Abstract

Background Despite showing remarkable performance on tasks across modalities, deep neural networks remain opaque, with limited insight on how they internally organize information. A learned hierarchy is suggested in vision (Zhou et al., 2015) and language (Tenney et al., 2019) models in deep learning. Bregman (1990) describes the process of auditory scene analysis (ASA), where the auditory system abstracts information based on perceptual properties such as pitch and timbre, among others, in order to construct auditory streams that are then integrated to form higher order percepts like musical genre. This implies that the representation of audio goes through a series of hierarchical abstractions in the process. While hierarchical representations are indicated in human auditory processing (Kell et al., 2018), the same has not been thoroughly demonstrated in deep learning models for audio. Aims We study how convolutional neural networks (CNNs) represent the hierarchy of audio tasks, hypothesizing that representations at earlier layers of a CNN perform better at lower level tasks, and those at later layers perform better at higher level tasks. Methods On the basis of Bregman’s ASA model, we choose tasks that are arranged in a hierarchy from the domains of speech and music. We select three hierarchical musical tasks - note identification, instrument classification, and genre classification, representative of low-, mid-, and high-level tasks. We use class-balanced subsets of NSynth (Engel et al., 2017) with 1800 instances across 12 classes, Medley-solos-DB (Lostanlen et al., 2019) with 965 instances across 7 classes , and GTZAN (Tzanetakis et al., 2001) with 1000 instances across 10 classes, for each of the respective tasks. We inspect three models - VGGish (Hershey et al., 2017), CLAP (Elizalde et al., 2023), and MobileNetV3 (Schmid et al., 2023). We use k-Nearest Neighbour classifiers with class-balanced five-fold cross-validation to assess the model's accuracy on each task with intermediate representations extracted at six equally spaced convolutional blocks and the first fully-connected layer. We repeat this for three hierarchically related speech tasks - consonant classification, keyword recognition, and speaker count estimation, using similarly class-balanced subsets of PCVC (Malekzadeh et al., 2020) with 1794 instances across 23 classes, Speech Commands (Warden, 2018) with 1750 instances across 35 classes, LibriCount (Stöter et al., 2018) with 1100 instances across 11 classes respectively. Results Our results show that the CNN’s early layers learned low-level tasks and later layers learned high-level tasks even without explicit training. On CLAP, for note identification, the first layer’s accuracy (47%) is greater than later layers (<30%). For instrument classification, we see peak accuracy at the last layer (97%), but a remarkable change occurs at the fourth layer (71%) compared to the third layer (55%). Genre classification accuracy at the last layer (75%) is better than earlier layers (<66%). These trends hold across all models. We see similar trends for the speech tasks. Discussion The results support our hypothesis, providing strong evidence that CNNs implicitly learn a hierarchy in sound. This mirrors the human brain’s hierarchical encoding and deserves further investigation if other deep learning architectures and training methodologies also encode similar hierarchies. We also see how low-level tasks are implicitly learnt.

Abstract

Recent advances in text-to-image models have enabled a new era of creative and controllable image generation. However, generating compositional scenes with multiple subjects and attributes remains a significant challenge. To enhance user control over subject placement, several layout-guided methods have been proposed. However, these methods face numerous challenges, particularly in compositional scenes. Unintended subjects often appear outside the layouts, generated images can be out-of-distribution and contain unnatural artifacts, or attributes bleed across subjects, leading to incorrect visual outputs. In this work, we propose MALeR, a method that addresses each of these challenges. Given a text prompt and corresponding layouts, our method prevents subjects from appearing outside the given layouts while being in-distribution. Additionally, we propose a masked, attribute-aware binding mechanism that prevents attribute leakage, enabling accurate rendering of subjects with multiple attributes, even in complex compositional scenes. Qualitative and quantitative evaluation demonstrates that our method achieves superior performance in compositional accuracy, generation consistency, and attribute binding compared to previous work. MALeR is particularly adept at generating images of scenes with multiple subjects and multiple attributes per subject.

