Brief articleA smoothness constraint on the development of object recognition
Introduction
Object recognition is one of the most important functions of the vertebrate visual system. To date, however, the development of object recognition is poorly understood. What environmental factors cause object recognition to emerge in the newborn brain? Does this ability emerge automatically, or do newborns require a specific type of visual input in order to develop accurate object recognition abilities? These types of questions are difficult to address with humans because human infants cannot be raised in strictly controlled environments from birth. In contrast, questions that concern the role of experience in development can be addressed directly with controlled-rearing studies of newborn animals. Here, I describe a high-throughput controlled-rearing experiment that examined whether the development of object recognition requires experience with temporally smooth visual objects.
Researchers have long theorized that biological visual systems leverage the temporal smoothness of natural visual environments to recognize objects (e.g., DiCarlo et al., 2012, Feldman and Tremoulet, 2006, Foldiak, 1991, Gibson, 1979, Stone, 1996, Wallis and Rolls, 1997, Wiskott and Sejnowski, 2002). In particular, when an object moves smoothly across the visual field, the object projects a series of gradually changing images on the retina. The visual system might take advantage of this natural tendency for temporally contiguous retinal images to belong to the same object by associating patterns of neuronal activity produced by successive retinal images of an object. When provided with temporally smooth visual input, this temporal association process should create object representations that are selective for object identity and tolerant to identity-preserving image transformations (e.g., changes in viewpoint).
A wealth of studies provide evidence that mature visual systems use temporal association mechanisms to create object representations. For example, when human adults are presented with sequential views of an object, the views come to be associated with one another in a manner that aids recognition (Cox et al., 2005, Liu, 2007, Stone, 1998, Vuong and Tarr, 2004, Wallis et al., 2009, Wallis and Bülthoff, 2001). Temporal association effects have also been found on the neurophysiological level in adult monkeys (Li and DiCarlo, 2008, Li and DiCarlo, 2010, Meyer and Olson, 2011, Miyashita, 1988). In the present study, I examined whether newborn visual systems create more accurate object representations when presented with temporally smooth objects compared to temporally non-smooth objects—as predicted by temporal association models (Wallis, 1998, Wallis and Bülthoff, 2001). Specifically, I examined the first visual object representation created by newborn subjects, before their visual systems had been shaped by any prior visual object experience.
This experiment required controlling all of the subjects’ visual experiences from the onset of vision and measuring their object recognition abilities across a range of test trials. To meet these requirements, I used a high-throughput controlled-rearing method (Wood, 2013). The method involves raising newborn chicks in strictly controlled environments and recording their behavior in response to pre-programmed animations (Fig. 1A). We use the term “high-throughput” to describe the method because the controlled-rearing chambers record all of the subjects’ behavior (24/7).
I used domestic chicks as an animal model because they are an ideal model system for studying the development of vision (Wood & Wood, 2015a). First, chicks can be raised in strictly controlled environments immediately after hatching, which makes it possible to control all of their visual object experiences. Second, chicks imprint to objects seen in the first days of life. This imprinting behavior can be used to test chicks’ object recognition abilities without training (Bateson, 2000, Horn, 2004). Third, birds and mammals process sensory input using homologous neural circuits with similar connectivity patterns (reviewed by Jarvis et al., 2005, Karten, 2013). Since birds and mammals use homologous neural mechanisms to process visual input, controlled-rearing studies of newborn chicks can inform our understanding of the development of both avian and mammalian vision. Finally, chicks develop visual recognition abilities rapidly (Vallortigara, 2012). For example, newborn chicks can begin recognizing objects (Wood, 2013, Wood, 2015), faces (Wood & Wood, 2015b), and actions (Goldman & Wood, 2015) at the onset of vision. Newborn chicks can also build integrated object representations with bound color-shape units (Wood, 2014).
In the first week of life (input phase), newborn chicks were raised in environments that contained no objects other than a single virtual object (Fig. 1A). For one group of chicks, the virtual object moved smoothly over time (Temporally Smooth Condition), whereas for another group of chicks, the virtual object moved non-smoothly over time (Temporally Non-Smooth Condition). In the second week of life (test phase), I used an automated two-alternative forced-choice procedure to test the chicks’ color recognition, shape recognition, and color-shape binding abilities.
