Canada

Systematic generalization is the ability to combine known parts into novel meaning; an important aspect of efficient human learning, but a weakness of neural network learning. In this work, we investigate how two well-known modeling principles --- modularity and data augmentation --- affect systematic generalization of neural networks in grounded language learning. We analyze how large the vocabulary needs to be to achieve systematic generalization and how similar the augmented data needs to be to the problem at hand. Our findings show that even in the controlled setting of a synthetic benchmark, achieving systematic generalization remains very difficult. After training on an augmented dataset with almost forty times more adverbs than the original problem, a non-modular baseline is not able to systematically generalize to a novel combination of a known verb and adverb. When separating the task into cognitive processes like perception and navigation, a modular neural network is able to utilize the augmented data and generalize more systematically, achieving 70% and 40% exact match increase over state-of-the-art on two gSCAN tests that have not previously been improved. We hope that this work gives insight into the drivers of systematic generalization, and what we still need to improve for neural networks to learn more like humans do.

CogSci 2022

Improving Systematic Generalization Through Modularity and Augmentation

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

poster

Join us for CogSci2022! Our 44th Annual Conference will be centered around the theme of diversity, with an accompanying range of keynotes and invited speakers who themselves highlight the diversity of our field. Our three symposia are Cognition across Diverse Populations, Cognition within Diverse Societies, and Global Challenges for Cognitive Science. The conference is being offered in a fully hybrid environment for onsite and online attendees.
 
 [**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2022/07/Cogsci-2022-PAAG-Detailed-July-12final-1.pdf)
 [**The Detailed Program can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2022/08/DeArmond-CogSci2022-program-v113-final.pdf)

This conference requires a registration to attend. Please use the link below to secure your ticket. The website will be available to all registered attendees on July 25th. 

[Register to Attend](https://cognitivesciencesociety.org/registration/)

Registration Is Required

The 44th Annual Meeting of the Cognitive Science Society highlighted research on the theme Cognitive Diversity in addition to the full breadth and diversity of research topics offered by the society’s membership. 





Although the relationship between the sound of each word and its meaning is generally arbitrary, onomatopoeias are said to have the unarbitrary link, which called sound-symbolism, between them. In this study, we investigated whether sound symbolic words are widely common in a class of onomatopoeias in some natural language. We conducted an experiment, which asked Japanese and non-Japanese speakers to match each given Japanese-onomatopoeia-like sound with a shape to which the sound referred. The result of analysis showed a similar structure for both speaker groups, in which round shapes were associated to a particular set of sounds and pointed shapes associated to the other set of sounds. Moreover, the round/ pointed shapes are correlated to pseudo onomatopoeias with sonorants/ fricative phonetic features. This finding supports the sound symbolic hypothesis asserting that the major component of Japanese onomatopoeias forms a bouba-kiki like sound-shape correspondence even for non-Japanese speakers.

Identifying a Phonetic Factors of Onomatopoeias Correlated to Sound Symbolic Commons between Japanese and Non-Japanese Speakers

English is often discussed with models that have a ‘Centre/Standard’ variety e.g., in US, and UK where English is spoken as a first language, and varieties of spoken elsewhere are regarded as ‘Outer’ or ‘Non-standard’. ‘Standard’ varieties are often regarded as prestigious and worthy of documentation, and to be taught in schools. This is problematic when the speech sounds in the local varieties of English differ from ‘Standard’ varieties. In Singapore, English is the working language among the multilingual population. Despite its importance, speech sounds in Singapore English are not well documented and not used in teaching. 60 Singaporean speakers of English completed a same/different judgement task of 206 words to investigate possible vowel mergers e.g., are ‘sheep/ship’ pronounced the ‘same’? They also read aloud the same words. We analyzed the acoustics (F1, F2) of their articulations and analyzed whether participants’ same/different judgements predicted the way they pronounced the words.

‘Sheep’ and ‘Ship’: An investigation into English vowel merger in multilingual Singapore

Contemporary models of subjective probability distortions assume that distortions arise during probability encoding. However, such assumptions are inconsistent with the ability of humans to retrieve probabilities veridically in some elicitation formats. We present a sampling-based model of probability judgment for risky prospects that assumes that probability distortions occur because people read out probability judgments as biased averages from working memory contents. Simulations demonstrate that this model shows the classic inverse-S shaped distortion of probability judgments using only retrieval-stage assumptions. The model further predicts that observers with greater working memory capacity would show larger probability distortions on average, which should lead to a particular fourfold pattern of risk preference as a function of working memory capacity. Using cognitive ability measurements as a proxy for working memory capacity, we conducted an experiment with human participants and found results consistent with the model's predictions as well as previous empirical studies. Our results support a role for sampling during assessment of risky prospects, which in turn explains differences in probability distortions seen across different elicitation methods.

Sampling-based probability construction explains individual differences in risk preference

It is intuitive to believe that humans take considerations of mental effort into account when making decisions. However, it has proved difficult to differentiate theories of mental effort in the absence of direct measurements of this psychological construct. Existing measurements of mental effort using response times and revealed preferences have low reliability. In this paper, we present a new experimental task - selecting between visuomotor lotteries using eye-tracking for sampling lotteries - that enables direct measurement of mental effort. Unlike response time-based measures, effort measurements in this task are not confounded by actual effort allocation. Unlike revealed preference-based measures, effort measurements in this task are acquired on a natural scale unitized by automatic visual selection processes. We also report results from a simple experiment conducted using this task, which reproduce existing findings of costly effort-aversion, and also demonstrate adaptive adjustment of mental effort.


Selecting between visuomotor lotteries to measure mental effort in risky decisions

Simple perceptual decision-making tasks such as the Stroop and flanker tasks are popular as a method of measuring individual variation in the processing of conflicting visual stimuli--for instance, the difference in accuracy on stimuli with and without conflict. A major challenge in applying these tasks, for instance, to compare two different populations of subjects, is the low reliability of the nonparametric measures of performance in the tasks. Here, we model dynamic adjustments in decision policies often seen in human behavior, thereby capturing trial-by-trial variation in decision policies, in addition to the classically used average statistics. We propose a recurrent network model to capture behavioral strategies in the task in a model-agnostic manner, and to overcome small-sample learning challenges by pooling across subjects. We show that by splitting the learning into a complex, shared meta-model and simple subject-specific parameters, we learn significantly better predictive models, and also identify latent dimensions indexing the decision policy that may serve as a better measure of individual differences in the task.

Meta-Learning of Dynamic Policy Adjustments in Inhibitory Control Tasks

In many real-life situations, we make decisions between a defined set of options, which can be either discrete (as when deciding between going on driving and stopping the car) or continuous (as when stirring the wheel, the possible range of angles goes from $-30$ to $30$ degrees). However, most computational models for decision-making focus on decisions between a discrete set of options. While there are a few sequential sampling models that can explain behavioral patterns (i.e., choices and response times) of decisions in a continuous option space (i.e., the CDM and the SCDM), these models have a few limitations. For example, these models assume no leakage in the evidence accumulation process and no spatial inhibition (i.e., inhibition among different areas of the option space depending on their distance to each other). In this paper, we propose a novel sequential sampling model based on an existing computational model (i.e., the leaky competing accumulator model) for decisions in a continuous option space. Our proposed model includes leakage and spatial inhibition and is thus more biologically plausible.

Radial Basis Leaky Competing Accumulator Model: A Biologically Plausible Framework for Decision-Making in a Continuous Option Space

Human reinforcement learning (RL) is characterized by different challenges. Exploration has been studied extensively in multi-armed bandits, while planning has been investigated in multi-step decision tasks. More recent work has added structured rewards to study generalization. However, past studies have often focused on a single one of these aspects, making it hard to compare results. We propose a generative model for constructing correlated trees to provide a unified and scalable method for studying exploration, planning, and generalization in a single task. In an online experiment, we find that people use structure (when provided) to generalize and perform uncertainty-directed exploration, with structure helping more in larger environments. In environments without structure, exploration becomes more random and more planning is needed. All behavioral effects are captured in a single model with recoverable parameters. In conclusion, our results connect past research on human RL in one framework using correlated trees.

Connecting Exploration, Generalization, and Planning in Correlated Trees

Many studies have shown that people may have the ability to extract summary statistics over objects/events in a set. Most of the investigations have mainly used simultaneous presented visual features. This experiment tested whether and how people could represent the average value over tactile stimuli presented in a temporal sequence. A voice-coil type vibrator was used to present the stimuli. In the averaging task, a sequence of stimuli was presented first, followed by a single stimulus. The sequence was composed of 4, 6, or 8 vibro-tactile stimuli. Participants judged whether the vibration intensity of the second stimulus was stronger or weaker than the estimated average value of the first sequence. The result demonstrated that 1) people can extract the average tactile stimuli, and 2) averaged intensity tend to be overestimated than the single stimuli. We will discuss characteristics of the summary statistical representation in the tactile modality.

Summary statistics extracted in tactile feature

Biological brains learn much more quickly than standard deep neural network reinforcement learning algorithms. One reason for this is that the deep neural networks need to learn a representation that is appropriate for the task at hand, whilst biological systems already possess an appropriate representation. Here, we bypass this problem by imposing on the neural network a representation based on what is observed in biology, such as grid cells. This study explores the impact of using a biologically-inspired grid-cell representation vs. a one-hot representation, on the speed at which a Temporal Difference-based Actor-Critic network learns to solve a simple 2D grid-world reinforcement learning task. The results suggest that the use of grid cells does promote faster learning. Furthermore, the grid cells implemented here have the potential for accurately representing unbounded continuous space. Thus, their promising performance on this discrete task acts as a first step in exploring their utility for reinforcement learning in continuous space.

Biologically-Based Neural Representations Enable Fast Online Shallow Reinforcement Learning

How do humans find their way in new environments, so quickly and efficiently? Humans reuse old knowledge to build new concepts in many non-spatial domains, such as language and drawing. Could people learn maps by a similar process, that extracts common structure to speed up learning, and generalize across maps? Understanding the computational cognitive mechanisms that support this efficiency can advance the study of the human mind and enable more efficient exploration algorithms. We hypothesize that human map learning relies on inferences over the structure of unobserved spaces, based on spatial priors informed by previous experience. We model this by combining Program Induction with a Hierarchical Bayesian framework that explicitly reasons about uncertainty through strong spatial priors. Using a new behavioral Map Induction Task, we demonstrate that this computational framework explains human exploration behavior better than non-inductive models and outperforms state-of-the-art planning algorithms in a realistic spatial navigation domain.

Premium content

Downloads

Next from CogSci 2022

Identifying a Phonetic Factors of Onomatopoeias Correlated to Sound Symbolic Commons between Japanese and Non-Japanese Speakers

Similar lecture

Systematic Generalization in Language Models Scales with Information Entropy