Have you ever wondered why artificial intelligence (AI) systems can recognize faces better than humans but make surprising mistakes like confusing an astronaut for a shovel?
Researchers at Brown University’s Carney Institute for Brain Science are working on this intriguing puzzle.
Just like the human brain, AI systems use unique strategies to understand and classify images. But what exactly goes on inside these systems is still largely unknown.
That’s why they’re often called ‘black boxes’.
Thomas Serre, a professor at Brown, compares the work his team does to shining a light inside these black boxes to uncover the similarities and differences between AI systems and human brains.
How does AI learn to recognize images, you ask? They’re trained on vast sets of data, like ImageNet, which contains over a million images categorized into thousands of object categories. The AI system then discovers its own strategy to process these images. However, evaluating the accuracy of this strategy is a challenge, but that’s where Thomas Serre and his team’s tool, CRAFT, comes in.
CRAFT, short for Concept Recursive Activation FacTorization for Explainability, is a tool that helps to understand what strategies AI systems use to process images.
For instance, if the AI is learning to recognize a type of fish called a tench, it will pick up key visual concepts such as fish fins, heads, tails, and eyeballs. Interestingly, AI can also pick up biases in the datasets.
For example, if there are many photos online of sports fishermen holding a tench, the AI might associate a white male face with the concept of a tench.
CRAFT is a big leap from the previous methods used to understand computer vision. Earlier methods mainly highlighted the prominent parts of an image that the AI focused on but didn’t explain what the AI was ‘seeing’.
CRAFT, on the other hand, also explains what concepts the AI is using to understand the image and how it’s ranking these concepts.
So, how does CRAFT help understand AI’s funny mistakes? Let’s take the example of an astronaut being mistaken for a shovel.
According to CRAFT, the AI system misclassified because it picked up on the concepts of “dirt” and “ski pants” usually associated with a shovel and couldn’t make the connection with the astronaut’s pants as it hadn’t seen a similar image during its training.
Understanding how AI sees images is critical for a few reasons. It can help enhance the performance of vision-based tools like facial recognition, making them more trustworthy and safer against cyber attacks.
It can also teach us about the differences between human and AI vision systems and help improve AI by incorporating human brain-inspired mechanisms.
While AI vision systems that mimic human vision systems might sound worrying, in reality, making AI more human-like makes them more reliable and less prone to errors.
In the future, the researchers aim to use this understanding of AI and human vision to tackle big scientific problems that humans currently can’t solve, like cancer diagnostics or fossil recognition. It’s indeed an exciting road ahead!
