Corentin Kervadec

Corentin Kervadec, Iuliia Lysova, Iuri Macocco et al.

Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate the subgraph of size s that best approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.

Emily Cheng, Corentin Kervadec, Marco Baroni

For a language model (LM) to faithfully model human language, it must compress vast, potentially infinite information into relatively few dimensions. We propose analyzing compression in (pre-trained) LMs from two points of view: geometric and information-theoretic. We demonstrate that the two views are highly correlated, such that the intrinsic geometric dimension of linguistic data predicts their coding length under the LM. We then show that, in turn, high compression of a linguistic dataset predicts rapid adaptation to that dataset, confirming that being able to compress linguistic information is an important part of successful LM performance. As a practical byproduct of our analysis, we evaluate a battery of intrinsic dimension estimators for the first time on linguistic data, showing that only some encapsulate the relationship between information-theoretic compression, geometric compression, and ease-of-adaptation.

Corentin Kervadec, Francesca Franzon, Marco Baroni

Language model prompt optimization research has shown that semantically and grammatically well-formed manually crafted prompts are routinely outperformed by automatically generated token sequences with no apparent meaning or syntactic structure, including sequences of vectors from a model's embedding space. We use machine-generated prompts to probe how models respond to input that is not composed of natural language expressions. We study the behavior of models of different sizes in multiple semantic tasks in response to both continuous and discrete machine-generated prompts, and compare it to the behavior in response to human-generated natural-language prompts. Even when producing a similar output, machine-generated and human prompts trigger different response patterns through the network processing pathways, including different perplexities, different attention and output entropy distributions, and different unit activation profiles. We provide preliminary insight into the nature of the units activated by different prompt types, suggesting that only natural language prompts recruit a genuinely linguistic circuit.

Corentin Kervadec, Iuliia Lysova, Marco Baroni et al.

Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs. Several studies on LLM efficiency optimization argue that it is possible to prune a significant portion of the parameters, while only marginally impacting performance. This suggests that the computation is not uniformly distributed across the parameters. We introduce here a technique to systematically quantify computation density in LLMs. In particular, we design a density estimator drawing on mechanistic interpretability. We experimentally test our estimator and find that: (1) contrary to what has been often assumed, LLM processing generally involves dense computation; (2) computation density is dynamic, in the sense that models shift between sparse and dense processing regimes depending on the input; (3) per-input density is significantly correlated across LLMs, suggesting that the same inputs trigger either low or high density. Investigating the factors influencing density, we observe that predicting rarer tokens requires higher density, and increasing context length often decreases the density. We believe that our computation density estimator will contribute to a better understanding of the processing at work in LLMs, challenging their symbolic interpretation.

Emily Cheng, Diego Doimo, Corentin Kervadec et al.

A language model (LM) is a mapping from a linguistic context to an output token. However, much remains to be known about this mapping, including how its geometric properties relate to its function. We take a high-level geometric approach to its analysis, observing, across five pre-trained transformer-based LMs and three input datasets, a distinct phase characterized by high intrinsic dimensionality. During this phase, representations (1) correspond to the first full linguistic abstraction of the input; (2) are the first to viably transfer to downstream tasks; (3) predict each other across different LMs. Moreover, we find that an earlier onset of the phase strongly predicts better language modelling performance. In short, our results suggest that a central high-dimensionality phase underlies core linguistic processing in many common LM architectures.

Nathanaël Carraz Rakotonirina, Corentin Kervadec, Francesca Franzon et al.

It has been widely observed that language models (LMs) respond in predictable ways to algorithmically generated prompts that are seemingly unintelligible. This is both a sign that we lack a full understanding of how LMs work, and a practical challenge, because opaqueness can be exploited for harmful uses of LMs, such as jailbreaking. We present the first thorough analysis of opaque machine-generated prompts, or autoprompts, pertaining to 6 LMs of different sizes and families. We find that machine-generated prompts are characterized by a last token that is often intelligible and strongly affects the generation. A small but consistent proportion of the previous tokens are prunable, probably appearing in the prompt as a by-product of the fact that the optimization process fixes the number of tokens. The remaining tokens fall into two categories: filler tokens, which can be replaced with semantically unrelated substitutes, and keywords, that tend to have at least a loose semantic relation with the generation, although they do not engage in well-formed syntactic relations with it. Additionally, human experts can reliably identify the most influential tokens in an autoprompt a posteriori, suggesting these prompts are not entirely opaque. Finally, some of the ablations we applied to autoprompts yield similar effects in natural language inputs, suggesting that autoprompts emerge naturally from the way LMs process linguistic inputs in general.

Pierre Marza, Corentin Kervadec, Grigory Antipov et al.

