United States

In this work, we develop convolutional neural generative coding (Conv-NGC), a generalization of predictive coding to the case of convolution/deconvolution-based computation. Specifically, we concretely implement a flexible neurobiologically-motivated algorithm that progressively refines latent state feature maps in order to dynamically form a more accurate internal representation/reconstruction model of natural images. The performance of the resulting sensory processing system is evaluated on complex datasets such as Color-MNIST, CIFAR-10, and Street House View Numbers (SVHN). We study the effectiveness of our brain-inspired model on the tasks of reconstruction and image denoising and find that it is competitive with convolutional auto-encoding systems trained by backpropagation of errors and outperforms them with respect to out-of-distribution reconstruction (including the full 90k CINIC-10 test set).


CogSci 2023

Convolutional Neural Generative Coding: Scaling Predictive Coding to Natural Images

technical paper

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.

Modeling human navigation and route selection in complex environments has used path planning in the past and more recently, dynamical models. The current study modeled the navigational movement trajectories of human herder agents tasked with corralling an autonomous target agent into a specified containment zone. Analyses of trajectories of participants completing a first-person herding game revealed that their movement entailed sequential (i) approach and then (ii) corral phases that were invariant to the initial location of the herder and target agents. More importantly, the trajectories of human herders could be captured using a low-dimensional environmentally coupled dynamical model. The novelty of this work is in extending the previous research in modeling navigational behaviors of humans to include moving targets that need to be manipulated into a desired position. Implications of the findings for the development of artificial agents for human-machine herding, as well as more general human-machine teaming are discussed.

Modeling Human Navigation in First-Person Herding Tasks

Educational technologies with AI are designed to personalize students’ learning, but also to alleviate teacher’s workload. However, acceptability of such technologies among teachers may be impacted by factors such as fear of replacement or ethical concerns. The purpose of the current study is to investigate attitudes of teachers towards educational tools with or without AI. The main hypothesis is that technologies with AI would be more negatively judged than technologies without AI, and thus intention to use would be weaker for technologies with AI. Results show that teachers seem to accurately perceive the potential benefit of AI technologies for reduction of workload, without feeling threatened by a replacement in the future. Ethical concerns are higher for AI technologies, but intention to use is pretty much similar. Differences between primary and secondary teachers are discussed.

Teachers’ attitudes towards AI: what is the difference with non-AI technologies?

Recent progress in artificial intelligence has renewed interest in
building machines that learn like animals. Almost all of the
work comparing learning across biological and artificial systems
comes from studies where animals and machines received
different training data, obscuring whether differences between
animals and machines emerged from differences in learning
mechanisms versus training data. We present an experimental
approach—a “newborn embodied Turing Test”—that allows
newborn animals and machines to be raised in the same
environments and tested with the same tasks, permitting direct
comparison of their learning abilities. To make this platform, we
first collected controlled-rearing data from newborn chicks, then
performed “digital twin” experiments in which machines were
raised in virtual environments that mimicked the rearing
conditions of the chicks. We found that 1) machines (deep
reinforcement learning agents with intrinsic motivation) can
spontaneously develop visually guided preference behavior, akin
to imprinting in newborn chicks, and 2) machines are still far
from newborn-level performance on object recognition tasks.
Almost all of the chicks developed view-invariant object
recognition, whereas the machines tended to develop
view-dependent recognition. Moreover, the learning outcomes
(developmental pathways) were far more constrained in the
chicks versus machines. Ultimately, we anticipate that this
approach will help researchers develop embodied AI systems
that learn like newborn animals.




A newborn embodied Turing test for view-invariant object recognition

What are the computational foundations of social grouping?
Traditional approaches to this question have focused on verbal
reasoning or simple (low-dimensional) quantitative models. In
the real world, however, social preferences emerge when
high-dimensional learning systems (brains and bodies) interact
with high-dimensional sensory inputs during an animal’s
embodied interactions with the world. A deep understanding of
social grouping will therefore require embodied models that
learn directly from sensory inputs using high-dimensional
learning mechanisms. To this end, we built artificial neural
networks (ANNs), embodied those ANNs in virtual fish bodies,
and raised the artificial fish in virtual fish tanks that mimicked
the rearing conditions of real fish. We then compared the social
preferences that emerged in real fish versus artificial fish. We
found that when artificial fish had two core learning
mechanisms (reinforcement learning and curiosity-driven
learning), artificial fish developed fish-like social preferences.
Like real fish, the artificial fish spontaneously learned to prefer
members of their own group over members of other groups. The
artificial fish also spontaneously learned to self-segregate with
their in-group, akin to self-segregation behavior seen in nature.
Our results suggest that social grouping can emerge in embodied
agents from three ingredients: (1) reinforcement learning, (2)
intrinsic motivation, and (3) early social experiences with
in-group members. This approach lays a foundation for reverse
engineering animal-like social behavior with image-computable
models, bridging the divide between high-dimensional sensory
inputs and social preferences.

