Lukas Muttenthaler

Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller et al. · berkeley, cambridge

Biological and artificial information processing systems form representations of the world that they can use to categorize, reason, plan, navigate, and make decisions. How can we measure the similarity between the representations formed by these diverse systems? Do similarities in representations then translate into similar behavior? If so, then how can a system's representations be modified to better match those of another system? These questions pertaining to the study of representational alignment are at the heart of some of the most promising research areas in contemporary cognitive science, neuroscience, and machine learning. In this Perspective, we survey the exciting recent developments in representational alignment research in the fields of cognitive science, neuroscience, and machine learning. Despite their overlapping interests, there is limited knowledge transfer between these fields, so work in one field ends up duplicated in another, and useful innovations are not shared effectively. To improve communication, we propose a unifying framework that can serve as a common language for research on representational alignment, and map several streams of existing work across fields within our framework. We also lay out open problems in representational alignment where progress can benefit all three of these fields. We hope that this paper will catalyze cross-disciplinary collaboration and accelerate progress for all communities studying and developing information processing systems.

Lukas Muttenthaler, Lorenz Linhardt, Jonas Dippel et al. · deepmind, stanford

Deep neural networks have reached human-level performance on many computer vision tasks. However, the objectives used to train these networks enforce only that similar images are embedded at similar locations in the representation space, and do not directly constrain the global structure of the resulting space. Here, we explore the impact of supervising this global structure by linearly aligning it with human similarity judgments. We find that a naive approach leads to large changes in local representational structure that harm downstream performance. Thus, we propose a novel method that aligns the global structure of representations while preserving their local structure. This global-local transform considerably improves accuracy across a variety of few-shot learning and anomaly detection tasks. Our results indicate that human visual representations are globally organized in a way that facilitates learning from few examples, and incorporating this global structure into neural network representations improves performance on downstream tasks.

Frieda Born, Tom Neuhäuser, Lukas Muttenthaler et al. · deepmind, stanford

Modern machine learning models typically represent inputs as fixed points in a high-dimensional embedding space. While this approach has been proven powerful for a wide range of downstream tasks, it fundamentally differs from the way humans process information. Because humans are constantly adapting to their environment, they represent objects and their relationships in a highly context-sensitive manner. To address this gap, we propose a method for context-sensitive similarity computation from neural network embeddings, applied to modeling a triplet odd-one-out task with an anchor image serving as simultaneous context. Modeling context enables us to achieve up to a 15% improvement in odd-one-out accuracy over a context-insensitive model. We find that this improvement is consistent across both original and "human-aligned" vision foundation models.

Lukas Muttenthaler, Klaus Greff, Frieda Born et al. · deepmind, stanford

Deep neural networks have achieved success across a wide range of applications, including as models of human behavior and neural representations in vision tasks. However, neural network training and human learning differ in fundamental ways, and neural networks often fail to generalize as robustly as humans do raising questions regarding the similarity of their underlying representations. What is missing for modern learning systems to exhibit more human-aligned behavior? We highlight a key misalignment between vision models and humans: whereas human conceptual knowledge is hierarchically organized from fine- to coarse-scale distinctions, model representations do not accurately capture all these levels of abstraction. To address this misalignment, we first train a teacher model to imitate human judgments, then transfer human-aligned structure from its representations to refine the representations of pretrained state-of-the-art vision foundation models via finetuning. These human-aligned models more accurately approximate human behavior and uncertainty across a wide range of similarity tasks, including a new dataset of human judgments spanning multiple levels of semantic abstractions. They also perform better on a diverse set of machine learning tasks, increasing generalization and out-of-distribution robustness. Thus, infusing neural networks with additional human knowledge yields a best-of-both-worlds representation that is both more consistent with human cognitive judgments and more practically useful, thus paving the way toward more robust, interpretable, and human-aligned artificial intelligence systems.

Lukas Muttenthaler, Jonas Dippel, Lorenz Linhardt et al.

Today's computer vision models achieve human or near-human level performance across a wide variety of vision tasks. However, their architectures, data, and learning algorithms differ in numerous ways from those that give rise to human vision. In this paper, we investigate the factors that affect the alignment between the representations learned by neural networks and human mental representations inferred from behavioral responses. We find that model scale and architecture have essentially no effect on the alignment with human behavioral responses, whereas the training dataset and objective function both have a much larger impact. These findings are consistent across three datasets of human similarity judgments collected using two different tasks. Linear transformations of neural network representations learned from behavioral responses from one dataset substantially improve alignment with human similarity judgments on the other two datasets. In addition, we find that some human concepts such as food and animals are well-represented by neural networks whereas others such as royal or sports-related objects are not. Overall, although models trained on larger, more diverse datasets achieve better alignment with humans than models trained on ImageNet alone, our results indicate that scaling alone is unlikely to be sufficient to train neural networks with conceptual representations that match those used by humans.