Abstract

Audio descriptions (ADs) narrate important visual details in movies, enabling Blind and Low Vision (BLV) users to understand narratives and appreciate visual details. Existing works in automatic AD generation mostly focus on few-second trimmed clips, and evaluate them by comparing against a single ground-truth reference AD. However, writing ADs is inherently subjective. Through alignment and analysis of two independent AD tracks for the same movies, we quantify the subjectivity in when and whether to describe, and what and how to highlight. Thus, we show that working with trimmed clips is inadequate. We propose ADQA, a QA benchmark that evaluates ADs at the level of few-minute long, coherent video segments, testing whether they would help BLV users understand the story and appreciate visual details. ADQA features visual appreciation (VA) questions about visual facts and narrative understanding (NU) questions based on the plot. Through ADQA, we show that current AD generation methods lag far behind human-authored ADs. We conclude with several recommendations for future work and introduce a public leaderboard for benchmarking.

Abstract

To understand a prompt, Vision-Language models (VLMs) must perceive the image, comprehend the text, and build associations within and across both modalities. For instance, given an `image of a red toy car', the model should associate this image to phrases like `car', `red toy', `red object', etc. Prior work proposed the Binding ID mechanism in LLMs, suggesting that the entity and its corresponding attribute tokens share a Binding ID in the model activations. We investigate this for image-text binding in VLMs using a synthetic dataset and task that requires models to associate 3D objects in an image with their descriptions in the text. Our experiments demonstrate that VLMs assign a distinct Binding ID to an object's image tokens and its textual references, enabling in-context association.

Abstract

Recent evaluations of LLMs on coreference resolution have revealed that traditional output formats and evaluation metrics do not fully capture the models' referential understanding. To address this, we introduce IdentifyMe, a new benchmark for mention resolution presented in a multiple-choice question (MCQ) format, commonly used for evaluating LLMs. IdentifyMe features long narratives and employs heuristics to exclude easily identifiable mentions, creating a more challenging task. The benchmark also consists of a curated mixture of different mention types and corresponding entities, allowing for a fine-grained analysis of model performance. We evaluate both closed- and open source LLMs on IdentifyMe and observe a significant performance gap (20-30%) between the state-of-the-art sub-10B open models vs. closed ones. We observe that pronominal mentions, which have limited surface information, are typically much harder for models to resolve than nominal mentions. Additionally, we find that LLMs often confuse entities when their mentions overlap in nested structures. The highest-scoring model, GPT-4o, achieves 81.9% accuracy, highlighting the strong referential capabilities of state-of-the-art LLMs while also indicating room for further improvement

Abstract

A fundamental aspect of compositional reasoning in a video is associating people and their actions across time. Recent years have seen great progress in general-purpose vision/video models and a move towards long-video understanding. While exciting, we take a step back and ask: are today’s models good at compositional reasoning on short videos? To this end, we introduce VELOCITI, a benchmark to study Video-LLMs by disentangling and assessing the comprehension of agents, actions, and their associations across multiple events. We adopt the Video-Language Entailment setup and propose StrictVLE that requires correct classification (rather than ranking) of the positive and negative caption. We evaluate several models and observe that even the best, LLaVA-OneVision (42.5%) and GPT-4o (44.3%), are far from human accuracy at 89.6%. Results show that action understanding lags behind agents, and negative captions created using entities appearing in the video perform worse than those obtained from pure text manipulation. We also present challenges with ClassicVLE and multiple-choice (MC) evaluation, strengthening our preference for StrictVLE. Finally, we validate that our benchmark requires visual inputs of multiple frames making it ideal to study video-language compositional reasoning.

Abstract

We study the connection between audio-visual observations and the underlying physics of a mundane yet intriguing everyday activity: pouring liquids. Given only the sound of liquid pouring into a container, our objective is to automatically infer physical properties such as the liquid level, the shape and size of the container, the pouring rate and the time to fill. To this end, we: (i) show in theory that these properties can be determined from the fundamental frequency (pitch); (ii) train a pitch detection model with supervision from simulated data and visual data with a physics-inspired objective; (iii) introduce a new large dataset of real pouring videos for a systematic study; (iv) show that the trained model can indeed infer these physical properties for real data; and finally, (v) we demonstrate strong generalization to various container shapes, other datasets, and in-the-wild YouTube videos. Our work presents a keen understanding of a narrow yet rich problem at the intersection of acoustics, physics, and learning. It opens up applications to enhance multisensory perception in robotic pouring.

