Linda Smith

Infants possess a remarkable ability to rapidly learn and process visual inputs. As an infant's mobility increases, so does the variety and dynamics of their visual inputs. Is this change in the properties of the visual inputs beneficial or even critical for the proper development of the visual system? To address this question, we used video recordings from infants wearing head-mounted cameras to train a variety of self-supervised learning models. Critically, we separated the infant data by age group and evaluated the importance of training with a curriculum aligned with developmental order. We found that initiating learning with the data from the youngest age group provided the strongest learning signal and led to the best learning outcomes in terms of downstream task performance. We then showed that the benefits of the data from the youngest age group are due to the slowness and simplicity of the visual experience. The results provide strong empirical evidence for the importance of the properties of the early infant experience and developmental progression in training. More broadly, our approach and findings take a noteworthy step towards reverse engineering the learning mechanisms in newborn brains using image-computable models from artificial intelligence.

Research on infant and toddler reaching has shown evidence for motor planning after the initiation of the reaching action. However, the reach action sequence does not begin after the initiation of a reach but rather includes the initial visual fixations onto the target object occurring before the reach. We developed a paradigm that synchronizes head‐mounted eye‐tracking and motion capture to determine whether the latency between the first visual fixation on a target object and the first reaching movement toward the object predicts subsequent reaching behavior in toddlers. In a corpus of over one hundred reach sequences produced by 17 toddlers, we found that longer fixation‐reach latencies during the pre‐reach phase predicted slower reaches. If the slowness of an executed reach indicates reach difficulty, then the duration of pre‐reach planning would be correlated with reach difficulty. However, no relation was …

Children’s early accuracy on place value (PV) tasks longitudinally predicts their later multidigit calculation skills. However, another window into children’s emerging base-ten concepts is the pattern of errors – “smart errors” – they exhibit on these measures. Past research has speculated that these smart errors – similar to invented spelling – might reflect children’s initial PV understanding that might be important for later learning of multidigit numbers and calculation. The current study examines the development of smart errors on Base-Ten Counting (invented counting errors) and Transcoding (expanded errors) in 279 U.S. kindergartners (Mage = 5.76 years) and investigated whether the presence of smart errors is associated with 1) higher concurrent levels of PV task accuracy, 2) greater growth in PV understanding over one year, 3) higher levels of multidigit calculation in second grade. Results indicate that the …

Words and images experienced by an infant wearing sensors during their daily life have led to efficient machine learning, pointing to the power of multimodal training signals and to the potentially exploitable statistics of real-life experience.

Traditionally, the exogenous control of gaze by external saliencies and the endogenous control of gaze by knowledge and context have been viewed as competing systems, with late infancy seen as a period of strengthening top‐down control over the vagaries of the input. Here we found that one‐year‐old infants control sustained attention through head movements that increase the visibility of the attended object. Freely moving one‐year‐old infants (n = 45) wore head‐mounted eye trackers and head motion sensors while exploring sets of toys of the same physical size. The visual size of the objects, a well‐documented salience, varied naturally with the infant's moment‐to‐moment posture and head movements. Sustained attention to an object was characterized by the tight control of head movements that created and then stabilized a visual size advantage for the attended object for sustained attention. The …

Visual systems must interpret the spatial arrangement of low-level properties of light in order to perceive the object created from those patterns. Extended experimental literature indicates that the spatial structure of images—and the centering of important visual information in a scene—supports feature extraction, discrimination, attention, and learning. Experiments on active egocentric vision suggest that perceivers purposely move their bodies, heads, eyes, and objects to place important information in the center of the head-centered visual field of view. However, there has been no direct quantitative assessment of the spatially-informed visual statistics of egocentric images at the scale of everyday life. Here, we introduce a new method that illustrates the perceivers' tendency to optimize contrast in the middle of head-centered images. For a grayscale image, we first quantify the local root mean square (RMS) contrast at …

Actions in the world elicit data for learning and do so in a stream of interconnected events. Here, we provide evidence on how toddlers with their parent sample information by acting on toys during exploratory play. We observed 10 min of free-flowing and unconstrained object exploration of by toddlers (mean age 21 months) and parents in a room with many available objects (n = 32). Borrowing concepts and measures from the study of narratives, we found that the toy selections are not a string of unrelated events but exhibit a suite of what we call coherence statistics: Zipfian distributions, burstiness and a network structure. We discuss the transient memory processes that underlie the moment-to-moment toy selections that create this coherence and the role of these statistics in the development of abstract and generalizable systems of knowledge. This article is part of the theme issue ‘Concepts in interaction: social …

