New AI Training Trick Makes Machines See More Like Humans
It could make image search, robots, and accessibility tools finally 'get' how you see.
Researchers found that fine-tuning pre-trained deep neural networks with Relational Knowledge Distillation—a method that transfers the relational structure of human similarity judgments into AI—brings DNN representations close enough to human mental representations that the two can be aligned at the level of individual objects, without any supervision. Using an unsupervised comparison method called Gromov-Wasserstein optimal transport, they tested this alignment on a curated set of concepts that did not overlap with the training data. The improvement was driven by a more human-like global structure in how the DNN organized coarse categories, while the local overlap of nearest neighbors with human judgments stayed largely unchanged.
- AI vision models can identify objects but organize them differently than people do.
- A training method called relational knowledge distillation (teaching by matching relationships) made AI's internal image map align with human perception.
- The AI even matched human perception on objects it had never seen before, suggesting a more human-like global structure.
Why It Matters
AI that sees more like we do could lead to better image search, more intuitive robots, and safer self-driving cars.