Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning , which drastically reduces the number of task-speciﬁc training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM). We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly eﬃcient training across multiple TPU Pods. We demonstrate continued beneﬁts of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the ﬁnetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A signiﬁcant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

PaLM: Scaling Language Modeling with Pathways

The past 3 years of work in NLP have been characterized by the development and deployment of ever larger language models, especially for English. BERT, its variants, GPT-2/3, and others, most recently Switch-C, have pushed the boundaries of the possible both through architectural innovations and through sheer size. Using these pretrained models and the methodology of fine-tuning them for specific tasks, researchers have extended the state of the art on a wide array of tasks as measured by leaderboards on specific benchmarks for English. In this paper, we take a step back and ask: How big is too big? What are the possible risks associated with this technology and what paths are available for mitigating those risks? We provide recommendations including weighing the environmental and financial costs first, investing resources into curating and carefully documenting datasets rather than ingesting everything on the web, carrying out pre-development exercises evaluating how the planned approach fits into research and development goals and supports stakeholder values, and encouraging research directions beyond ever larger language models.

https://dl.acm.org/doi/pdf/10.1145/3442188.3445922

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? &#x1f99c;

Authored by Dr. Safiya U. Noble, an assistant professor at the University of Southern California Annenberg School of Communication, this text is one of the preeminent works that explicitly addresse...

Algorithms of Oppression: How Search Engines Reinforce Racism

We address the problem of Visual Question Answering (VQA), which requires joint image and language understanding to answer a question about a given photograph. Recent approaches have applied deep image captioning methods based on convolutional-recurrent networks to this problem, but have failed to model spatial inference. To remedy this, we propose a model we call the Spatial Memory Network and apply it to the VQA task. Memory networks are recurrent neural networks with an explicit attention mechanism that selects certain parts of the information stored in memory. Our Spatial Memory Network stores neuron activations from different spatial regions of the image in its memory, and uses attention to choose regions relevant for computing the answer. We propose a novel question-guided spatial attention architecture that looks for regions relevant to either individual words or the entire question, repeating the process over multiple recurrent steps, or “hops”. To better understand the inference process learned by the network, we design synthetic questions that specifically require spatial inference and visualize the network’s attention. We evaluate our model on two available visual question answering datasets and obtain improved results.

Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question Answering

We address the problem of Visual Question Answering (VQA), which requires joint image and language understanding to answer a question about a given photograph. Recent approaches have applied deep image captioning methods based on convolutional-recurrent networks to this problem, but have failed to model spatial inference. To remedy this, we propose a model we call the Spatial Memory Network and apply it to the VQA task. Memory networks are recurrent neural networks with an explicit attention mechanism that selects certain parts of the information stored in memory. Our Spatial Memory Network stores neuron activations from different spatial regions of the image in its memory, and uses the question to choose relevant regions for computing the answer, a process of which constitutes a single "hop" in the network. We propose a novel spatial attention architecture that aligns words with image patches in the first hop, and obtain improved results by adding a second attention hop which considers the whole question to choose visual evidence based on the results of the first hop. To better understand the inference process learned by the network, we design synthetic questions that specifically require spatial inference and visualize the attention weights. We evaluate our model on two published visual question answering datasets, DAQUAR [1] and VQA [2], and obtain improved results compared to a strong deep baseline model (iBOWIMG) which concatenates image and question features to predict the answer [3].

Building equitable and inclusive NLP technologies demands consideration of whether and how social attitudes are represented in ML models. In particular, representations encoded in models often inadvertently perpetuate undesirable social biases from the data on which they are trained. In this paper, we present evidence of such undesirable biases towards mentions of disability in two different English language models: toxicity prediction and sentiment analysis. Next, we demonstrate that the neural embeddings that are the critical first step in most NLP pipelines similarly contain undesirable biases towards mentions of disability. We end by highlighting topical biases in the discourse about disability which may contribute to the observed model biases; for instance, gun violence, homelessness, and drug addiction are over-represented in texts discussing mental illness.

