A Developmental Approach to Machine Learning?

Abstract

Visual learning depends on both the algorithms and the training material. This essay considers the natural statistics of infant- and toddler-egocentric vision. These natural training sets for human visual object recognition are very different from the training data fed into machine vision systems. Rather than equal experiences with all kinds of things, toddlers experience extremely skewed distributions with many repeated occurrences of a very few things. And though highly variable when considered as a whole, individual views of things are experienced in a specific order – with slow, smooth visual changes moment-to-moment, and developmentally ordered transitions in scene content. We propose that the skewed, ordered, biased visual experiences of infants and toddlers are the training data that allow human learners to develop a way to recognize everything, both the pervasively present entities and the rarely encountered ones. The joint consideration of real-world statistics for learning by researchers of human and machine learning seems likely to bring advances in both disciplines.

Development of spatial suppression surrounding the focus of visual attention

“…Deep learning networks have been successful in yielding state-of-the-art image recognition and have done so through a hierarchical organization of feature extraction similar to the cortical layers of the human visual system (Cadieu et al, 2014), lending support to its applicability to understand human vision and its relevance to artificial intelligence. However, natural training and visual experience occurring in human development is very different from the training data fed into current machine learning systems (Smith & Slone, 2017). For instance, unlike machine learning algorithms, human infants experience an egocentric view where not all items in the environment are in view and with many repeated occurrences of a very few items (Smith & Slone, 2017).…”

“…However, natural training and visual experience occurring in human development is very different from the training data fed into current machine learning systems (Smith & Slone, 2017). For instance, unlike machine learning algorithms, human infants experience an egocentric view where not all items in the environment are in view and with many repeated occurrences of a very few items (Smith & Slone, 2017). If machine learning aims to mimic human learning, the current study highlights yet another important factor that should be taken into consideration-that developmental differences in this key mechanism of visual attention, and other perhaps interacting anatomical and physiological mechanisms of vision (Siu & Murphy, 2018), result in ever changing and diverse visual experiences or input across development, but despite this, visual learning and the maturity of the visual system still takes place.…”

Does Exposure to Hostile Environments Predict Enhanced Emotion Detection?

“…This could be done either objectively, by quantifying the relevant statistics of people's natural environments (i.e., base rates and the extent to which emotional expressions predict particular behaviors), or subjectively by measuring people's expectations about such statistics (i.e., their priors about base rates and the predictive value of emotions). More generally, to better understand the development and utility of emotion perception styles, the field would benefit from greater investment in quantifying the actual statistics of people's lived developmental and current environments (e.g., Smith & Slone, 2017).…”

Rethinking the input: Skewed distributions of exemplars result in broad generalization in category learning

“…First, these results suggest that current categorization theories-which only consider one type of input distribution-might underestimate category generalization and the generalization process. Second, these theories derived from experimental studies using normal distributions of exemplars might be critically misaligned with human category learning in the real world, as most input distributions in the real world are skewed (Boyd & Goldberg, 2009;Clerkin et al, 2017;Smith et al, 2018;Smith & Slone, 2017). For example, when children see a skewed input distribution of objects, where there is a clear most frequent item (their favorite sippy cup) but a long tail of substantially less frequent items of the same category (wine glasses; Clerkin et al, 2017), they are creating broad, more inclusive categories.…”