We provide a window into the process of constructing dataset for machine learning (ML) applications by reflecting on building World Wide Dishes (WWD), an image and text consisting culinary dishes their associated customs from around world. WWD takes participatory approach to creation: community members guide design research engage in crowdsourcing efforts build dataset. responds calls ML address limitations web-scraped Internet datasets with curated, high-quality data incorporating localised...

Abstract Africa has a very poor doctor-to-patient ratio. At busy clinics, doctors could see 30+ patients per day—a heavy patient burden compared with developed countries—but productivity tools such as clinical automatic speech recognition (ASR) are lacking for these overworked clinicians. However, ASR is mature, even ubiquitous, in nations, and clinician-reported performance of commercial systems generally satisfactory. Furthermore, the recent general domain approaching human accuracy....

Generative models trained with Differential Privacy (DP) are becoming increasingly prominent in the creation of synthetic data for downstream applications. Existing literature, however, primarily focuses on basic benchmarking datasets and tends to report promising results only elementary metrics relatively simple distributions. In this paper, we initiate a systematic analysis how DP generative perform their natural application scenarios, specifically focusing real-world gene expression data....

Generative models trained with Differential Privacy (DP) are becoming increasingly prominent in the creation of synthetic data for downstream applications. Existing literature, however, primarily focuses on basic benchmarking datasets and tends to report promising results only elementary metrics relatively simple distributions. In this paper, we initiate a systematic analysis how DP generative perform their natural application scenarios, specifically focusing real-world gene expression data....

These are the proceedings of 4th workshop on Machine Learning for Developing World (ML4D), held as part Thirty-fourth Conference Neural Information Processing Systems (NeurIPS) Saturday, December 12th 2020.

This paper describes foundational efforts with SautiDB-Naija, a novel corpus of non-native (L2) Nigerian English speech. We describe how the was created and curated as well preliminary experiments accent classification learning embeddings. The initial version includes over 900 recordings from L2 speakers languages, such Yoruba, Igbo, Edo, Efik-Ibibio, Igala. further demonstrate fine-tuning on pre-trained model like wav2vec can yield representations suitable for related speech tasks...

Foundation models are increasingly ubiquitous in our daily lives, used everyday tasks such as text-image searches, interactions with chatbots, and content generation. As use increases, so does concern over the disparities performance fairness of these for different people parts world. To assess growing regional disparities, we present World Wide Dishes, a mixed text image dataset consisting 765 dishes, dish names collected 131 local languages. Dishes has been purely through human...

Recent advances in speech synthesis have enabled many useful applications like audio directions Google Maps, screen readers, and automated content generation on platforms TikTok. However, these systems are mostly dominated by voices sourced from data-rich geographies with personas representative of their source data. Although 3000 the world's languages domiciled Africa, African under-represented systems. As becomes increasingly democratized, it is desirable to increase representation English...

Recent strides in automatic speech recognition (ASR) have accelerated their application the medical domain where performance on accented named entities (NE) such as drug names, diagnoses, and lab results, is largely unknown. We rigorously evaluate multiple ASR models a clinical English dataset of 93 African accents. Our analysis reveals that despite some achieving low overall word error rates (WER), errors are higher, potentially posing substantial risks to patient safety. To empirically...

Recent strides in automatic speech recognition (ASR) have accelerated their application the medical domain where performance on accented named entities (NE) such as drug names, diagnoses, and lab results, is largely unknown. We rigorously evaluate multiple ASR models a clinical English dataset of 93 African accents. Our analysis reveals that despite some achieving low overall word error rates (WER), errors are higher, potentially posing substantial risks to patient safety. To empirically...

Recent advances in speech synthesis have enabled many useful applications like audio directions Google Maps, screen readers, and automated content generation on platforms TikTok. However, these systems are mostly dominated by voices sourced from data-rich geographies with personas representative of their source data. Although 3000 the world's languages domiciled Africa, African under-represented systems. As becomes increasingly democratized, it is desirable to increase representation English...

Gene regulatory networks (GRNs) represent the causal relationships between transcription factors (TFs) and target genes in single-cell RNA sequencing (scRNA-seq) data. Understanding these is crucial for uncovering disease mechanisms identifying therapeutic targets. In this work, we investigate potential of large language models (LLMs) GRN discovery, leveraging their learned biological knowledge alone or combination with traditional statistical methods. We develop a task-based evaluation...

Generating tabular data under differential privacy (DP) protection ensures theoretical guarantees but poses challenges for training machine learning models, primarily due to the need capture complex structures noisy supervision signals. Recently, pre-trained Large Language Models (LLMs) -- even those at scale of GPT-2 have demonstrated great potential in synthesizing data. However, their applications DP constraints remain largely unexplored. In this work, we address gap by applying...

It is perhaps no longer surprising that machine learning models, especially deep neural networks, are particularly vulnerable to attacks. One such vulnerability has been well studied model extraction: a phenomenon in which the attacker attempts steal victim's by training surrogate mimic decision boundaries of victim model. Previous works have demonstrated effectiveness an attack and its devastating consequences, but much this work done primarily for image text processing tasks. Our first...

These are the proceedings of 5th workshop on Machine Learning for Developing World (ML4D), held as part Thirty-fifth Conference Neural Information Processing Systems (NeurIPS) December 14th, 2021.

Useful conversational agents must accurately capture named entities to minimize error for downstream tasks, example, asking a voice assistant play track from certain artist, initiating navigation specific location, or documenting laboratory result patient. However, where such as ``Ukachukwu`` (Igbo), ``Lakicia`` (Swahili), ``Ingabire`` (Rwandan) are spoken, automatic speech recognition (ASR) models' performance degrades significantly, propagating errors systems. We model this problem...

The potential of realistic and useful synthetic data is significant. However, current evaluation methods for tabular generation predominantly focus on downstream task usefulness, often neglecting the importance statistical properties. This oversight becomes particularly prominent in low sample scenarios, accompanied by a swift deterioration these measures. In this paper, we address issue conducting an three state-of-the-art generators based their marginal distribution, column-pair...

Africa has a very low doctor-to-patient ratio. At busy clinics, doctors could see 30+ patients per day -- heavy patient burden compared with developed countries but productivity tools such as clinical automatic speech recognition (ASR) are lacking for these overworked clinicians. However, ASR is mature, even ubiquitous, in nations, and clinician-reported performance of commercial systems generally satisfactory. Furthermore, the recent general domain approaching human accuracy. several gaps...

This is the proceedings of 3rd ML4D workshop which was help in Vancouver, Canada on December 13, 2019 as part Neural Information Processing Systems conference.