We investigate the problem of simultaneous machine translation of long-form speech content. We target a continuous speech-to-text scenario, generating translated captions for a live audio feed, such as a lecture or play-by-play commentary. As this scenario allows for revisions to our incremental translations, we adopt a re-translation approach to simultaneous translation, where the source is repeatedly translated from scratch as it grows. This approach naturally exhibits very low latency and high final quality, but at the cost of incremental instability as the output is continuously refined. We experiment with a pipeline of industry-grade speech recognition and translation tools, augmented with simple inference heuristics to improve stability. We use TED Talks as a source of multilingual test data, developing our techniques on English-to-German spoken language translation. Our minimalist approach to simultaneous translation allows us to easily scale our final evaluation to six more target languages, dramatically improving incremental stability for all of them.

It's commonly known that single sign-on (SSO) reduces risk associated with poor password practices and improves security posture.

Researchers are developing'digital twins' that combine computational models and machine learning to predict vehicle health and enable …

Apple uses machine learning more and more, and not just for Siri.

Artificial intelligence has had a profound impact on finance. In the span of a few decades, it has made finance faster, more accessible, more profitable, and more efficient in many ways. Despite all the significant benefits made possible by financial artificial intelligence, it also presents serious risks and implications for law, business, and society. My recent article, 'Artificial Intelligence, Finance, and the Law', published in the Fordham Law Review, offers a study of those risks and implications. It provides a broad examination of the inherent risks and larger implications of financial artificial intelligence.

The major one, I feel, is an easy way to train multiple models in parallel. I can run one model on one GPU, one model on multiple GPUs, and multiple models on multiple GPUs. However, it's currently very inefficient to run multiple small models on a single GPU. Even though a model being used for an embedded application might be well under a meg, the CUDA instance needed to run it is almost a gig in size, so a GPU with 6 gigs of spare VRAM can only run 6 models in parallel, when it should be able to run a much greater number with greater efficiency. Basically, anything in general that will let you run multiple models on the same GPU without large computational bottlenecks would be great.

The "Artificial Intelligence in IoT (AIoT) Convergence: Technologies, Platforms, Applications and AIoT Services in Industry Verticals 2019 - 2024" report has been added to ResearchAndMarkets.com's offering. This research provides a multi-dimensional view into the AI market including analysis of embedded devices and components, embedded software, and AI platforms. This research also assesses the combined Artificial Intelligence (AI) marketplace including embedded IoT and non-IoT devices, embedded components (including AI chipsets), embedded software and AI platforms, and related services. This research evaluates leading solution providers including hardware, software, integrated platforms, and services. It includes quantitative analysis with forecasts covering AI technology and systems by type, use case, application, and industry vertical. The forecast also covers each major market sector including consumer, enterprise, industrial, and government.

Partial Dependence Plots (PDPs) are a standard model inspection technique. It turns out that for decision trees, they can be computed very efficiently. This post explains how PDPs are computed in general, and goes into the details of the optimized version for tree models.