What a perceiver sees and learns about the visual world depends intimately on their own moment-to-moment behaviors. Head, eye and postural movements–even small ones–significantly alter the images projected to the eye. These images jointly carry both the low-level visual information fundamental to cortical visual processing and high-level semantic information. Infant visual experiences are crucial to the development of all levels of the cortical hierarchy and their feed-forward and feed-back connections. Visual experiences have been assumed to be more or less uniform in terms of their low-level statistics. However, the dramatic changes in infant motor control and behavior that occur over the first year have to been shown to yield marked changes in the higher level content (eg, object categories, faces and their views) of visual experience (Adolph & Hoch, 2019; Smith et al., 2018). Here we present evidence on …

BackgroundThe Age Friendly Health System (AFHS) 4Ms framework is designed to be clinically implemented to provide reliable, high-quality age-friendly care to older adults with reduced harm and hospital-associated complications. The 4M concept of ‘What Matters to You’(WMTY) refers to engaging clinicians in assessing and aligning a patient’s specific health outcome goals and care preferences to ensure patient-centered care. There is limited empirical evidence on the WMTY concept and none describing the implementation and impact when used by clinical nurses in hospital settings.PurposeThe purpose is to gain an in-depth understanding of the knowledge, perceptions, and experiences of nurses who implement and use the WMTY concept and to evaluate the implementation and impact for older adults.MethodsA qualitative, descriptive, phenomenology study will be conducted using focus group discussions to capture key themes and achieve study aims. Research subjects will be recruited from six hospitals who have implemented the AFHS 4 Ms model including implementation leaders and a representative sample of clinical nurses who use the WMTY concept in their daily practice. After informed consent, subjects will complete a demographic survey and participate in a 90 minute virtual focus group using videoconferencing software with recording. Groups will be led by a skilled moderator using a discussion guide based on the 4Ms literature with loose, broad, and open-ended questions. Recordings will be transcribed without identifiers and uploaded into NVivo software.ResultsDescriptive statistics will be used to identify similarities and …

Visual perception emerges from a cortical hierarchy that extracts increasingly complex features, from edges to categories. Considerable computational and neural research suggests that the visual system is biased to extract slowly changing features in the input. However, little is known about the visual statistics of infant experience during early stages of receptive field formation. Here we provide evidence on the rate of change in low-level visual features and semantic features in infant everyday experience. Infants (2 to 12 months of age, n= 27) wore head cameras at home, collecting 120 hours of egocentric videos. We measured the rate of change at three levels of stimulus description: 1) raw pixels, 2) edge features (GIST, a measure of the edges at various orientations and scales), and 3) semantic features (derived from a trained CNN object classifier). For all measures, we calculated the Euclidean distance between …

Early object name learning is often conceptualized as a problem of mapping heard names to referents. However, infants do not hear object names as discrete events but rather in extended interactions organized around goal-directed actions on objects. The present study examined the statistical structure of the nonlinguistic events that surround parent naming of objects. Parents and 12-month-old infants were left alone in a room for 10 minutes with 32 objects available for exploration. Parent and infant handling of objects and parent naming of objects were coded. The four measured statistics were from measures used in the study of coherent discourse: (i) a frequency distribution in which actions were frequently directed to a few objects and more rarely to other objects; (ii) repeated returns to the high-frequency objects over the 10-minute play period; (iii) clustered repetitions and continuity of actions on objects; and (iv …

Human cognition is shaped by our bodies and actions. The influence of embodiment on cognition is particularly crucial during early development. Recent evidence shows that infants use actions to accomplish cognitive and social tasks that may later be solved internally. In our study, we propose that a sensorimotor mechanism to hand-eye coordination is through a full path from manual action, to visual saliency, and to visual attention. To provide a rigorous test of this pathway, we analyzed multimodal behavioral data collected from parent-infant toy play. We focused on linking infants’ manual actions with visual properties in the infant’s view and attention. Further, we extended our analyses to quantify the effects of manual actions on one’s own visual attention, infant’s actions on parent attention, and parent’s actions on infant attention. Results suggest that both infants’ and parents’ actions create visual saliency of objects to support visual attention and joint attention.

Purpose Human infants start to learn about and interact with their environment during the first postnatal months. Immaturities in their motor responses and spatial vision constrain their visual behavior during this period of rapid development. The goal of this study was to determine how these immaturities interact to define the visual behavior of developing infants in a naturalistic setting. Methods Participants aged 2-12 months wore head-mounted scene and binocular eye-tracking cameras (a modified PupilLabs Core system) while engaging in naturalistic behavior in a home-like environment in the lab. Calibrated eye movements were corrected for optical distortion in the head-camera images and then compiled into gaze position distributions over recordings of at least 4 minutes. These distributions were then fit with a cumulative density function over eccentricity from primary gaze. Results Videos from infants in age …

