Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A neuro-cognitive model of efficient letter recognition
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Janos Pauli1, Benjamin Gagl2; 1University of Cologne, Self Learning Systems Lab
Understanding how the brain transforms printed letter strings into meaningful linguistic units is a central challenge for cognitive science. Reading a word and accessing its meaning critically depend on identifying its constituent letters. Expert readers excel at this, as they can read around 200-300 words per minute, irrespective of font. This highlights that they process letters with remarkable efficiency, as the visual appearance of written script is often highly variable. However, an explicit computational account of how font-invariant letter recognition is achieved remains absent from predominant models. Cognitive models of visual word recognition tend to operate on abstract features, ignoring issues of initial visual processing, thereby circumventing the idea of font invariance. Modern connectionist models, such as Artificial Neural Networks, achieve empirical font invariance but impose no a priori constraints on representational format, thereby lacking the interpretability required for an explicit neuro-cognitive account. To challenge this gap, we develop and implement a transparent, image-computable neuro-cognitive model of letter recognition. We assume that visual letter information is decomposed into features that describe distinct letter characteristics and shapes. These features then activate prototype representations that capture font invariance. Therefore, the model identifies letters based on a succession of pixel-, feature-, and memory-processing. The model's transparency enables us to assess which sensory processing account best implements font-invariant letter recognition. We therefore tested model variants that implement either (i) purely bottom-up processing or (ii) top-down optimization based on the assumptions of predictive coding. Furthermore, we compare these transparent theoretical variations with an (iii) artificial neuronal network model. To evaluate the computational models, we used a behavioral letter identification task with increasing levels of noise and different fonts. We demonstrate that the predictive-coding-based model variant best simulates behavioral response patterns of all implementations. To investigate whether the assumed representations can be identified neurally in the model's predicted succession, we conducted an additional EEG study and found that pixel-level prediction-error representations are implemented between 100-150 ms at posterior electrodes, before we identify a component reflecting memory access around 200 ms. Thus, based on these findings, we conclude that letter recognition is optimized on the sensory level by top-down predictions, yielding a representation optimal for memory access.
Topic Areas: Reading, Computational Approaches