Introducing v0.5 of the AI Safety Benchmark from MLCommons

This paper introduces v0.5 of the AI Safety Benchmark, which has been created by the MLCommons AI Safety Working Group. The AI Safety Benchmark has been designed to assess the safety risks of AI systems that use chat-tuned language models. We introduce a principled approach to specifying and constructing the benchmark, which for v0.5 covers only a single use case (an adult chatting to a general-purpose assistant in English), and a limited set of personas (i.e., typical users, malicious users, and vulnerable users). We created a new taxonomy of 13 hazard categories, of which 7 have tests in the v0.5 benchmark. We plan to release version 1.0 of the AI Safety Benchmark by the end of 2024. The v1.0 benchmark will provide meaningful insights into the safety of AI systems. However, the v0.5 benchmark should not be used to assess the safety of AI systems. We have sought to fully document the limitations, flaws, and challenges of v0.5. This release of v0.5 of the AI Safety Benchmark includes (1) a principled approach to specifying and constructing the benchmark, which comprises use cases, types of systems under test (SUTs), language and context, personas, tests, and test items; (2) a taxonomy of 13 hazard categories with definitions and subcategories; (3) tests for seven of the hazard categories, each comprising a unique set of test items, i.e., prompts. There are 43,090 test items in total, which we created with templates; (4) a grading system for AI systems against the benchmark; (5) an openly available platform, and downloadable tool, called ModelBench that can be used to evaluate the safety of AI systems on the benchmark; (6) an example evaluation report which benchmarks the performance of over a dozen openly available chat-tuned language models; (7) a test specification for the benchmark.

All Should Be Equal in the Eyes of Language Models: Counterfactually Aware Fair Text Generation

Fairness in Language Models (LMs) remains a longstanding challenge, given the inherent biases in training data that can be perpetuated by models and affect the downstream tasks. Recent methods employ expensive retraining or attempt debiasing during inference by constraining model outputs to contrast from a reference set of biased templates or exemplars. Regardless, they dont address the primary goal of fairness to maintain equitability across different demographic groups. In this work, we posit that inferencing LMs to generate unbiased output for one demographic under a context ensues from being aware of outputs for other demographics under the same context. To this end, we propose Counterfactually Aware Fair InferencE (CAFIE), a framework that dynamically compares the model understanding of diverse demographics to generate more equitable sentences. We conduct an extensive empirical evaluation using base LMs of varying sizes and across three diverse datasets and found that CAFIE outperforms strong baselines. CAFIE produces fairer text and strikes the best balance between fairness and language modeling capability

Towards Estimating Transferability using Hard Subsets

As transfer learning techniques are increasingly used to transfer knowledge from the source model to the target task, it becomes important to quantify which source models are suitable for a given target task without performing computationally expensive fine tuning. In this work, we propose HASTE (HArd Subset TransfErability), a new strategy to estimate the transferability of a source model to a particular target task using only a harder subset of target data. By leveraging the internal and output representations of model, we introduce two techniques, one class agnostic and another class specific, to identify harder subsets and show that HASTE can be used with any existing transferability metric to improve their reliability. We further analyze the relation between HASTE and the optimal average log likelihood as well as negative conditional entropy and empirically validate our theoretical bounds. Our experimental results across multiple source model architectures, target datasets, and transfer learning tasks show that HASTE modified metrics are consistently better or on par with the state of the art transferability metrics.

One-Shot Doc Snippet Detection: Powering Search in Document Beyond Text

Active consumption of digital documents has yielded scope for research in various applications, including search. Traditionally, searching within a document has been cast as a text matching problem ignoring the rich layout and visual cues commonly present in structured documents, forms, etc. To that end, we ask a mostly unexplored question: "Can we search for other similar snippets present in a target document page given a single query instance of a document snippet?". We propose MONOMER to solve this as a one-shot snippet detection task. MONOMER fuses context from visual, textual, and spatial modalities of snippets and documents to find query snippet in target documents. We conduct extensive ablations and experiments showing MONOMER outperforms several baselines from one-shot object detection (BHRL), template matching, and document understanding (LayoutLMv3). Due to the scarcity of relevant data for the task at hand, we train MONOMER on programmatically generated data having many visually similar query snippets and target document pairs from two datasets - Flamingo Forms and PubLayNet. We also do a human study to validate the generated data.

On Conditioning the Input Noise for Controlled Image Generation with Diffusion Models

Conditional image generation has paved the way for several breakthroughs in image editing, generating stock photos and 3-D object generation. This continues to be a significant area of interest with the rise of new state-of-the-art methods that are based on diffusion models. However, diffusion models provide very little control over the generated image, which led to subsequent works exploring techniques like classifier guidance, that provides a way to trade off diversity with fidelity. In this work, we explore techniques to condition diffusion models with carefully crafted input noise artifacts. This allows generation of images conditioned on semantic attributes. This is different from existing approaches that input Gaussian noise and further introduce conditioning at the diffusion model's inference step. Our experiments over several examples and conditional settings show the potential of our approach.

TRACE: Transform Aggregate and Compose Visiolinguistic Representations for Image Search with Text Feedback

The ability to efficiently search for images over an indexed database is the cornerstone for several user experiences. Incorporating user feedback, through multi-modal inputs provide flexible and interaction to serve fine-grained specificity in requirements. We specifically focus on text feedback, through descriptive natural language queries. Given a reference image and textual user feedback, our goal is to retrieve images that satisfy constraints specified by both of these input modalities. The task is challenging as it requires understanding the textual semantics from the text feedback and then applying these changes to the visual representation. To address these challenges, we propose a novel architecture TRACE which contains a hierarchical feature aggregation module to learn the composite visio-linguistic representations. TRACE achieves the SOTA performance on 3 benchmark datasets: FashionIQ, Shoes, and Birds-to-Words, with an average improvement of at least ~5.7%, ~3%, and ~5% respectively in R@K metric. Our extensive experiments and ablation studies show that TRACE consistently outperforms the existing techniques by significant margins both quantitatively and qualitatively.

Retrospective Loss: Looking Back to Improve Training of Deep Neural Networks

Deep neural networks (DNNs) are powerful learning machines that have enabled breakthroughs in several domains. In this work, we introduce a new retrospective loss to improve the training of deep neural network models by utilizing the prior experience available in past model states during training. Minimizing the retrospective loss, along with the task-specific loss, pushes the parameter state at the current training step towards the optimal parameter state while pulling it away from the parameter state at a previous training step. Although a simple idea, we analyze the method as well as to conduct comprehensive sets of experiments across domains - images, speech, text, and graphs - to show that the proposed loss results in improved performance across input domains, tasks, and architectures.

SieveNet: A Unified Framework for Robust Image-Based Virtual Try-On

Image-based virtual try-on for fashion has gained considerable attention recently. The task requires trying on a clothing item on a target model image. An efficient framework for this is composed of two stages: (1) warping (transforming) the try-on cloth to align with the pose and shape of the target model, and (2) a texture transfer module to seamlessly integrate the warped try-on cloth onto the target model image. Existing methods suffer from artifacts and distortions in their try-on output. In this work, we present SieveNet, a framework for robust image-based virtual try-on. Firstly, we introduce a multi-stage coarse-to-fine warping network to better model fine-grained intricacies (while transforming the try-on cloth) and train it with a novel perceptual geometric matching loss. Next, we introduce a try-on cloth conditioned segmentation mask prior to improve the texture transfer network. Finally, we also introduce a dueling triplet loss strategy for training the texture translation network which further improves the quality of the generated try-on results. We present extensive qualitative and quantitative evaluations of each component of the proposed pipeline and show significant performance improvements against the current state-of-the-art method.

AdvGAN++ : Harnessing latent layers for adversary generation

Adversarial examples are fabricated examples, indistinguishable from the original image that mislead neural networks and drastically lower their performance. Recently proposed AdvGAN, a GAN based approach, takes input image as a prior for generating adversaries to target a model. In this work, we show how latent features can serve as better priors than input images for adversary generation by proposing AdvGAN++, a version of AdvGAN that achieves higher attack rates than AdvGAN and at the same time generates perceptually realistic images on MNIST and CIFAR-10 datasets.