Finetuning primate visual representations with word recognition

Reading accurately requires precise encoding of letters and their relative positions, especially to distinguish similar words like TRIAL and TRAIL. While reading expertise leads to the formation of the Visual Word Form Area (VWFA), the neuronal changes enabling invariant word recognition remain unknown. In this study, we trained two macaque primates to process orthographic inputs using a same-different task. The neural responses to a large set of 4-letter strings were wirelessly recorded from the Inferior Temporal (IT) cortex before and after 5 days of training. A subset of these stimuli was designed as “words,” while others were used as “nonwords.” During each training session, the monkeys performed at least 360 task trials. The sample stimuli were always words, and their positions varied to test for invariant word recognition. The test stimuli, which could be either words or nonwords, were always centered. Neural data was also recorded during the training sessions. As expected, the monkeys’ performance (response time and error rate) improved over the five days of training, suggesting that word recognition became increasingly position-invariant. Furthermore, they became sensitive to the orthographic structures of words, and their performance on non-matching conditions degraded with an increase in similarity to words. Neural recordings revealed that many cells were tuned to specific letters. Using a letter x position-based model, we observed that most neurons encoded retinotopic position in the contralateral visual field, but a few neurons also encoded the ordinal position of letters. With training, the representation space expanded, and receptive field size of few neurons broadened, eventually leading to enhanced position invariance. Overall, we observed finer discriminability between letter strings with training, which could potentially be a consequence of a shift from retinotopic to ordinal position encoding among a few units.