News

All decisions have been sent out and the list of accepted papers and the updated workshop schedule can be found below. We are super excited to see you all in Vienna!

About

The question of What makes a good representation? in machine learning can be addressed in one of several ways: By evaluating downstream behavior, by inspecting internal representations, or by characterizing a system’s inductive biases. Each of these methodologies involves measuring the alignment of an artificial intelligence (AI) system to a ground truth system (usually a human or a population of humans) at some level of analysis (be it behavior, internal representation, or something in between). However, despite this shared goal, the machine learning, neuroscience, and cognitive science communities that study alignment among artificial and biological intelligence systems currently lack a shared framework for conveying insights across methodologies and disciplines.

This workshop aims to bridge this gap by defining, evaluating, and understanding the implications of representational alignment among biological & artificial systems. We invite researchers across the machine learning, neuroscience, and cognitive science communities to contribute to this discussion in the form of invited talks, contributed papers, and structured discussions that address questions such as:

• How can we measure representational alignment among biological and artificial intelligence (AI) systems?
• Can representational alignment tell us if AI systems use the same strategies to solve tasks as humans do?
• What are the consequences (positive, neutral, and negative) of representational alignment?
• How does representational alignment connect to behavioral alignment and value alignment, as understood in AI safety and interpretability & explainability?
• How can we increase (or decrease) representational alignment of an AI system?
• How does the degree of representational alignment between two systems impact their ability to compete, cooperate, and communicate?

In collaboration with other researchers, the organizers have prepared a position paper (📄 paper) that collects prior work and highlights key issues on the topic of representational alignment. A concrete goal of the workshop is to expand this paper with any new insights generated during the workshop.

Contributed papers

Below you will find a list of all papers accepted as contributed talks and posters. The papers and their corresponding reviews can be found on the workshop's OpenReview's page.

Contributed Talks

• Beyond Sight: Probing Alignment Between Image Models and Blind V1
• Disentangling Recurrent Neural Dynamics with Stochastic Representational Geometry
• Inter-animal transforms as a guide to model-brain comparison

Poster Presentations

Poster session 1:

• The Curious Case of Representational Alignment: Unravelling Visio-Linguistic Tasks in Emergent Communication
• How aligned are different alignment metrics?
• Unsupervised alignment reveals structural commonalities and differences in neural representations of natural scenes across individuals and brain areas
• Wild Comparisons: A Study of how Representation Similarity Changes when Input Data is Drawn from a Shifted Distribution
• Biased Causal Strength Judgments in Humans and Large Language Models
• Human-like Geometric Abstraction in Large Pre-Trained Neural Networks
• Context-Sensitive Semantic Reasoning in Large Language Models
• Differentiable Optimization of Similarity Scores Between Models and Brains
• Human and Deep Neural Network Alignment in Navigational Affordance Perception
• Topoformer: brain-like topographic organization in Transformer language models through spatial querying and reweighting
• Explaining Human Comparisons using Alignment-Importance Heatmaps
• Enriching ConvNets with pre-cortical processing enhances alignment with human brain responses
• Is my "red" your "red"?: Unsupervised alignment of qualia structures via optimal transport
• Less is More: Discovering Concise Network Explanations
• The role of shared labels and experiences in representational alignment
• What Causes Polysemanticity? An Alternative Origin Story of Mixed Selectivity from Incidental Causes
• Measuring Human-CLIP Alignment at Different Abstraction Levels
• The benefits of Incorporating Shape Priors in Contrastive Learning
• ReAlnet: Achieving More Human Brain-Like Vision via Human Neural Representational Alignment
• An Analysis of Human Alignment of Latent Diffusion Models
• Unveiling the Dynamics of Transfer Learning Representations
• Lessons learned in the study of representational alignment in physical reasoning
• Measuring Mechanistic Interpretability at Scale Without Humans
• Simplicity in Complexity
• MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data
• Does resistance to style-transfer equal Global Shape Bias? Measuring network sensitivity to global shape configuration
• On convex decision regions in deep network representations
• Disentangling Recurrent Neural Dynamics with Stochastic Representational Geometry
• Modality-Agnostic fMRI Decoding of Vision and Language

Poster sessions 2:

• Improving Intervention Efficacy via Concept Realignment in Concept Bottleneck Models
• A case for sparse positive alignment of neural systems
• Self-supervised learning facilitates neural representation structures that can be unsupervisedly aligned to human behaviors
• The impact of task structure, representational geometry, and learning mechanism on compositional generalization
• Categories vs Semantic Features: What shape the similarities people discern in photographs of objects?
• Correcting Biased Centered Kernel Alignment Measures in Biological and Artificial Neural Networks
• Koopman Operator Based Dynamical Similarity Analysis for Data-driven Quantification of Distance between Dynamics
• Do Foundation Models Smell Like Humans?
• Saliency Suppressed, Semantics Surfaced: Visual Transformations in Neural Networks and the Brain
• Temperature-scaling surprisal estimates improve fit to human reading times – But does it do so for the "right reasons"?
• On the universality of neural encodings in CNNs
• Harmony in Diversity: Merging Neural Networks with Canonical Correlation Analysis
• Removing High Frequency Information Improves DNN Behavioral Alignment
• Learning and Aligning Structured Random Feature Networks
• Immediate generalisation in humans but a generalisation lag in deep neural networks—evidence for representational divergence?
• Comparing Abstraction in Humans and Large Language Models Using Multimodal Serial Reproduction
• Symbolic Variables in Distributed Networks that Count
• Humans diverge from language models when predicting spoken language
• Learning to Abstract Visuomotor Mappings using Meta-Reinforcement Learning
• Texture bias in primate ventral visual cortex
• Can Generative Multimodal Models Count to Ten?
• Towards neural foundation models for vision: Aligning EEG, MEG and fMRI representations to perform decoding, encoding and modality conversion
• Inferring DNN-brain alignment using representational similarity analyses can be problematic
• Identifying and Interpreting Non-Aligned Human Conceptual Representations using Language Modeling
• Brain-inspired L_p-Convolution benefits large kernels and aligns better with visual cortex?
• Inter-animal transforms as a guide to model-brain comparison
• Beyond Sight: Probing Alignment Between Image Models and Blind V1

Poster Format

Please follow the official ICLR guidelines regarding the size of workshop posters. Note that these are different from posters presented at the main conference.

Remote Presentation

Our goal is to allow every accepted paper to be presented in person. If our assigned space at the conference does not have the capacity for all papers to be presented as posters, we will feature any papers we cannot accommodate via presentation in an asynchronous virtual format (3 minute long videos hosted on online platform). This option will also be available to presenters who are unable to travel to Vienna due to visa issues or restrictions in terms of their available funding. Please contact us as soon as possible if you need to make use of the remote presentation format.

Speakers and Panelists

Schedule
Please see the ICLR.cc virtual site for the most up-to-date schedule. An ICLR 2024 registration is required to access the virtual site.

Organizers

Sponsor

Contact
Reach out to representational-alignment@googlegroups.com for any questions.
Cover image provided by Fabian Lackner via CC BY-SA 3.0.