The AI program spoke in a foreign language in which it was not trained. These inexplicable behaviors are called emergent traits, and the AI learns new skills unexpectedly. A recent example is an AI program recently adapted to Bengali based on some prompts. AI can now fully translate Bengali with just a few prompts.
Learning a language that an AI program has not been trained in raises many questions. Why does this happen and what is this phenomenon called? According to recent reports, the only plausible explanation seems to be a phenomenon known as AI black boxes.
Simply put, an AI black box is a system whose inputs and operations are invisible to users and other parties. It is known to be an impenetrable system. The striking element of this is that black-box AI models reach their conclusions without providing a rationale for how they made their decisions.
To understand AI black boxes, we first need to know how human or machine intelligence works. Learning by example is what drives most intelligence, whether human or machine. For example, children learn to recognize letters and different animals. Just show examples of letters and animals and you’ll be able to identify them in no time.
According to Samir Lawashdeh, a professor at the University of Michigan at Dearborn, the human brain is essentially a trend-spotting machine, identifying qualities when exposed to examples and ultimately classifying them autonomously and unconsciously. can. AI expert Lawashdeh says it’s easy, but explaining how it’s done is almost impossible.
Deep learning works in much the same way because it is trained the same way children are trained. These systems are fed correct examples of what they should be able to recognize. A proprietary trend detection mechanism then immediately evaluates the neural network to classify the corresponding objects. Searching for the same object in the search bar correctly displays the object or image. Similar to human intelligence, we don’t really know how deep learning systems reach their conclusions.
What Sundar Pichai said about black boxes
“There is an aspect that we all call the black box in this field. We have some ideas and our ability to understand this will improve over time, but that’s where the state of the art lies,” Google CEO Sundar Pichai said in January this year. Minutes,” Scott Perry said.
When Perry interjected, “You don’t fully understand how it works, and you let society loose it?” I don’t think we fully understand how the human mind works.”
Why is the black box problem a concern?
AI can do many things humans cannot, but the problem of AI black boxes can lead to mistrust and uncertainty about AI-powered tools. For data scientists and programmers, AI black boxes can be a challenge because they are self-directed and there is no data available about their inner workings.
One of the most prominent issues is AI bias. Conscious or unconscious biases of developers can introduce biases into algorithms, but in black boxes these can creep in undetected. Deep learning systems are now being used to make decisions about human treatment, loan eligibility, or who should get a particular job. In such cases, the AI system already exhibits bias. And the black box problem can exacerbate this and make it difficult for many people to use certain services.
Lack of transparency and accountability can also cause many problems. The complexity of black-box neural networks makes these systems poorly auditable. This can pose problems in areas such as healthcare, banking and financial services, and criminal justice. Additionally, they have numerous security flaws and are vulnerable to attacks from various threat actors. This can be explained by a scenario in which a malicious party modifies the model’s input data to influence the model’s judgment and make potentially dangerous decisions.
How can we combat the threat of AI black boxes?
Experts say there are two approaches to the black box problem. Creating a regulatory framework and finding ways to dig deeper into the box. Since the output and the decisions behind it are incomprehensible, a deeper look into the inner workings may alleviate the problem. This is where Explainable AI comes in, an emerging field of AI that aims at transparent and accountable deep learning.
Although AI black boxes present many challenges, systems using such architectures have already proven useful in many applications. Such systems can still identify complex patterns in data with a high level of accuracy. Reach conclusions relatively quickly and using less computing power. The only problem is that it can be difficult to understand exactly how they came to that conclusion.