Lukas Muttenthaler, Charles Y. Zheng, Patrick McClure et al.

A central goal in the cognitive sciences is the development of numerical models for mental representations of object concepts. This paper introduces Variational Interpretable Concept Embeddings (VICE), an approximate Bayesian method for embedding object concepts in a vector space using data collected from humans in a triplet odd-one-out task. VICE uses variational inference to obtain sparse, non-negative representations of object concepts with uncertainty estimates for the embedding values. These estimates are used to automatically select the dimensions that best explain the data. We derive a PAC learning bound for VICE that can be used to estimate generalization performance or determine a sufficient sample size for experimental design. VICE rivals or outperforms its predecessor, SPoSE, at predicting human behavior in the triplet odd-one-out task. Furthermore, VICE's object representations are more reproducible and consistent across random initializations, highlighting the unique advantage of using VICE for deriving interpretable embeddings from human behavior.

Lukas Muttenthaler, Robert A. Vandermeulen, Qiuyi Zhang et al.

Model overconfidence and poor calibration are common in machine learning and difficult to account for when applying standard empirical risk minimization. In this work, we propose a novel method to alleviate these problems that we call odd-$k$-out learning (OKO), which minimizes the cross-entropy error for sets rather than for single examples. This naturally allows the model to capture correlations across data examples and achieves both better accuracy and calibration, especially in limited training data and class-imbalanced regimes. Perhaps surprisingly, OKO often yields better calibration even when training with hard labels and dropping any additional calibration parameter tuning, such as temperature scaling. We demonstrate this in extensive experimental analyses and provide a mathematical theory to interpret our findings. We emphasize that OKO is a general framework that can be easily adapted to many settings and a trained model can be applied to single examples at inference time, without significant run-time overhead or architecture changes.

Maximilian Alber, Timo Milbich, Alexandra Carpen-Amarie et al.

Pathology foundation models substantially advanced the possibilities in computational pathology -- yet tradeoffs in terms of performance, robustness, and computational requirements remained, which limited their clinical deployment. In this report, we present Atlas 2, Atlas 2-B, and Atlas 2-S, three pathology vision foundation models which bridge these shortcomings by showing state-of-the-art performance in prediction performance, robustness, and resource efficiency in a comprehensive evaluation across eighty public benchmarks. Our models were trained on the largest pathology foundation model dataset to date comprising 5.5 million histopathology whole slide images, collected from three medical institutions Charité - Universtätsmedizin Berlin, LMU Munich, and Mayo Clinic.

Laure Ciernik, Marco Morik, Lukas Thede et al.

With the rise of large-scale foundation models, efficiently adapting them to downstream tasks remains a central challenge. Linear probing, which freezes the backbone and trains a lightweight head, is computationally efficient but often restricted to last-layer representations. We show that task-relevant information is distributed across the network hierarchy rather than solely encoded in any of the last layers. To leverage this distribution of information, we apply an attentive probing mechanism that dynamically fuses representations from all layers of a Vision Transformer. This mechanism learns to identify the most relevant layers for a target task and combines low-level structural cues with high-level semantic abstractions. Across 20 diverse datasets and multiple pretrained foundation models, our method achieves consistent, substantial gains over standard linear probes. Attention heatmaps further reveal that tasks different from the pre-training domain benefit most from intermediate representations. Overall, our findings underscore the value of intermediate layer information and demonstrate a principled, task aware approach for unlocking their potential in probing-based adaptation.

Shobhita Sundaram, Stephanie Fu, Lukas Muttenthaler et al.

Humans judge perceptual similarity according to diverse visual attributes, including scene layout, subject location, and camera pose. Existing vision models understand a wide range of semantic abstractions but improperly weigh these attributes and thus make inferences misaligned with human perception. While vision representations have previously benefited from alignment in contexts like image generation, the utility of perceptually aligned representations in more general-purpose settings remains unclear. Here, we investigate how aligning vision model representations to human perceptual judgments impacts their usability across diverse computer vision tasks. We finetune state-of-the-art models on human similarity judgments for image triplets and evaluate them across standard vision benchmarks. We find that aligning models to perceptual judgments yields representations that improve upon the original backbones across many downstream tasks, including counting, segmentation, depth estimation, instance retrieval, and retrieval-augmented generation. In addition, we find that performance is widely preserved on other tasks, including specialized out-of-distribution domains such as in medical imaging and 3D environment frames. Our results suggest that injecting an inductive bias about human perceptual knowledge into vision models can contribute to better representations.

Laure Ciernik, Lorenz Linhardt, Marco Morik et al.