Archaeological remains can preserve some proteins into deep time, offering remarkable opportunities for probing past events in human history. Recovering functional proteins from skeletal tissues could uncover a molecular memory related to the life-history of the associated remains. We demonstrate affinity purification of whole antibody molecules from medieval human teeth, dating to the 13th–15th centuries, from skeletons with different putative pathologies. Purified antibodies are intact retaining disulphide-linkages, are amenable to primary sequences analysis, and demonstrate apparent immunoreactivity against contemporary EBV antigen on western blot. Our observations highlight the potential of ancient antibodies to provide insights into the long-term association between host immune factors and ancient microbes, and more broadly retain a molecular memory related to the natural history of human health and …

This interdisciplinary seminar brought together 18 academic and industry computer science researchers in artificial intelligence, computer vision and machine learning with 19 researchers from developmental psychology, neuroscience and linguistics. The objective was to catalyze connections between these communities, through discussions on both how the use of developmental insights can spur advances in machine learning, and how computational models and data-driven learning can lead to novel tools and insights for studying child development. The seminar consisted of tutorials, working groups, and a series of talks and discussion sessions. The main outcomes of this seminar were 1) The founding of DevelopmentalAI (http://www. developmentalai. com), an online research community to serve as a venue for communication and collaboration between develpomental and machine learning researchers, as well as a place collect and organize relevant research papers and talks; 2) Working group outputs-summaries of in-depth discussions on research questions at the intersection of developmental and machine learning, including the role of information bottlenecks and multimodality, as well as proposals for novel developmentally motivated benchmarks.

BackgroundThe scale of COVID-19 and its well documented long-term sequelae support a need to understand long-term outcomes including frailty.MethodsThis prospective cohort study recruited adults who had survived hospitalisation with clinically diagnosed COVID-19 across 35 sites in the UK (PHOSP-COVID). The burden of frailty was objectively measured using Fried's Frailty Phenotype (FFP). The primary outcome was the prevalence of each FFP group—robust (no FFP criteria), pre-frail (one or two FFP criteria) and frail (three or more FFP criteria)—at 5 months and 1 year after discharge from hospital. For inclusion in the primary analysis, participants required complete outcome data for three of the five FFP criteria. Longitudinal changes across frailty domains are reported at 5 months and 1 year post-hospitalisation, along with risk factors for frailty status. Patient-perceived recovery and health-related quality of …

The current longitudinal study measured 279 kindergartners’ (Mage=5.76 years; SD = 0.55; 135 females) place value understanding using both approximate and syntactic measures based on previous evidence that early variation in performance on approximate measures was associated with subsequent syntactic place value understanding. In the present study, we used the same dataset but traced kindergarten variation on both approximate and syntactic measures to children’s multidigit calculation skill in second grade. Path analyses indicated that latent approximate place value knowledge predicted later multidigit calculation performance but more precise understanding of counts and base-ten units did not. However, among the individual task predictors, Base-Ten Counting was the strongest predictor followed by a task asking for specific count of units (e.g. ‘Which number has [two tens]?') and reading and …

Toddlers learn words in the context of speech from adult social partners. The present studies quantitatively describe the temporal context of parent speech to toddlers about objects in individual real-world interactions. We show that at the temporal scale of a single play episode, parent talk to toddlers about individual objects is predominantly, but not always, clustered. Clustered speech is characterized by repeated references to the same object close in time, interspersed with lulls in speech about the object. Clustered temporal speech patterns mirror temporal patterns observed at longer timescales, and persisted regardless of play context. Moreover, clustered speech about individual novel objects predicted toddlers' learning of those objects' novel names. Clustered talk may be optimal for toddlers' word learning because it exploits domain-general principles of human memory and attention, principles that may have …

Place value concepts were measured longitudinally from kindergarten (2017) to first grade (2018) in a diverse sample (n = 279; Mage = 5.76 years, SD = 0.55; 135 females; 41% Black, 38% White, 8% Asian, 12% Latino). Children completed three syntactic tasks that required an explicit understanding of base‐10 symbols and three approximate tasks that could be completed without this explicit understanding. Approximate performance was significantly better in both age groups. A factor analysis confirmed that syntactic and approximate tasks tapped separate latent variables in kindergarten, but not in first grade. Path analyses indicated that only kindergarten approximate performance predicted overall first‐grade place value understanding. These findings suggest that explicit understanding of base‐10 principles develops from implicit, partial knowledge of multidigit numbers.

How do humans, including toddlers, take knowledge from past experiences and apply this knowledge in new ways? Current approaches to human and artificial intelligence (AI) fail to offer satisfactory explanations. We suggest the explanation will be found in the coherence statistics of the individual time-extended episodes of human experience and the cognitive processes those statistics engage.