United States

Human physical scene understanding requires more than simply localizing and recognizing objects — we can quickly adapt our predictions about how a scene will unfold by incorporating objects&#39; latent physics properties, such as the masses of the objects in the scene. What are the underlying computational mechanisms that allow humans to infer these physical properties and adapt their physical predictions so efficiently from visual inputs? One hypothesis is that general intuitive physics knowledge can be learned from enough raw data, instantiated as computational models that predict future video frames in large datasets of complex scenes. To test this hypothesis, we evaluate existing state-of-the-art video models. We measured both model and human performance on Physion++, a novel dataset and benchmark that rigorously evaluates visual physical prediction in humans and machines, under circumstances where accurate physical prediction relies on accurate estimates of the latent physical properties of objects in the scene. Specifically, we tested scenarios where accurate prediction relied on accurate estimates of objects&#39; mechanical properties, including masses, friction, elasticity and deformability, and the values of these mechanical properties could only be inferred by observing how these objects moved and interacted with other objects and/or fluids. We found that models that encode objectness and physical states tend to perform better, yet there is still a huge gap compared to human performance. We also found most models&#39; predictions correlate poorly with that made by humans. These results show that current deep learning models that succeed in some settings nevertheless fail to achieve human-level physical prediction in other cases, especially those where latent property inference is required.

CogSci 2023

Physion++: Evaluating Physical Scene Understanding with Objects Consisting of Different Physical Attributes in Humans and Machines

Human physical scene understanding requires more than simply localizing and recognizing objects — we can quickly adapt our predictions about how a scene will unfold by incorporating objects' latent physics properties, such as the masses of the objects in the scene. What are the underlying computational mechanisms that allow humans to infer these physical properties and adapt their physical predictions so efficiently from visual inputs? One hypothesis is that general intuitive physics knowledge can be learned from enough raw data, instantiated as computational models that predict future video frames in large datasets of complex scenes. To test this hypothesis, we evaluate existing state-of-the-art video models. We measured both model and human performance on Physion++, a novel dataset and benchmark that rigorously evaluates visual physical prediction in humans and machines, under circumstances where accurate physical prediction relies on accurate estimates of the latent physical properties of objects in the scene. Specifically, we tested scenarios where accurate prediction relied on accurate estimates of objects' mechanical properties, including masses, friction, elasticity and deformability, and the values of these mechanical properties could only be inferred by observing how these objects moved and interacted with other objects and/or fluids. We found that models that encode objectness and physical states tend to perform better, yet there is still a huge gap compared to human performance. We also found most models' predictions correlate poorly with that made by humans. These results show that current deep learning models that succeed in some settings nevertheless fail to achieve human-level physical prediction in other cases, especially those where latent property inference is required.

poster

The 45th Annual Meeting of the Cognitive Science Society will take place in Sydney, Australia as a modified hybrid conference. Presentations from the main-session room will be live-streamed every day for all events. This will include all invited talks and award sessions. Recordings of all CogSci 2023 sessions will be made available for on-demand viewing after the conference.<br>
The in-person component will be held at the International Convention Centre Sydney at Darling Harbour. The Conference will take place Wednesday July 26th – Saturday July 29th, 2023 (Wednesday, July 26th is the Workshops Day).

Join us!

<br>[**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2023/07/COgSci-PAAG-V2.pdf)
<br>[**The Detailed Program can be found here**](https://online.fliphtml5.com/ebtyf/mswz/)
<br><br>

#### **CogSci 2023 Program Book**
<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/tzgk/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>
<br>

The 45th Annual Meeting of the Cognitive Science Society presents the latest research across cognitive science and highlights the theme of Cognition in Context.

Agents trained using deep reinforcement learning (DRL) are capable of meeting or exceeding human-levels of performance in multi-agent tasks. However, the behaviors exhibited by these agents are not guaranteed to be human-like or human-compatible. This poses a problem if the goal is to design agents capable of collaborating with humans in cooperative or team-based tasks. Previous approaches to encourage the development of human-compatible agents have relied on pre-recorded human data during training. However, such data is not available for the majority of everyday tasks. Importantly, research on human perceptual-motor behavior has found that task-directed behavior is often low-dimensional and can be decomposed into a defined set of dynamical perceptual-motor primitives (DPMPs). Accordingly, we propose a hierarchical approach to simplify DRL training by defining the action dynamics of agents using DPMPs at the lower level, while using DRL to train the decision-making dynamics of agents at the higher level. We evaluate our approach using a multi-agent shepherding task used to study human and multi-agent coordination. Our hierarchical, DRL-DPMP approach resulted in agents which trained faster than vanilla, black-box DRL agents. Further, the hierarchical agents reached higher levels of performance not only when interacting with each other during self-play, but also when completing the task alongside agents embodying models of novice and expert human behavior. Finally, the hierarchical DRL-DPMP agents developed decision-making policies that outperformed heuristic-based agents used in previous research in human-agent coordination.

