Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers that, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination" and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated, the reliance on the visual prompt decreases, and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by 25% and 28%, respectively, and improve the accuracy on VQA benchmarks such as POPE by 21% and 24%.
Quantifying the degree of similarity between images is a key copyright issue for image-based machine learning. In legal doctrine however, determining the degree of similarity between works requires subjective analysis, and fact-finders (judges and juries) can demonstrate considerable variability in these subjective judgement calls. Images that are structurally similar can be deemed dissimilar, whereas images of completely different scenes can be deemed similar enough to support a claim of copying. We seek to define and compute a notion of "conceptual similarity" among images that captures high-level relations even among images that do not share repeated elements or visually similar components. The idea is to use a base multi-modal model to generate "explanations" (captions) of visual data at increasing levels of complexity. Then, similarity can be measured by the length of the caption needed to discriminate between the two images: Two highly dissimilar images can be discriminated early in their description, whereas conceptually dissimilar ones will need more detail to be distinguished. We operationalize this definition and show that it correlates with subjective (averaged human evaluation) assessment, and beats existing baselines on both image-to-image and text-to-text similarity benchmarks. Beyond just providing a number, our method also offers interpretability by pointing to the specific level of granularity of the description where the source data are differentiated.
We propose to extract meaning representations from autoregressive language models by considering the distribution of all possible trajectories extending an input text. This strategy is prompt-free, does not require fine-tuning, and is applicable to any pre-trained autoregressive model. Moreover, unlike vector-based representations, distribution-based representations can also model asymmetric relations (e.g., direction of logical entailment, hypernym/hyponym relations) by using algebraic operations between likelihood functions. These ideas are grounded in distributional perspectives on semantics and are connected to standard constructions in automata theory, but to our knowledge they have not been applied to modern language models. We empirically show that the representations obtained from large models align well with human annotations, outperform other zero-shot and prompt-free methods on semantic similarity tasks, and can be used to solve more complex entailment and containment tasks that standard embeddings cannot handle. Finally, we extend our method to represent data from different modalities (e.g., image and text) using multimodal autoregressive models.
We describe a first step towards learning general-purpose visual representations of physical scenes using only image prediction as a training criterion. To do so, we first define "physical scene" and show that, even though different agents may maintain different representations of the same scene, the underlying physical scene that can be inferred is unique. Then, we show that NeRFs cannot represent the physical scene, as they lack extrapolation mechanisms. Those, however, could be provided by Diffusion Models, at least in theory. To test this hypothesis empirically, NeRFs can be combined with Diffusion Models, a process we refer to as NeRF Diffusion, used as unsupervised representations of the physical scene. Our analysis is limited to visual data, without external grounding mechanisms that can be provided by independent sensory modalities.
We investigate whether prompts learned independently for different tasks can be later combined through prompt algebra to obtain a model that supports composition of tasks. We consider Visual Language Models (VLM) with prompt tuning as our base classifier and formally define the notion of prompt algebra. We propose constrained prompt tuning to improve performance of the composite classifier. In the proposed scheme, prompts are constrained to appear in the lower dimensional subspace spanned by the basis vectors of the pre-trained vocabulary. Further regularization is added to ensure that the learned prompt is grounded correctly to the existing pre-trained vocabulary. We demonstrate the effectiveness of our method on object classification and object-attribute classification datasets. On average, our composite model obtains classification accuracy within 2.5% of the best base model. On UTZappos it improves classification accuracy over the best base model by 8.45% on average.
We study the geometry of linear networks with one-dimensional convolutional layers. The function spaces of these networks can be identified with semi-algebraic families of polynomials admitting sparse factorizations. We analyze the impact of the network's architecture on the function space's dimension, boundary, and singular points. We also describe the critical points of the network's parameterization map. Furthermore, we study the optimization problem of training a network with the squared error loss. We prove that for architectures where all strides are larger than one and generic data, the non-zero critical points of that optimization problem are smooth interior points of the function space. This property is known to be false for dense linear networks and linear convolutional networks with stride one.
We introduce Train/Test-Time Adaptation with Retrieval (${\rm T^3AR}$), a method to adapt models both at train and test time by means of a retrieval module and a searchable pool of external samples. Before inference, ${\rm T^3AR}$ adapts a given model to the downstream task using refined pseudo-labels and a self-supervised contrastive objective function whose noise distribution leverages retrieved real samples to improve feature adaptation on the target data manifold. The retrieval of real images is key to ${\rm T^3AR}$ since it does not rely solely on synthetic data augmentations to compensate for the lack of adaptation data, as typically done by other adaptation algorithms. Furthermore, thanks to the retrieval module, our method gives the user or service provider the possibility to improve model adaptation on the downstream task by incorporating further relevant data or to fully remove samples that may no longer be available due to changes in user preference after deployment. First, we show that ${\rm T^3AR}$ can be used at training time to improve downstream fine-grained classification over standard fine-tuning baselines, and the fewer the adaptation data the higher the relative improvement (up to 13%). Second, we apply ${\rm T^3AR}$ for test-time adaptation and show that exploiting a pool of external images at test-time leads to more robust representations over existing methods on DomainNet-126 and VISDA-C, especially when few adaptation data are available (up to 8%).
We investigate compositional structures in vector data embeddings from pre-trained vision-language models (VLMs). Traditionally, compositionality has been associated with algebraic operations on embeddings of words from a pre-existing vocabulary. In contrast, we seek to approximate label representations from a text encoder as combinations of a smaller set of vectors in the embedding space. These vectors can be seen as "ideal words" which can be used to generate new concepts in an efficient way. We present a theoretical framework for understanding linear compositionality, drawing connections with mathematical representation theory and previous definitions of disentanglement. We provide theoretical and empirical evidence that ideal words provide good compositional approximations of composite concepts and can be more effective than token-based decompositions of the same concepts.
We introduce \`A-la-carte Prompt Tuning (APT), a transformer-based scheme to tune prompts on distinct data so that they can be arbitrarily composed at inference time. The individual prompts can be trained in isolation, possibly on different devices, at different times, and on different distributions or domains. Furthermore each prompt only contains information about the subset of data it was exposed to during training. During inference, models can be assembled based on arbitrary selections of data sources, which we call "\`a-la-carte learning". \`A-la-carte learning enables constructing bespoke models specific to each user's individual access rights and preferences. We can add or remove information from the model by simply adding or removing the corresponding prompts without retraining from scratch. We demonstrate that \`a-la-carte built models achieve accuracy within $5\%$ of models trained on the union of the respective sources, with comparable cost in terms of training and inference time. For the continual learning benchmarks Split CIFAR-100 and CORe50, we achieve state-of-the-art performance.