Visual Program Distillation: Distilling Tools and Programmatic Reasoning into Vision-Language Models
Solving complex visual tasks such as "Who invented the musical instrument on the right?" involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs. Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data.
相關內容
Target speaker extraction (TSE) aims to extract the target speaker's voice from the input mixture. Previous studies have concentrated on high-overlapping scenarios. However, real-world applications usually meet more complex scenarios like variable speaker overlapping and target speaker absence. In this paper, we introduces a framework to perform continuous TSE (C-TSE), comprising a target speaker voice activation detection (TSVAD) and a TSE model. This framework significantly improves TSE performance on similar speakers and enhances personalization, which is lacking in traditional diarization methods. In detail, unlike conventional TSVAD deployed to refine the diarization results, the proposed Attention-target speaker voice activation detection (A-TSVAD) directly generates timestamps of the target speaker. We also explore some different integration methods of A-TSVAD and TSE by comparing the cascaded and parallel methods. The framework's effectiveness is assessed using a range of metrics, including diarization and enhancement metrics. Our experiments demonstrate that A-TSVAD outperforms conventional methods in reducing diarization errors. Furthermore, the integration of A-TSVAD and TSE in a sequential cascaded manner further enhances extraction accuracy.
Imagining the future trajectory is the key for robots to make sound planning and successfully reach their goals. Therefore, text-conditioned video prediction (TVP) is an essential task to facilitate general robot policy learning. To tackle this task and empower robots with the ability to foresee the future, we propose a sample and computation-efficient model, named \textbf{Seer}, by inflating the pretrained text-to-image (T2I) stable diffusion models along the temporal axis. We enhance the U-Net and language conditioning model by incorporating computation-efficient spatial-temporal attention. Furthermore, we introduce a novel Frame Sequential Text Decomposer module that dissects a sentence's global instruction into temporally aligned sub-instructions, ensuring precise integration into each frame of generation. Our framework allows us to effectively leverage the extensive prior knowledge embedded in pretrained T2I models across the frames. With the adaptable-designed architecture, Seer makes it possible to generate high-fidelity, coherent, and instruction-aligned video frames by fine-tuning a few layers on a small amount of data. The experimental results on Something Something V2 (SSv2), Bridgedata and EpicKitchens-100 datasets demonstrate our superior video prediction performance with around 480-GPU hours versus CogVideo with over 12,480-GPU hours: achieving the 31% FVD improvement compared to the current SOTA model on SSv2 and 83.7% average preference in the human evaluation.
Counterfactual Explanations (CE) is the de facto method for providing insight and interpretability in black-box decision-making models by identifying alternative input instances that lead to different outcomes. This paper extends the concept of CEs to a distributional context, broadening the scope from individual data points to entire input and output distributions, named Distributional Counterfactual Explanation (DCE). In DCE, our focus shifts to analyzing the distributional properties of the factual and counterfactual, drawing parallels to the classical approach of assessing individual instances and their resulting decisions. We leverage Optimal Transport (OT) to frame a chance-constrained optimization problem, aiming to derive a counterfactual distribution that closely aligns with its factual counterpart, substantiated by statistical confidence. Our proposed optimization method, DISCOUNT, strategically balances this confidence across both input and output distributions. This algorithm is accompanied by an analysis of its convergence rate. The efficacy of our proposed method is substantiated through a series of illustrative case studies, highlighting its potential in providing deep insights into decision-making models.
This paper presents MoE-Infinity, a cost-efficient mixture-of-expert (MoE) serving system that realizes activation-aware expert offloading. MoE-Infinity features sequence-level expert activation tracing, a new approach adept at identifying sparse activations and capturing the temporal locality of MoE inference. By analyzing these traces, MoE-Infinity performs novel activation-aware expert prefetching and caching, substantially reducing the latency overheads usually associated with offloading experts for improved cost performance. Extensive experiments in a cluster show that MoE-Infinity outperforms numerous existing systems and approaches, reducing latency by 4 - 20X and decreasing deployment costs by over 8X for various MoEs. MoE-Infinity's source code is publicly available at //github.com/TorchMoE/MoE-Infinity
Loopable music generation systems enable diverse applications, but they often lack controllability and customization capabilities. We argue that enhancing controllability can enrich these models, with emotional expression being a crucial aspect for both creators and listeners. Hence, building upon LooperGP, a loopable tablature generation model, this paper explores endowing systems with control over conveyed emotions. To enable such conditional generation, we propose integrating musical knowledge by utilizing multi-granular semantic and musical features during model training and inference. Specifically, we incorporate song-level features (Emotion Labels, Tempo, and Mode) and bar-level features (Tonal Tension) together to guide emotional expression. Through algorithmic and human evaluations, we demonstrate the approach's effectiveness in producing music conveying two contrasting target emotions, happiness and sadness. An ablation study is also conducted to clarify the contributing factors behind our approach's results.
Guitar tablature is a form of music notation widely used among guitarists. It captures not only the musical content of a piece, but also its implementation and ornamentation on the instrument. Guitar Tablature Transcription (GTT) is an important task with broad applications in music education, composition, and entertainment. Existing GTT datasets are quite limited in size and scope, rendering models trained on them prone to overfitting and incapable of generalizing to out-of-domain data. In order to address this issue, we present a methodology for synthesizing large-scale GTT audio using commercial acoustic and electric guitar plugins. We procure SynthTab, a dataset derived from DadaGP, which is a vast and diverse collection of richly annotated symbolic tablature. The proposed synthesis pipeline produces audio which faithfully adheres to the original fingerings and a subset of techniques specified in the tablature, and covers multiple guitars and styles for each track. Experiments show that pre-training a baseline GTT model on SynthTab can improve transcription performance when fine-tuning and testing on an individual dataset. More importantly, cross-dataset experiments show that pre-training significantly mitigates issues with overfitting.
Image generation using generative AI is rapidly becoming a major new source of visual media, with billions of AI generated images created using diffusion models such as Stable Diffusion and Midjourney over the last few years. In this paper we collect and analyse over 3 million prompts and the images they generate. Using natural language processing, topic analysis and visualisation methods we aim to understand collectively how people are using text prompts, the impact of these systems on artists, and more broadly on the visual cultures they promote. Our study shows that prompting focuses largely on surface aesthetics, reinforcing cultural norms, popular conventional representations and imagery. We also find that many users focus on popular topics (such as making colouring books, fantasy art, or Christmas cards), suggesting that the dominant use for the systems analysed is recreational rather than artistic.
Images can convey rich semantics and induce various emotions in viewers. Recently, with the rapid advancement of emotional intelligence and the explosive growth of visual data, extensive research efforts have been dedicated to affective image content analysis (AICA). In this survey, we will comprehensively review the development of AICA in the recent two decades, especially focusing on the state-of-the-art methods with respect to three main challenges -- the affective gap, perception subjectivity, and label noise and absence. We begin with an introduction to the key emotion representation models that have been widely employed in AICA and description of available datasets for performing evaluation with quantitative comparison of label noise and dataset bias. We then summarize and compare the representative approaches on (1) emotion feature extraction, including both handcrafted and deep features, (2) learning methods on dominant emotion recognition, personalized emotion prediction, emotion distribution learning, and learning from noisy data or few labels, and (3) AICA based applications. Finally, we discuss some challenges and promising research directions in the future, such as image content and context understanding, group emotion clustering, and viewer-image interaction.
In the last years, Artificial Intelligence (AI) has achieved a notable momentum that may deliver the best of expectations over many application sectors across the field. For this to occur, the entire community stands in front of the barrier of explainability, an inherent problem of AI techniques brought by sub-symbolism (e.g. ensembles or Deep Neural Networks) that were not present in the last hype of AI. Paradigms underlying this problem fall within the so-called eXplainable AI (XAI) field, which is acknowledged as a crucial feature for the practical deployment of AI models. This overview examines the existing literature in the field of XAI, including a prospect toward what is yet to be reached. We summarize previous efforts to define explainability in Machine Learning, establishing a novel definition that covers prior conceptual propositions with a major focus on the audience for which explainability is sought. We then propose and discuss about a taxonomy of recent contributions related to the explainability of different Machine Learning models, including those aimed at Deep Learning methods for which a second taxonomy is built. This literature analysis serves as the background for a series of challenges faced by XAI, such as the crossroads between data fusion and explainability. Our prospects lead toward the concept of Responsible Artificial Intelligence, namely, a methodology for the large-scale implementation of AI methods in real organizations with fairness, model explainability and accountability at its core. Our ultimate goal is to provide newcomers to XAI with a reference material in order to stimulate future research advances, but also to encourage experts and professionals from other disciplines to embrace the benefits of AI in their activity sectors, without any prior bias for its lack of interpretability.
We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (at a rate of up to 50 characters per second). We apply our iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduce a new domain to study adversarial examples.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191