Abstract

Image captioning systems are unable to generate fine-grained captions as they are trained on data that is either noisy (alt-text) or generic (human annotations). This is further exacerbated by maximum likelihood training that encourages generation of frequently occurring phrases. Previous works have tried to address this limitation by fine-tuning captioners with a self-retrieval (SR) reward. However, we find that SR fine-tuning has a tendency to reduce caption faithfulness and even hallucinate. In this work, we circumvent this bottleneck by improving the MLE initialization of the captioning system and designing a curriculum for the SR fine-tuning process. To this extent, we present (1) Visual Caption Boosting, a novel framework to instill fine-grainedness in generic image captioning datasets while remaining anchored in human annotations; and (2) BagCurri, a carefully designed training curriculum that more optimally leverages the contrastive nature of the self-retrieval reward. Jointly, they enable the captioner to describe fine-grained aspects in the image while preserving faithfulness to ground-truth captions. Our approach outperforms previous work by +8.9% on SR against 99 random distractors (RD100) (Dessi et al., 2023); and +7.6% on ImageCoDe. Additionally, existing metrics to evaluate captioning systems fail to reward diversity or evaluate a model's fine-grained understanding ability. Our third contribution addresses this by proposing self-retrieval from the lens of evaluation. We introduce TrueMatch, a benchmark comprising bags of highly similar images that uses SR to assess the captioner's ability to capture subtle visual distinctions. We evaluate and compare several state-of-the-art open-source MLLMs on TrueMatch, and find that our SR approach outperforms them all by a significant margin (e.g. +4.8% - 7.1% over Cambrian) while having 1-2 orders of magnitude fewer parameters. We also outperform vanilla SR by +14.4% to +19.5%.

Abstract

Understanding what makes a video memorable has important applications in advertising or education technology. Towards this goal, we investigate spatio-temporal attention mechanisms underlying video memorability. Different from previous works that fuse multiple features, we adopt a simple CNN+Transformer architecture that enables analysis of spatio-temporal attention while matching state-of-the-art (SoTA) performance on video memorability prediction. We compare model attention against human gaze fixations collected through a small-scale eye-tracking study where humans perform the video memory task. We uncover the following insights: (i) Quantitative saliency metrics show that our model, trained only to predict a memorability score, exhibits similar spatial attention patterns to human gaze, especially for more memorable videos. (ii) The model assigns greater importance to initial frames in a video, mimicking human attention patterns. (iii) Panoptic segmentation reveals that both (model and humans) assign a greater share of attention to things and less attention to stuff as compared to their occurrence probability.

Abstract

Visual Question Answering (VQA) with multiple choice questions enables a vision-centric evaluation of Multimodal Large Language Models (MLLMs). Although it reliably checks the existence of specific visual abilities, it is easier for the model to select an answer from multiple choices (VQA evaluation) than to generate the answer itself. In this work, we offer a novel perspective: we evaluate how well an MLLM understands a specific visual concept by its ability to uniquely describe two extremely similar images that differ only in the targeted visual concept. Specifically, we assess the ability of MLLMs to capture specific points of visual differences using self-retrieval, i.e., by retrieving the target image using its generated caption against the other image in the pair serving as the distractor. We curate 247 highly similar image pairs as part of the D3 benchmark. For each image pair, the model is prompted to: (1) Detect a specific visual difference, and (2) Describe the target image uniquely such that it (3) Discriminates the target image from the distractor. Self-retrieval within D3 enables whitebox evaluation across six different visual patterns, revealing that current models struggle to independently discern fine-grained visual differences, with open-source models failing to outperform random guess.