As in many tasks combining vision and language, both modalities play a crucial role in Visual Question Answering (VQA). To properly solve the task, a given model should both understand the content of the proposed image and the nature of the question. While the fusion between modalities, which is another obviously important part of the problem, has been highly studied, the vision part has received less attention in recent work. Current state-of-the-art methods for VQA mainly rely on off-the-shelf object detectors delivering a set of object bounding boxes and embeddings, which are then combined with question word embeddings through a reasoning module. In this paper, we propose an in-depth study of the vision-bottleneck in VQA, experimenting with both the quantity and quality of visual objects extracted from images. We also study the impact of two methods to incorporate the information about objects necessary for answering a question, in the reasoning module directly, and earlier in the object selection stage. This work highlights the importance of vision in the context of VQA, and the interest of tailoring vision methods used in VQA to the task at hand.

Corentin Kervadec, Christian Wolf, Grigory Antipov et al.

Methods for Visual Question Anwering (VQA) are notorious for leveraging dataset biases rather than performing reasoning, hindering generalization. It has been recently shown that better reasoning patterns emerge in attention layers of a state-of-the-art VQA model when they are trained on perfect (oracle) visual inputs. This provides evidence that deep neural networks can learn to reason when training conditions are favorable enough. However, transferring this learned knowledge to deployable models is a challenge, as much of it is lost during the transfer. We propose a method for knowledge transfer based on a regularization term in our loss function, supervising the sequence of required reasoning operations. We provide a theoretical analysis based on PAC-learning, showing that such program prediction can lead to decreased sample complexity under mild hypotheses. We also demonstrate the effectiveness of this approach experimentally on the GQA dataset and show its complementarity to BERT-like self-supervised pre-training.

Corentin Kervadec, Theo Jaunet, Grigory Antipov et al.

Since its inception, Visual Question Answering (VQA) is notoriously known as a task, where models are prone to exploit biases in datasets to find shortcuts instead of performing high-level reasoning. Classical methods address this by removing biases from training data, or adding branches to models to detect and remove biases. In this paper, we argue that uncertainty in vision is a dominating factor preventing the successful learning of reasoning in vision and language problems. We train a visual oracle and in a large scale study provide experimental evidence that it is much less prone to exploiting spurious dataset biases compared to standard models. We propose to study the attention mechanisms at work in the visual oracle and compare them with a SOTA Transformer-based model. We provide an in-depth analysis and visualizations of reasoning patterns obtained with an online visualization tool which we make publicly available (https://reasoningpatterns.github.io). We exploit these insights by transferring reasoning patterns from the oracle to a SOTA Transformer-based VQA model taking standard noisy visual inputs via fine-tuning. In experiments we report higher overall accuracy, as well as accuracy on infrequent answers for each question type, which provides evidence for improved generalization and a decrease of the dependency on dataset biases.

Theo Jaunet, Corentin Kervadec, Romain Vuillemot et al.

Visual Question Answering systems target answering open-ended textual questions given input images. They are a testbed for learning high-level reasoning with a primary use in HCI, for instance assistance for the visually impaired. Recent research has shown that state-of-the-art models tend to produce answers exploiting biases and shortcuts in the training data, and sometimes do not even look at the input image, instead of performing the required reasoning steps. We present VisQA, a visual analytics tool that explores this question of reasoning vs. bias exploitation. It exposes the key element of state-of-the-art neural models -- attention maps in transformers. Our working hypothesis is that reasoning steps leading to model predictions are observable from attention distributions, which are particularly useful for visualization. The design process of VisQA was motivated by well-known bias examples from the fields of deep learning and vision-language reasoning and evaluated in two ways. First, as a result of a collaboration of three fields, machine learning, vision and language reasoning, and data analytics, the work lead to a better understanding of bias exploitation of neural models for VQA, which eventually resulted in an impact on its design and training through the proposition of a method for the transfer of reasoning patterns from an oracle model. Second, we also report on the design of VisQA, and a goal-oriented evaluation of VisQA targeting the analysis of a model decision process from multiple experts, providing evidence that it makes the inner workings of models accessible to users.

Corentin Kervadec, Grigory Antipov, Moez Baccouche et al.

Since its appearance, Visual Question Answering (VQA, i.e. answering a question posed over an image), has always been treated as a classification problem over a set of predefined answers. Despite its convenience, this classification approach poorly reflects the semantics of the problem limiting the answering to a choice between independent proposals, without taking into account the similarity between them (e.g. equally penalizing for answering cat or German shepherd instead of dog). We address this issue by proposing (1) two measures of proximity between VQA classes, and (2) a corresponding loss which takes into account the estimated proximity. This significantly improves the generalization of VQA models by reducing their language bias. In particular, we show that our approach is completely model-agnostic since it allows consistent improvements with three different VQA models. Finally, by combining our method with a language bias reduction approach, we report SOTA-level performance on the challenging VQAv2-CP dataset.

Corentin Kervadec, Grigory Antipov, Moez Baccouche et al.