Parallel development of social preferences in fish and machines

Recent years have seen remarkable advances in the development and use of Artificial Intelligence (AI) in image classification, driving cars, and writing scientific articles. Although AI can outperform humans in many tasks, there remain domains where humans and AI working together can outperform either working alone. For humans and AI to work together effectively, the human must trust the AI bot to the right degree (calibrated). If the human does not trust the bot sufficiently, or conversely trusts the bot more than is warranted, the human-bot team will not perform as well as they could. We report three experiments examining trust in human-AI teaming. While existing studies typically collect binary responses (to trust, or not to trust), we present a novel paradigm that quantifies trust in a bot-recommendation in a continuous fashion. These data allow better precision, and in the future the development of more refined models of human-bot trust.

Trust in Human-bot Teaming: Applications of the Judge Advisor System

With AIs becoming prevalent in our daily lives, it is still controversial whether they are a reliable conversational assistant. We, therefore, developed a chatbot, a Friendly-Bot, where users can share interpersonal experiences (e.g., friendship and romantic relationships). We then manipulated chatbot appearances into two types: robot and human-looking. We found that chatbot appearance predicted users' trust. Participants preferred the robot-looking chatbot over the human-looking one. Participants also showed higher trust in robot-looking chatbots when conversing about positive interpersonal relationships. Participants showed a different pattern of trust depending on the appearance condition: positive experience led to higher trust in robot-looking condition, whereas the opposite was observed in human-looking condition. Our findings show how chatbot appearance influences rapport-building in AI-assisted interactions (e.g., counseling).

Friendly-Bot: The Impact of Chatbot Appearance and Relationship Style on User Trust

People can understand images that vary in visual abstraction---from detailed illustrations to schematic icons. 
To what degree are current vision algorithms sensitive to such variation when encoding their meaning? 
We first obtained >80K human-generated sketches produced under different time limits (4s, 8s, 16s, 32s; N=5,563 participants) and AI-generated sketches (Vinker et al., 2022) produced under different ink limits (8, 16, 32 strokes) of 2,048 real-world objects spanning 128 categories from the THINGS dataset (Hebart et al.,2019).
We then evaluated how 13 state-of-the-art vision algorithms varying in architecture and training represented the meaning of these sketches, both at the category and exemplar levels, and without additional training.
We found that these models were generally sensitive to variation in visual abstraction, although some models were substantially more sensitive than others.
Together, these findings suggest promising avenues for building better models of human visual abstraction.

Evaluating machine comprehension of sketch meaning at different levels of abstraction

Although object detection AI plays an important role in many critical systems, corresponding Explainable AI (XAI) methods remain very limited. Here we first developed FullGrad-CAM and FullGrad-CAM++ by extending traditional gradient-based methods to generate object-specific explanations with higher plausibility. Since human attention may reflect features more in-terpretable to humans, we explored the possibility to use it as guidance to learn how to combine the explanatory information in the detector model to best present as an XAI saliency map that is interpretable (plausible) to humans. Interestingly, we found that human attention maps had higher faithfulness for explaining the detector model than existing saliency-based XAI methods. By using trainable activation functions and smoothing kernels to maximize the XAI saliency map similarity to human attention maps, the generated map had higher faithfulness and plausibility than both existing XAI methods and human atten-tion maps. The learned functions were model-specific, well generalizable to other databases.

Human Attention-Guided Explainable AI for Object Detection

Punishment is a cost imposed on a target, in response to an undesirable action. Yet choosing to punish also reveals information about the authority’s own motives and values. 
We propose that observers jointly infer the wrongness of the action and the authority's motivations. Using hypothetical scenarios in unfamiliar societies, we experimentally manipulated observers' prior beliefs and measured human observers' inferences after observing punishment. These inferences were recapitulated in a formal model that inverts an intuitive causal model of authorities who make rational choices about punishment by weighing its costs and benefits (i.e. utilities). An essential component of this model, driving these inferences, is that legitimate authorities consider the utility of a proportional response to harmful actions, which depends on the balance between the wrongness of the act and the severity of the punishment.

What people learn from punishment: joint inference of wrongness and punisher's motivations from observation of punitive choices

Political polarization is driven by many factors, but the role of moral values as both a signal of political identity and a source of internal conflict is understudied. We report an agent-based computational model of polarization that fills this gap. Agents seek to differentiate in- and outgroup neighbors with a slight preference for the former. However, they must do so by inferring neighbors’ identities from visible but transient moral signals. Moreover, agents experience conflicts within their own values, and if difficult to resolve internally, can copy the values of their ingroup or disengage (i.e., act immorally). Results show that liberals form larger, more homogeneous clusters, are happier, and experience less moral conflict than conservatives. Conservatives experience more and higher levels of conflict, and morally disengage significantly more often than liberals.


Premium content

Convolutional Neural Generative Coding: Scaling Predictive Coding to Natural Images

Downloads

Next from CogSci 2023

Modeling Human Navigation in First-Person Herding Tasks

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES