We present Imagen, a text-to-image diffusion model with an unprecedented degree of photorealism and deep level language understanding. Imagen builds on the power large transformer models in understanding text hinges strength high-fidelity image generation. Our key discovery is that generic (e.g. T5), pretrained text-only corpora, are surprisingly effective at encoding for synthesis: increasing size boosts both sample fidelity image-text alignment much more than model. achieves new...

We present techniques for speeding up the test-time evaluation of large convolutional networks, designed object recognition tasks. These models deliver impressive accuracy but each image requires millions floating point operations, making their deployment on smartphones and Internet-scale clusters problematic. The computation is dominated by convolution operations in lower layers model. exploit linear structure within filters to derive approximations that significantly reduce required...

Generating video frames that accurately predict future world states is challenging. Existing approaches either fail to capture the full distribution of outcomes, or yield blurry generations, both. In this paper we introduce an unsupervised generation model learns a prior uncertainty in given environment. Video are generated by drawing samples from and combining them with deterministic estimate frame. The approach simple easily trained end-to-end on variety datasets. Sample generations both...

We present a new model DrNET that learns disentangled image representations from video. Our approach leverages the temporal coherence of video and novel adversarial loss to learn representation factorizes each frame into stationary part temporally varying component. The can be used for range tasks. For example, applying standard LSTM time-vary components enables prediction future frames. evaluate our on synthetic real videos, demonstrating ability coherently generate hundreds steps future.

Building equitable and inclusive NLP technologies demands consideration of whether how social attitudes are represented in ML models. In particular, representations encoded models often inadvertently perpetuate undesirable biases from the data on which they trained. this paper, we present evidence such towards mentions disability two different English language models: toxicity prediction sentiment analysis. Next, demonstrate that neural embeddings critical first step most pipelines similarly...

Although essential to revealing biased performance, well intentioned attempts at algorithmic auditing can have effects that may harm the very populations these measures are meant protect. This concern is even more salient while biometric systems such as facial recognition, where data sensitive and technology often used in ethically questionable manners. We demonstrate a set of fiveethical concerns particular case commercial processing technology, highlighting additional design considerations...

Datasets that power machine learning are often used, shared, and reused with little visibility into the processes of deliberation led to their creation. As artificial intelligence systems increasingly used in high-stakes tasks, system development deployment practices must be adapted address very real consequences how model data is constructed practice. This includes greater transparency about data, accountability for decisions made when developing it. In this paper, we introduce a rigorous...

In response to growing concerns of bias, discrimination, and unfairness perpetuated by algorithmic systems, the datasets used train evaluate machine learning models have come under increased scrutiny. Many these examinations focused on contents datasets, finding glaring underrepresentation minoritized groups. contrast, relatively little work has been done examine norms, values, assumptions embedded in datasets. this work, we conceptualize as a type informational infrastructure, motivate...

Data is a crucial component of machine learning. The field reliant on data to train, validate, and test models. With increased technical capabilities, learning research has boomed in both academic industry settings, one major focus been computer vision. Computer vision popular domain increasingly pertinent real-world applications, from facial recognition policing object detection for autonomous vehicles. Given vision's propensity shape impact human life, we seek understand disciplinary...

Large language models (LLMs) trained on real-world data can inadvertently reflect harmful societal biases, particularly toward historically marginalized communities. While previous work has primarily focused harms related to age and race, emerging research shown that biases disabled communities exist. This study extends prior exploring the existence of by identifying categories LLM-perpetuated disability community. We conducted 19 focus groups, during which 56 participants with disabilities...

We examine the way race and racial categories are adopted in algorithmic fairness frameworks. Current methodologies fail to adequately account for socially constructed nature of race, instead adopting a conceptualization as fixed attribute. Treating an attribute, rather than structural, institutional, relational phenomenon, can serve minimize structural aspects unfairness. In this work, we focus on history turn critical theory sociological work ethnicity ground conceptualizations research,...

Understanding the content of user's image posts is a particularly interesting problem in social networks and web settings. Current machine learning techniques focus mostly on curated training sets image-label pairs, perform classification given pixels within image. In this work we instead leverage wealth information available from users: firstly, employ user hashtags to capture description content; secondly, make use valuable contextual about user. We show how metadata (age, gender, etc.)...

There is a tendency across different subfields in AI to valorize small collection of influential benchmarks. These benchmarks operate as stand-ins for range anointed common problems that are frequently framed foundational milestones on the path towards flexible and generalizable systems. State-of-the-art performance these widely understood indicative progress long-term goals. In this position paper, we explore limits such order reveal construct validity issues their framing functionally...

The ethical concept of fairness has recently been applied in machine learning (ML) settings to describe a wide range constraints and objectives. When considering the relevance concepts subset selection problems, diversity inclusion are additionally applicable order create outputs that account for social power access differentials. We introduce metrics based on these concepts, which can be together, separately, tandem with additional constraints. Results from human subject experiments lend...

Human annotated data plays a crucial role in machine learning (ML) research and development. However, the ethical considerations around processes decisions that go into dataset annotation have not received nearly enough attention. In this paper, we survey an array of literature provides insights crowdsourced annotation. We synthesize these insights, lay out challenges space along two layers: (1) who annotator is, how annotators' lived experiences can impact their annotations, (2)...

This paper presents a community-centered study of cultural limitations text-to-image (T2I) models in the South Asian context. We theorize these failures using scholarship on dominant media regimes representations and locate them within participants' reporting their existing social marginalizations. thus show how generative AI can reproduce an outsiders gaze for viewing cultures, shaped by global regional power inequities. By centering communities as experts soliciting perspectives T2I...

This paper reports on disability representation in images output from text-to-image (T2I) generative AI systems. Through eight focus groups with 25 people disabilities, we found that models repeatedly presented reductive archetypes for different disabilities. Often these representations reflected broader societal stereotypes and biases, which our participants were concerned to see reproduced through T2I. Our discussed further challenges using including the current reliance prompt engineering...

Abstract Summary: The rapidly increasing research activity focused on chromatin-mediated regulation of epigenetic mechanisms is generating waves data writers, readers and erasers the histone code, such as protein methyltransferases, bromodomains or deacetylases. To make these easily accessible to communities scientists coming from diverse horizons, we have created ChromoHub, an online resource where users can map phylogenetic trees disease associations, structures, chemical inhibitors,...

We introduce a simple semi-supervised learning approach for images based on in-painting using an adversarial loss. Images with random patches removed are presented to generator whose task is fill in the hole, surrounding pixels. The in-painted then discriminator network that judges if they real (unaltered training images) or not. This acts as regularizer standard supervised of discriminator. Using our we able directly train large VGG-style networks fashion. evaluate STL-10 and PASCAL...