STSBench dataset boosts primate dorsal stream modeling with 2,000+ neurons

The primate visual system splits into two streams: the ventral stream for object recognition and the dorsal stream for spatial relations and motion. While convolutional neural networks (CNNs) pretrained on object recognition have proven remarkably accurate at predicting ventral stream activity, models of the dorsal stream have lagged behind due to a lack of large-scale neural data. To close this gap, researchers from Stanford and other institutions have released STSBench, a dataset of single-neuron recordings from over 2,000 neurons in the superior temporal sulcus (STS) of Rhesus macaques. The recordings were collected while the animals viewed thousands of unique, natural videos — a setup that captures the rich visual dynamics processed by the dorsal stream.

At 50 times the size of previous datasets, STSBench is a game changer for computational neuroscience. The team demonstrated that STSBench can be used to benchmark encoding models that predict dorsal stream responses and even reconstruct visual inputs from neural firing patterns. Accepted at the NeurIPS 2025 Datasets and Benchmarks Track, the dataset is publicly available and comes with 21 pages of methodology and analysis. This work opens the door to building more accurate AI models of motion perception, spatial awareness, and visually guided behavior — areas where current AI still struggles compared to biological vision.