Models for Visual Question Answering (VQA) are notorious for their tendency to rely on dataset biases, as the large and unbalanced diversity of questions and concepts involved and tends to prevent models from learning to reason, leading them to perform educated guesses instead. In this paper, we claim that the standard evaluation metric, which consists in measuring the overall in-domain accuracy, is misleading. Since questions and concepts are unbalanced, this tends to favor models which exploit subtle training set statistics. Alternatively, naively introducing artificial distribution shifts between train and test splits is also not completely satisfying. First, the shifts do not reflect real-world tendencies, resulting in unsuitable models; second, since the shifts are handcrafted, trained models are specifically designed for this particular setting, and do not generalize to other configurations. We propose the GQA-OOD benchmark designed to overcome these concerns: we measure and compare accuracy over both rare and frequent question-answer pairs, and argue that the former is better suited to the evaluation of reasoning abilities, which we experimentally validate with models trained to more or less exploit biases. In a large-scale study involving 7 VQA models and 3 bias reduction techniques, we also experimentally demonstrate that these models fail to address questions involving infrequent concepts and provide recommendations for future directions of research.

Corentin Kervadec, Grigory Antipov, Moez Baccouche et al.

The large adoption of the self-attention (i.e. transformer model) and BERT-like training principles has recently resulted in a number of high performing models on a large panoply of vision-and-language problems (such as Visual Question Answering (VQA), image retrieval, etc.). In this paper we claim that these State-Of-The-Art (SOTA) approaches perform reasonably well in structuring information inside a single modality but, despite their impressive performances , they tend to struggle to identify fine-grained inter-modality relationships. Indeed, such relations are frequently assumed to be implicitly learned during training from application-specific losses, mostly cross-entropy for classification. While most recent works provide inductive bias for inter-modality relationships via cross attention modules, in this work, we demonstrate (1) that the latter assumption does not hold, i.e. modality alignment does not necessarily emerge automatically, and (2) that adding weak supervision for alignment between visual objects and words improves the quality of the learned models on tasks requiring reasoning. In particular , we integrate an object-word alignment loss into SOTA vision-language reasoning models and evaluate it on two tasks VQA and Language-driven Comparison of Images. We show that the proposed fine-grained inter-modality supervision significantly improves performance on both tasks. In particular, this new learning signal allows obtaining SOTA-level performances on GQA dataset (VQA task) with pre-trained models without finetuning on the task, and a new SOTA on NLVR2 dataset (Language-driven Comparison of Images). Finally, we also illustrate the impact of the contribution on the models reasoning by visualizing attention distributions.

Valentin Vielzeuf, Corentin Kervadec, Stéphane Pateux et al.

This paper presents a novel approach to the facial expression generation problem. Building upon the assumption of the psychological community that emotion is intrinsically continuous, we first design our own continuous emotion representation with a 3-dimensional latent space issued from a neural network trained on discrete emotion classification. The so-obtained representation can be used to annotate large in the wild datasets and later used to trained a Generative Adversarial Network. We first show that our model is able to map back to discrete emotion classes with a objectively and subjectively better quality of the images than usual discrete approaches. But also that we are able to pave the larger space of possible facial expressions, generating the many moods of emotion. Moreover, two axis in this space may be found to generate similar expression changes as in traditional continuous representations such as arousal-valence. Finally we show from visual interpretation, that the third remaining dimension is highly related to the well-known dominance dimension from psychology.

Valentin Vielzeuf, Corentin Kervadec, Stéphane Pateux et al.

This paper presents a light-weight and accurate deep neural model for audiovisual emotion recognition. To design this model, the authors followed a philosophy of simplicity, drastically limiting the number of parameters to learn from the target datasets, always choosing the simplest earning methods: i) transfer learning and low-dimensional space embedding allows to reduce the dimensionality of the representations. ii) The isual temporal information is handled by a simple score-per-frame selection process, averaged across time. iii) A simple frame selection echanism is also proposed to weight the images of a sequence. iv) The fusion of the different modalities is performed at prediction level (late usion). We also highlight the inherent challenges of the AFEW dataset and the difficulty of model selection with as few as 383 validation equences. The proposed real-time emotion classifier achieved a state-of-the-art accuracy of 60.64 % on the test set of AFEW, and ranked 4th at he Emotion in the Wild 2018 challenge.

Corentin Kervadec, Valentin Vielzeuf, Stéphane Pateux et al.

Numerous models describing the human emotional states have been built by the psychology community. Alongside, Deep Neural Networks (DNN) are reaching excellent performances and are becoming interesting features extraction tools in many computer vision tasks.Inspired by works from the psychology community, we first study the link between the compact two-dimensional representation of the emotion known as arousal-valence, and discrete emotion classes (e.g. anger, happiness, sadness, etc.) used in the computer vision community. It enables to assess the benefits -- in terms of discrete emotion inference -- of adding an extra dimension to arousal-valence (usually named dominance). Building on these observations, we propose CAKE, a 3-dimensional representation of emotion learned in a multi-domain fashion, achieving accurate emotion recognition on several public datasets. Moreover, we visualize how emotions boundaries are organized inside DNN representations and show that DNNs are implicitly learning arousal-valence-like descriptions of emotions. Finally, we use the CAKE representation to compare the quality of the annotations of different public datasets.