Katherine R. Storrs's research while affiliated with University of Auckland and other places

... In PredNet, higher layers do not predict the activity of the representation units of the layer below, but of the error units (Lotter et al., 2016). Moreover, more recent investigations deemed the motion illusion results inconsistent with human-like illusory motion (Kirubeswaran & Storrs, 2023). ...

... In this study, we leveraged the flexibility of DNNs to construct an image-computational model of human motion processing. While recent studies successfully used DNNs to elaborate the understanding of the neural mechanism of visual motion processing [19][20][21][22][23], we aimed to make a model that can explain a broader range of physiological and psychophysical phenomena, including those whose neural mechanisms are not yet clear. From an engineering standpoint, we aimed to make a human-aligned optic flow algorithm while maintaining a flow estimation capability comparable to the state-of-the-art (SOTA) CV models. ...

... Consider, for example, the act of solving a metal puzzle, where one of the rings must be removed from the rest of the tangled metal loops. A representationalist approach, borrowed from traditional cognitive psychology, might assume that a mental representation of the metal puzzle must rst be generated in the brain of the observer and then the cognitive operations -such as mental rotation [36] and other transformations -would all be performed on that mental representation. However, that is not at all how one actually solves a metal puzzle in the real world. ...

... Market-oriented research prompted AI engineers to endorse the "if it works, it works" motto which essentially meant that they did not really care about the neuroscientific details when designing and training DL models as long as these models surpass human performance and/or turn out to be lucrative. Hassabis et al. (2017) were among the first to plea for renewing the vows between AI research and neuroscience, with cognitive neuroscientists follow-ing suit and claiming that DL models are indeed a valuable toolkit for learning about the brain, especially given their predictive power and previous success in modeling primate visual cortex (see Storrs & Kriegeskorte, 2019 for an overview, Yamins & DiCarlo, 2016). Concrete examples of models from recent years bear witness to the beneficial exchange between these fields. ...

... This suggests that setting dw = 70 as 100% of the other's destination face allows for sufficient observation of the boundary between self and other within the facial distance (the latent vector distance dw = 70) used in this experiment. Particularly, between 20 and 60%, we observed a region where subjective self-face recognition and the morphing rate are approximately linear, as reported previously 47 , confirming the validity of this facial distance we proposed. ...

... The such design mirrors the temporal predictive coding of the human visual system [28,35], wherein higher order neurons send feedback signals to lower order neurons and thus control their behaviors. This mechanism implicitly forces neurons to learn motion perception in a dynamic environment [40]. Integration of the above modules yields a lightweight network with only 2.50 M parameters, which handles causal sequences of arbitrary length to estimate optical flow. ...

... Adelson argued stuff is important because of its ubiquity in everyday life. Ritchie et al. [25] describe material perception in two parts. The first part is categorical recognition of what something is made of. ...

... Although effective, existing multimodal algorithms often lack a dynamical perception of the informativeness of each modality for different samples, which could otherwise enhance the trustworthiness, stability, and explainability of these methods. Evidence has shown that the informativeness of a modality typically varies across different samples [18], [19], which motivates efforts to model modality informativeness to enhance fusion effectiveness [20]. ...

... Thus, VGG-19 models that lack any knowledge of natural image structure show the same patterns of bias. Our findings involving V1 neural predictivity can be contrasted with a prior study that evaluated CNN models of neural responses to complex objects, for which a variety of control networks with randomized weights (including VGG-16) showed very similar levels of predictive performance from layer 1 onward (Storrs, Kietzmann, Walther, Mehrer, & Kriegeskorte, 2021). We performed an additional control analysis to evaluate the potential contributions of nonlinear complexity; this was done by removing every ReLU operation from VGG-19 except for the final one to occur in the feedforward pipeline of a layer-specific analysis. ...

... Second, the VAE model family, specifically hierarchical VAEs, is broad with other generative models, such as diffusion models, understood as special cases of hierarchical VAEs [37][38][39]. Finally, VAEs learn representations that are similar to cortex [1, 2, 40], exhibit cortex-like topographic organization [41,42], and make perceptual errors that mimic those of humans [43], indicating a significant degree of neural, organizational, and psychophysical alignment with the brain. ...