AI Inspired by Human Development
My collaborators and I investigate the mechanisms underlying the remarkable performance of deep neural networks, while also considering persistent limitations not observed in the human brain, such as robust generalization. Adopting a developmental approach, we treat the network as an immature and inexperienced visual system. This strategy enables us to meticulously control the type of visual experience, in this case, the training regimen, that the network undergoes throughout its development. This method has already demonstrated success, as we derive testable theories from our research on the development of the human visual system, operating under the assumption that the human brain has evolved to optimize its visual experience, thereby facilitating robust visual learning.
Key Findings
1
Color Perception: Impact of early visual experience on later usage of color cues
Children treated for persistent congenital cataracts can distinguish colors from the outset of vision. However, their object recognition performance significantly drops when color information is removed, unlike their typically developing peers who can recognize objects regardless of color cues. We propose that the immediate access to rich color information may lead the late-sighted children to over-rely on color cues. This is in contrast to the typically developing neonate, whose immature retinal cones lead to an initial period of reduced color signaling early in development. Our computational simulations with DNNs support this theory, emphasizing the adaptive significance of the typical developmental trajectory of color vision and offering pathways for improving machine vision algorithms through biomimetic training regimens.
Update: manuscript revisions are under review. Be on the lookout for publication updates soon…
2
Learning invariance by Deep Neural Networks
Spearheaded by the talented Xavier Boix, in this project a pivotal question in the deep learning community: understanding the mechanisms that underlie robust recognition by DNNs, despite image-level transformations (e.g. varying resolution). Additionally, we explore how we can leverage this understanding to facilitate further robustness to untrained transformation. Our analysis focuses on training paradigms where only some object categories undergo transformations. We assess whether the DNN extends its robustness to categories that weren't subject to transformations. The findings reveal that while the network can develop robustness even without invariant representations, increased robustness occurs with invariance, particularly as the number of transformed categories in the training set rises. This is particularly true for local transformations like blurring and high-pass filtering but not for geometric transformations affecting spatial arrangements, such as rotation and thinning.
Read our paper:
Jang, H., Zaidi, S. S. A., Boix, X., Prasad, N., Gilad-Gutnick, S., Ben-Ami, S., & Sinha, P. (2023). Robustness to Transformations Across Categories: Is Robustness Driven by Invariant Neural Representations? Neural Computation, 1-28.
3
The downsides of High Initial Acuity (HIA)
In studies of children treated for congenital cataracts, an intriguing discovery emerged: when tested on face individuation years after treatment, they showed impairments in configural processing compared to normally developing children. The key difference lies in the visual acuity trajectory, with newborns experiencing a period of low acuity and those with late sight-onset immediately encountering high acuity post-cataract removal. Rather than attributing this to a face-specific critical period, we propose a simpler explanation—deficits in configural processing result from abnormally high initial acuity (HIA). Computational simulations support this idea, revealing that training with low-resolution images induces larger receptive fields and better generalization across resolutions, essential for configural processing. Our findings challenge the notion that training on optimal (high-resolution) input may hinder face individuation development, emphasizing the adaptive role of the acuity trajectory in normal development and suggesting enhancements for computational face recognition systems. Our recent empirical work is consistent with the predictions set forth by this hypothesis.
Read our papers:
Vogelsang, L.*, Gilad-Gutnick, S*., Ehrenberg, E., Yonas, A., Diamond, S., Held, R. and Sinha, P. (2018). Potential downside of high initial visual acuity. Proc. Natl. Acad. Sci., 115 (44) 11333-11338.
*Equally contributing first-authors
Vogelsang, L.*, Gilad-Gutnick, S*., Diamond, S., Yonas, A. and Sinha, P. (2019). Response to Katzhendler and Weinshall: Initial visual degradation during development may be adaptive. Proceedings of the National Academy of Sciences, 116 (38) 18767-18768.
*Equally contributing first-authors