Section snippets
Subjects
Twenty-two domestic chicks of unknown sex were tested. No subjects were excluded from the analyses. The eggs were obtained from a local distributor and incubated in darkness in an OVA-Easy incubator (Brinsea Products Inc., Titusville, FL). After hatching, the chicks were moved from the incubation room to the controlled-rearing chambers in complete darkness. Each chick was raised singly within its own chamber. Ten chicks were raised with a temporally smooth object and 12 chicks were raised with
Recognition performance
The results are shown in Fig. 3. For each test trial type, I computed the percent of time each chick spent with the imprinted object compared to the unfamiliar object. A repeated measures ANOVA with Test Trial Type as a within-subjects factor and Condition (Temporally Smooth vs. Temporally Non-Smooth) as a between-subjects factor revealed a significant main effect of Test Trial Type (F(6, 120) = 17.08, p < 0.001) and Condition (F(1, 20) = 10.99, p = .003). The interaction was not significant (F(6, 120) =
Discussion
I used a high-throughput controlled-rearing method to examine whether newborn chicks need visual experience with temporally smooth objects to develop object recognition abilities. The chicks raised with the temporally smooth objects and the chicks raised with the temporally non-smooth objects were exposed to the same individual images, and the objects were equally predictive in terms of the transitional probabilities between images; nevertheless, there were significant differences in
Acknowledgements
This research was funded by National Science Foundation CAREER Grant BCS-1351892. I thank Brian W. Wood for assistance with the supplementary movies and Samantha M.W. Wood for helpful comments on the manuscript.
References (35)
- et al.
Visual statistical learning in the newborn infant
Cognition
(2011) - et al.
How does the brain solve visual object recognition?
Neuron
(2012) - et al.
Individuation of visual objects over time
Cognition
(2006) Neocortical evolution: Neuronal circuits arise independently of lamination
Current Biology
(2013)- et al.
Visual statistical learning in infancy: Evidence for a domain general learning mechanism
Cognition
(2002) - et al.
Unsupervised natural visual experience rapidly reshapes size-invariant object representation in inferior temporal cortex
Neuron
(2010) Object recognition using spatio-temporal signatures
Vision Research
(1998)- et al.
Rotation direction affects object recognition
Vision Research
(2004) - et al.
Invariant face and object recognition in the visual system
Progress in Neurobiology
(1997) Development of motion-specific cortical responses in infancy
Vision Research
(1991)
The development of maximum displacement limits for discrimination of motion direction in infancy
Vision Research
‘Breaking’ position-invariant object recognition
Nature Neuroscience
Learning invariance from transformation sequences
Neural Computation
The ecological approach to visual perception
An automated controlled-rearing method for studying the origins of movement recognition in newly hatched chicks
Animal Cognition
Pathways of the past: The imprint of memory
Nature Reviews Neuroscience
Cited by (28)
A predictive model of indoor PM<inf>2.5</inf> considering occupancy level in a hospital outpatient hall
2022, Science of the Total EnvironmentCitation Excerpt :Deep-learning models can express more complex transformation relationships by combining simple composite functions. Object recognition is an important function of the vertebrate visual system (Wood, 2016). It aims to identify specific objects by extracting contours, textures, and shapes from complex image content.
Regularizing disentangled representations with anatomical temporal consistency
2022, Biomedical Image Synthesis and Simulation: Methods and ApplicationsOne-shot learning of view-invariant object representations in newborn chicks
2020, CognitionCitation Excerpt :Unlike humans and commonly-used animal models in psychology (e.g., rats, pigeons, monkeys), newborn chicks are precocial, require no parental care, and can be raised in strictly controlled environments from the onset of vision (Fig. 1A). With chicks it is therefore possible to study how specific visual inputs shape the development of object recognition (Wood, 2016; Wood & Wood, 2016). Newborn chicks can also be observed and tested continuously (24/7) for long periods of time, using automated image-based tracking software.