Scaffolding Deep Reinforcement Learning Agents using Dynamical Perceptual-Motor Primitives

Our memories are stored not in isolation, but in richly structured networks of knowledge. How are individual memories represented within these knowledge structures, and how do new experiences become integrated into existing structures? To address these questions, we exposed participants to novel objects from a hierarchical category structure. After participants learned these objects to criterion, we introduced new objects from the same underlying structure. Participants learned the new objects much more quickly than the old set, suggesting that recently learned knowledge rapidly scaffolded the learning of new, related information. However, fine-grained analyses of participants’ memory errors suggest these newly learned objects may not be fully integrated into prior learned category structure, but rather are held separate. These results suggest that existing knowledge can facilitate the learning of new information in the absence of assimilation of that information. Ongoing work investigates how these dynamics unfold over time and with sleep.


Measuring the rapid acquisition and integration of structured knowledge

One of the most ancient and universal forms of cognitive offloading is the outsourcing of cognitive operations onto other humans. We explored mechanisms underlying this ability by designing a novel computerised memory task where participants memorised target locations and recalled them after a brief delay. Next, they watched two virtual people compete in another memory game, where one demonstrated a reliable memory, and the other demonstrated an unreliable memory. Finally, participants completed the initial memory task again, with either the reliable-memory or unreliable-memory person being available to help with recall on each trial. Through observation and without direct instruction, participants acquired beliefs about the virtual people’s cognitive proficiencies and could readily draw upon these beliefs to inform offloading decisions. Participants were more likely to seek help from the reliable-memory person, and this tendency interacted with factors known to drive cognitive offloading more generally, like task difficulty and unaided cognitive ability.

Can you help me? Using others to offload cognition

How much can we infer about the past and future, given our knowledge of the present? Unlike temporally symmetric inferences about simple sequences, inferences about our own lives are asymmetric: we are better able to infer the past than the future, since we remember our past but not our future (i.e., the psychological arrow of time). What happens when both the past and future are unobserved, as when we make inferences about other people’s lives? We had participants view segments of a character-driven television drama. They wrote out what would happen just before or after each just-watched segment. Participants were better at inferring past (versus future) events. This asymmetry was driven by participants’ reliance on characters’ conversational references in the narrative, which tended to favor the past. Our work reveals a temporal asymmetry in how observations of other people’s behaviors can inform us about the past and future.

Temporal asymmetries in inferring unobserved past and future events

Many empirical domains center around acts of abstract perception. In this work, we are exploring the periphery of perceptual experiences, specifically the ways professional mathematicians develop their internal visualizations within domains of particular expertise. In their core domains, mathematicians report having highly vivid perceptual experiences of impossible objects such as infinite sets or multi-dimensional spaces. We apply an empirical lens which has formerly been used to understand deliberate visualization practices in Shamanic and Christian contexts to understand how mathematicians develop vividness in their visualizations, and how they exert control over them. 
We find that every mathematician in our sample engaged in these multimodal visualizations. We found that shape, movement, and color were much more common qualities than tactile and olfactory qualities. Many mathematicians reported their visualizations to be dynamic and at least partially under their control. Furthermore, visualizations tended to increase in vividness across experience within a domain.

