the 1960s, today’s deep neural network (DNN) algorithms derive their power from their ability to learn from data rather than being explicitly programmed. Through supervised training, these systems excel at classifying objects, such as distinguishing between images of cats and dogs. When fed ample high-quality data, deep learning algorithms achieve remarkable accuracy in recognizing these distinct classes.

When Self-Doubt Becomes a Strength

However, when faced with an unfamiliar entity, the DNN may falter. Let’s say a mouse enters the picture. Since the DNN was not trained to recognize mice, it will have to guess and categorize the mouse based on the closest available class.

Despite their connection to the architecture and physiology of the nervous system, DNNs differ from biological brains in one crucial aspect. Brains possess a “two-way street” that connects low-level features, such as pixels in an image or words in an email, to higher-level concepts like the class of an image or the meaning of a phrase.

Why Should AI Embrace Doubt?

Nature avoids waste, and the abundance of feedback projections in the brain is evident from Paul’s example. In the months leading up to that email, Paul had learned to interpret Heather’s emails and developed expectations regarding their content. These expectations, constantly operating in the background of our minds, contribute to our intelligence as humans and animals. They enable us to leverage our learning to predict what will happen next.

Although still not commonplace in AI research and application, “doubt” holds value when it comes to training and implementing DNNs. During training, if a DNN can confidently classify objects—or, in this case, animals—such as cats and dogs but also produce an “I don’t know” signal, developers can evaluate whether the system has truly mastered the learning process for a particular class.

In Doubt, There’s Opportunity

Currently, designing DNNs to possess this functionality presents challenges. However, emerging approaches are incorporating the concept of doubt into their architectures naturally. Like brains, these new technologies employ feedback mechanisms to assess whether an input aligns or deviates from what they have learned. This not only explains why these DNNs, as well as humans, experience doubt but also sheds light on a crucial aspect of human learning: our constant evaluation of whether our predictions align with reality.

Now that we have grasped the concept of doubt, let’s explore why AI’s ability to recognize doubt holds importance for business leaders. Experts estimate that poor quality can cost manufacturing and service companies up to 30% of their gross sales annually, amounting to millions of dollars. Therefore, there is significant motivation to enhance quality inspection and prevent product defects early in the process. Many quality issues in products could have been avoided if AI models used in inspections had the capability to doubt when faced with deviations from what they have learned as “good.”

Consider industrial manufacturing settings where computer vision systems are traditionally deployed for quality inspections. Detecting defects like deformed bottle caps can be challenging, but with the assistance of vision AI software, the computer vision system can be trained on a set of “good” caps. When faced with an anomaly or an unknown product, the network will exhibit doubt about its own judgment. This doubt signals a potential defect or deviation from the norm, allowing for appropriate categorization and handling of the product.

In manufacturing and similar use cases, an inspection tool equipped with doubt can have a tremendous impact, saving time, money, and resources throughout the supply chain—from raw material intake to shipping.

Remember: Doubt is what grants brains their superpowers.