The Platonic Representation Hypothesis claims that recent foundation models are converging to a shared representation space as a function of their downstream task performance, irrespective of the objectives and data modalities used to train these models (Huh et al., 2024). Representational similarity is generally measured for individual datasets and is not necessarily consistent across datasets. Thus, one may wonder whether this convergence of model representations is confounded by the datasets commonly used in machine learning. Here, we propose a systematic way to measure how representational similarity between models varies with the set of stimuli used to construct the representations. We find that the objective function is a crucial factor in determining the consistency of representational similarities across datasets. Specifically, self-supervised vision models learn representations whose relative pairwise similarities generalize better from one dataset to another compared to those of image classification or image-text models. Moreover, the correspondence between representational similarities and the models' task behavior is dataset-dependent, being most strongly pronounced for single-domain datasets. Our work provides a framework for analyzing similarities of model representations across datasets and linking those similarities to differences in task behavior.

Florian P. Mahner, Lukas Muttenthaler, Umut Güçlü et al.

Determining the similarities and differences between humans and artificial intelligence (AI) is an important goal both in computational cognitive neuroscience and machine learning, promising a deeper understanding of human cognition and safer, more reliable AI systems. Much previous work comparing representations in humans and AI has relied on global, scalar measures to quantify their alignment. However, without explicit hypotheses, these measures only inform us about the degree of alignment, not the factors that determine it. To address this challenge, we propose a generic framework to compare human and AI representations, based on identifying latent representational dimensions underlying the same behavior in both domains. Applying this framework to humans and a deep neural network (DNN) model of natural images revealed a low-dimensional DNN embedding of both visual and semantic dimensions. In contrast to humans, DNNs exhibited a clear dominance of visual over semantic properties, indicating divergent strategies for representing images. While in-silico experiments showed seemingly consistent interpretability of DNN dimensions, a direct comparison between human and DNN representations revealed substantial differences in how they process images. By making representations directly comparable, our results reveal important challenges for representational alignment and offer a means for improving their comparability.

Lukas Muttenthaler, Isabelle Augenstein, Johannes Bjerva

It is challenging to automatically evaluate the answer of a QA model at inference time. Although many models provide confidence scores, and simple heuristics can go a long way towards indicating answer correctness, such measures are heavily dataset-dependent and are unlikely to generalize. In this work, we begin by investigating the hidden representations of questions, answers, and contexts in transformer-based QA architectures. We observe a consistent pattern in the answer representations, which we show can be used to automatically evaluate whether or not a predicted answer span is correct. Our method does not require any labeled data and outperforms strong heuristic baselines, across 2 datasets and 7 domains. We are able to predict whether or not a model's answer is correct with 91.37% accuracy on SQuAD, and 80.7% accuracy on SubjQA. We expect that this method will have broad applications, e.g., in the semi-automatic development of QA datasets

Lukas Muttenthaler, Nora Hollenstein, Maria Barrett

Cognitively inspired NLP leverages human-derived data to teach machines about language processing mechanisms. Recently, neural networks have been augmented with behavioral data to solve a range of NLP tasks spanning syntax and semantics. We are the first to exploit neuroscientific data, namely electroencephalography (EEG), to inform a neural attention model about language processing of the human brain. The challenge in working with EEG data is that features are exceptionally rich and need extensive pre-processing to isolate signals specific to text processing. We devise a method for finding such EEG features to supervise machine attention through combining theoretically motivated cropping with random forest tree splits. After this dimensionality reduction, the pre-processed EEG features are capable of distinguishing two reading tasks retrieved from a publicly available EEG corpus. We apply these features to regularise attention on relation classification and show that EEG is more informative than strong baselines. This improvement depends on both the cognitive load of the task and the EEG frequency domain. Hence, informing neural attention models with EEG signals is beneficial but requires further investigation to understand which dimensions are the most useful across NLP tasks.

Lukas Muttenthaler

Understanding subjectivity demands reasoning skills beyond the realm of common knowledge. It requires a machine learning model to process sentiment and to perform opinion mining. In this work, I've exploited a recently released dataset for span-selection Question Answering, namely SubjQA. SubjQA is the first QA dataset that contains questions that ask for subjective opinions corresponding to review paragraphs from six different domains. Hence, to answer these subjective questions, a learner must extract opinions and process sentiment for various domains, and additionally, align the knowledge extracted from a paragraph with the natural language utterances in the corresponding question, which together enhance the difficulty of a QA task. The primary goal of this thesis was to investigate the inner workings (i.e., latent representations) of a Transformer-based architecture to contribute to a better understanding of these not yet well understood "black-box" models. Transformer's hidden representations, concerning the true answer span, are clustered more closely in vector space than those representations corresponding to erroneous predictions. This observation holds across the top three Transformer layers for both objective and subjective questions and generally increases as a function of layer dimensions. Moreover, the probability to achieve a high cosine similarity among hidden representations in latent space concerning the true answer span tokens is significantly higher for correct compared to incorrect answer span predictions. These results have decisive implications for down-stream applications, where it is crucial to know about why a neural network made mistakes, and in which point, in space and time the mistake has happened (e.g., to automatically predict correctness of an answer span prediction without the necessity of labeled data).