Expert Mathematicians Experience Multi-Modal Visualizations that Develop Across Time

The ‘Self’ has a prioritized cognitive status, attributed to an automatic bottom-up attentional enhancement for self-relevant stimuli. Two predictions follow if self-relevant information is automatically boosted. First, processing should be enhanced for self- comparative to other-relevant targets. Second, interference should be greater for self- compared to other-relevant distractors. To investigate these predictions, we adapted a motion discrimination task. Participants first learnt to associate a colour (pink, blue) with themselves and a stranger (other), then viewed a label (YOU or OTHER) and two different coloured superimposed random dot kinematograms (RDKs; blue, pink). A response dial recorded participants’ reproduced direction of motion for the coloured RDK associated with the presented label. Facilitation / interference for self- and other-labelled stimuli was assessed by the angular difference between the reported and true direction of motion (signed error magnitude). Further regression-based analyses quantified the weight self and other-related stimuli influenced responses (decision weights). Contrary to predictions, the mean signed error magnitude and distractor weights did not significantly differ between self and other-relevant stimuli, suggesting self-relevance did not modulate the degree of interference and self-relevant stimuli did not automatically capture attention. Nevertheless, decision weights for target stimuli showed a significant advantage for self- compared with other-relevant motions. This suggests attention enhances feature-based selective attention for self-relevant motion vectors when task relevant. Hence, the ‘Self’s prioritized status is potentially due to target facilitation.

Self-relevance Facilitates Attention to Self-associated Targets on Feature-based Selective Attention Tasks

A commonly held assumption is that demonstration and pantomime differ from
ordinary action in that the movements are slowed down and exaggerated to be better
understood by intended receivers. This claim has, however, been based on meager
empirical support. This article provides evidence that different
functional demands of demonstration and pantomime result in motion characteristics
that differ from those for praxic action. In the experiment, participants were dressed
in motion capture suits and asked to (1) perform an action, (2) demonstrate this
action so that somebody else could learn how to perform it, (3) pantomime this
action without using the object so that somebody else could learn how to perform it,
and (4) pantomime this action without using the object so that somebody else could
distinguish it from another action. The results confirm that actors exaggerate and
slow down their movements in demonstrations and pantomimes when compared to
ordinary actions.


Praxis, Demonstration and Pantomime: A Motion Capture Investigation of Differences in Action Performances

Stimuli that are physically salient—e.g., brighter or differently colored to others in the visual scene—capture eye gaze and attention. Many studies have shown that color-singleton distractors slow visual search for a target, even when color is irrelevant to the task at hand. In the current study we investigated whether participants could use pre-trial information about the color of an upcoming distractor to prevent themselves from getting distracted by it. To ensure that participants were motivated to try and use the pre-trial information, we also provided performance-contingent reward on some trials. Results showed that reward reduced distraction by physically salient stimuli relative to the baseline (distractor-absent condition). However, participants could not capitalize on information provided about the color of the upcoming distractor to further improve visual search performance. This study highlights the fast and reflexive nature of attentional capture by physically salient distractors, which is difficult to control strategically, even when motivated to do so.

Physically salient stimuli capture attention despite external motivation to ignore

Existing studies have suggested that there may be a link between perception and production of speech sounds. In this paper we examine this link using data collected on Singapore English-Mandarin bilingual adults. We used a 2AFC task to measure perception of the Singapore Mandarin [tsʰuə́n]-[ʈʂʰuə́n] alveolar-retroflex contrast, and measured production using centre of gravity (CoG) measurements on speech recordings of alveolar and retroflex words. We found that a perception-production link may exist in our participants, with participants who exhibit less retroflexion in speech showing less sensitivity to retroflexion in ambiguous speech sounds.

Do you say what you hear? Perception-production link of a phoneme contrast in Singapore Mandarin

We propose a novel approach that utilizes virtual reality (VR) and simulation environments to quantify the impact of visual impairments (VIs) on daily tasks, e.g., to what extent does glaucoma slow people down in wiping a table or chopping vegetables? We utilize clinical data from patients to develop visual field models, allowing VR to mimic the visual perception of VI patients as if it is seen through their eyes. Additionally, we leverage BEHAVIOR, the state-of-the-art simulation environment, to recreate a household environment for six daily activities. By measuring the disparity between the subjects' performance with and without VI, we can accurately quantify the impact of VIs. We further quantify the effects of VIs on body and eye movements, and model how movement strategies affect task performance under the influence of VIs. We hope our results can provide valuable insights into the challenges faced by individuals with VIs.

Downloads

Next from CogSci 2023

Scaffolding Deep Reinforcement Learning Agents using Dynamical Perceptual-Motor Primitives

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from CogSci 2023

Scaffolding Deep Reinforcement Learning Agents using Dynamical Perceptual-Motor Primitives

Downloads