Social Biases in NLP Models as Barriers for Persons with Disabilities

The challenges faced by blind people in their everyday lives are not well understood. In this paper, we report on the findings of a large-scale study of the visual questions that blind people would like to have answered. As part of this year-long study, 5,329 blind users asked 40,748 questions about photographs that they took from their iPhones using an application called VizWiz Social. We present a taxonomy of the types of questions asked, report on a number of features of the questions and accompanying photographs, and discuss how individuals changed how they used VizWiz Social over time. These results improve our understanding of the problems blind people face, and may help motivate new projects more accurately targeted to help blind people live more independently in their everyday lives.

/pdf/visual-challenges-in-the-everyday-lives-of-blind-people-2fkvvicv93.pdf

Visual challenges in the everyday lives of blind people

In this paper we introduce JustSpeak, a universal voice control solution for non-visual access to the Android operating system. JustSpeak offers two contributions as compared to existing systems. First, it enables system wide voice control on Android that can accommodate any application. JustSpeak constructs the set of available voice commands based on application context; these commands are directly synthesized from on-screen labels and accessibility metadata, and require no further intervention from the application developer. Second, it provides more efficient and natural interaction with support of multiple voice commands in the same utterance. We present the system design of JustSpeak and describe its utility in various use cases. We then discuss the system level supports required by a service like JustSpeak on other platforms. By eliminating the target locating and pointing tasks, JustSpeak can significantly improve experience of graphic interface interaction for blind and motion-impaired users.

JustSpeak: enabling universal voice control on Android

Recent work has shown the potential of having remote humans answer visual questions that blind users have. On the surface social networking sites (SNSs) offer an attractive free source of human-powered answers that can be personalized to the user. In this paper, we explore the potential of blind users asking visual questions to their social networks. We present the first formal study of how blind people use social networking sites via a survey of 191 blind adults. We also explore whether blind users find SNSs an appropriate venue for Q&A through a log analysis of questions asked using VizWiz Social, an iPhone app with over 5,000 users, which lets blind users ask questions to either the crowd or friends. We then report findings of a field experiment with 23 blind VizWiz Social users, which explored question asking on VizWiz Social in the presence of monetary costs for non-social sources. We find that blind people have a large presence on social networking sites, but do not see them as an appropriate venue for asking questions due to high perceived social costs.

/pdf/investigating-the-appropriateness-of-social-network-question-4awp4001m8.pdf

Investigating the appropriateness of social network question asking as a resource for blind users

Blind people face a range of accessibility challenges in their everyday lives, from reading the text on a package of food to traveling independently in a new place. Answering general questions about one's visual surroundings remains well beyond the capabilities of fully automated systems, but recent systems are showing the potential of engaging on-demand human workers (the crowd) to answer visual questions. The input to such systems has generally been a single image, which can limit the interaction with a worker to one question; or video streams where systems have paired the end user with a single worker, limiting the benefits of the crowd. In this paper, we introduce Chorus:View, a system that assists users over the course of longer interactions by engaging workers in a continuous conversation with the user about a video stream from the user's mobile device. We demonstrate the benefit of using multiple crowd workers instead of just one in terms of both latency and accuracy, then conduct a study with 10 blind users that shows Chorus:View answers common visual questions more quickly and accurately than existing approaches. We conclude with a discussion of users' feedback and potential future work on interactive crowd support of blind users.

/pdf/answering-visual-questions-with-conversational-crowd-nhnlt6ap2d.pdf

Yu Zhong

Papers

Social Biases in NLP Models as Barriers for Persons with Disabilities

Visual challenges in the everyday lives of blind people

JustSpeak: enabling universal voice control on Android

Investigating the appropriateness of social network question asking as a resource for blind users

Answering visual questions with conversational crowd assistants