This recent study utilized various metrics to evaluate the similarity between different neural network representations, analyzing diverse architectures, training objectives, and data modalities. The findings reveal that different models, regardless of their architecture or objectives, can achieve aligned representations, and this alignment improves with model scale and performance.
One key aspect of our research was the use of the mutual nearest-neighbor metric, which assesses the overlap in nearest-neighbor sets between representations from different models or systems. This approach highlighted that neural networks show substantial alignment with biological representations in the brain.
It becomes incredibly challenging for humans to imagine concepts entirely detached from known reality and experience. Our own experiences are minuscule compared to the potentially higher dimensions in nature. At this scale, as neural networks begin to approximate human top-level cognition in higher dimensions, neural networks may face the same imaginative limits as the human brain. This raises an intriguing question: What methods can help us break through these imaginative limits? π€
