Grace W. Lindsay's research while affiliated with New York University and other places

... To address this question, we need to be able to manipulate experimentally how neural networks form, as they learn to achieve behavioural objectives, to establish the causality of these relationships. Computational models allow us to do this 8 . They have shown that network modularity can arise through the spatial cost Our understanding of how the brain's structure and function interact largely comes from observing differences in brain structure, such as across individuals 6 or following brain injury 7 , and then systematically linking these differences to brain function or behavioural outcomes. ...

... Benchmark model. Systems neuroscience lacks a benchmark model that captures our existing knowledge about the nature and origin of vision (Poggio and Serre, 2013;Golan et al., 2023). Despite this lack of a benchmark model, we know many benchmark features relevant to vision. ...

... Notably, the advent of this new source has strengthened the tendency of neuroscientific research to cast aside the phenomenological source by implicitly assuming a close correspondence between physiological processes and conscious experience. In DL-based computational neuroscience, numerous research completely dismisses the phenomenological source and compares ANN processes to brain processes (comparison which does seem quite natural given DL's initial structural inspiration being the brain); examples include Yamins' paradigmatic research on vision in brains and in ANNs (Yamins & DiCarlo, 2016), other recent investigations (Kumar, et al., 2022;Millet, et al., 2022) and even proposed research programs (Doerig, et al., 2022;Cohen, et al., 2022). ...

... Black and red arrows represent feedforward and recurrent connections, respectively. Adapted fromLindsay et al. (2022). ...

... For example, the Allen Brain Institute released an SDK that simplifies retrieval of and interaction with extensive collections of NWB standardized data recorded with cutting edge electrophysiology and imaging tools. Such initiatives greatly expand opportunities to reuse data in education [44][45][46], basic research [47], and bench-marking of new computational models [48]. ...

... In addition, experimental evidence supporting the presence of inhibition of return both at the behavioral and neural levels remains controversial (19,20). On the other hand, neural circuit models for explaining the neural effects of top-down attention often treat them as static inhibitory (21) or excitatory inputs (22,23) to local circuits; doing so, thus, completely ignores the dynamical fluctuations of attention. Therefore, despite widespread investigations, the fundamental questions of the neural circuit mechanism underlying attention fluctuations and their functional role remain unclear. ...

... The role that attention plays in the brain encourages its addition to artificial neural networks, specifically in DL. Much like the apparent need for attention in the brain, it is a means of making neural systems more flexible [48]. ...

... Some artificial intelligence architectures, and in particular Deep Convolutional Neural Networks (DCNN) can fulfill this role. DCNNs and the visual cortex have similarities in how they process information [19,20]. Accordingly, patterns of neuronal activations within a DCNN have been shown to predict those recorded in the visual cortex of humans [20][21][22][23][24]. Furthermore, previous studies have shown that metrics summarizing the processing of visual information by DCNNs correlate with self-reported assessment of this processing; for example, the mean activation per layer predicts the perceived complexity of an image [25]. ...

... For example, DCNNs rely on texture information more than shape in object recognition [65]. Yet for many neuroscientists, the emergence in DCNNs of the most fundamental properties observed in animal vision indicates that they are still powerful tools for modelling and understanding this perceptual modality [66,67]. Regarding the second assumption, we acknowledge that purely feed-forward models like VGG16 fall short in describing the full complexity of the judgment of beauty. ...

... It has been vastly studied in psychology and neuroscience [Posner and Petersen, 1990, Cohen et al., 1990, Phaf et al., 1990, Bundesen, 1990, Desimone et al., 1995, Mozer and Sitton, 1998, Corbetta and Shulman, 2002, O'Reilly and Frank, 2006, Petersen and Posner, 2012, Moore and Zirnsak, 2017 and more recently by Flesch et al. [2022], Dekker et al. [2022. These studies have acted as a source of inspiration for several artificial intelligence models [Khosla et al., 2007, Lindsay andMiller, 2018] including the ones proposed in this thesis. ...