A new deep learning model of mental rotation — the ability to compare objects seen from different viewpoints — closely matches human performance and reaction times, according to a study published on arXiv by researchers from the University of Toronto and several European institutions. The model, built from three stacked components, was validated using virtual reality experiments where participants could sometimes manipulate objects themselves.
The research
Raymond Khazoum, Daniela Fernandes, Aleksandr Krylov, Qin Li, and Stephane Deny designed a mechanistic model of mental rotation that combines three components: an equivariant neural encoder that creates 3D spatial representations from images, a neuro-symbolic object encoder that converts those representations into symbolic descriptions, and a neural decision agent that compares these descriptions and prescribes rotation simulations in 3D latent space via a recurrent pathway. The team used VR experiments (sample size not specified in the abstract) where participants could sometimes manipulate objects to compare, complementing existing experimental literature. The model captured performance, response times, and behavior of participants in both their own and others' experiments. Ablation studies showed that each component is necessary for the model's accuracy.
Why it matters
Mental rotation is a core cognitive skill linked to spatial reasoning, STEM performance, and everyday tasks like navigation or assembling furniture. Understanding how the brain accomplishes this can lead to better brain training programs. This research suggests that the brain may use a combination of deep neural representations and symbolic reasoning — a hybrid approach that is more efficient than pure deep learning. For readers, this means that mental rotation might be trainable by practicing both visual rotation and symbolic thinking.
What you can do
Practice mental rotation with simple 2D or 3D shape comparisons, like those found in spatial reasoning puzzles. Try rotating objects in your mind before physically moving them — this strengthens the neural pathways the model simulates. The IQGenio platform offers spatial reasoning exercises to help you improve.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.