Search Papers | Poster Sessions | All Posters
Poster C6 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Recurrent issues with deep neural networks of visual recognition
Timothée Maniquet1 (), Hans Op de Beeck1, Andrea Ivan Costantino1; 1KU Leuven
The human ventral stream is equipped with recurrent connectivity allowing it to deal with the noise and uncertainty of visual inputs. Including recurrence in Deep Neural Networks (DNNs) is a promising way of modelling this non-feedforward connectivity. However, just like the role of recurrent processing in the brain remains elusive, it is unclear how making DNNs recurrent makes them more human-like. Here, we put to the test a wide range of DNN models equipped with various recurrent connections. We compared them to human behaviour facing challenging object recognition, and found recurrent model performance and consistency with humans to be mediated by size. Moreover, we found recurrent DNN confusion matrices to be less similar to that of humans than feedforward ones. These findings give perspective on the implementation of recurrence and the benchmarks used to assess it.
Keywords: recurrence visual recognition