Скачать или смотреть AI 360: 22/02/2021. Superhuman ATARI performance, the world's largest Turing test and more!

For the full experience, and links to everything referenced, visit our website: https://lamaai.io

Superhuman performance on ATARI
GoogleAI introduce DreamerV2, a world model-based reinforcement learning algorithm which now currently holds the state of the art on Atari. DreamerV2 demonstrates a substantial performance improvement over existing model-based algorithms, and has been shown to outperform existing model-free algorithms when given the same compute resources. At a very high level, the work, which was done in collaboration with DeepMind and the University of Toronto behaves within the latent space of a discrete world model trained from pixels. Building upon the Recurrent State-Space Model (introduced in PlaNet and used in DreamerV1), the authors replace Gaussian variables with categorical variables, and also introduce a technique known as KL balancing. KL balancing enables posteriors to more accurately perform predictions during training, when priors (and their related networks) are still sub-optimial during training.

3 updates from HuggingFace!
HuggingFace have provided us with 3 updates this week. The first of which is a new zero-shot text classification tool. Given your custom set of labels, and an unlabeled collection of text documents, you can feed both into the model to obtain classifications for each of your text documents. Secondly, HuggingFace now supports MBart and MBart-50. MBart is a multi-lingual translation tool, where the base variant enables translating between English, and one of another 49 languages. The more recent MBart-50 enables enables translating between any pair of the 50 supported languages. Find more information here Datasets 1.3.0 has also been released! This brings the total datasets that Datasets contains to over 600, over 400 different languages! In this release, 26 new datasets are supported. A couple of new features in the update also include: on the fly data transformations, datasets versioning and S3 support for downloading and uploading datasets

3 updates from PapersWithCode!
PapersWithCode also gifted us with 3 updates. The first of which is a huge milestone they've hit! Their machine learning (ML) Methods now contains over 1000 ML components. A method is essentially a building block/component of an artificial intelligence (AI) system - such as an attention layer or optimization function. PapersWithCode also announce PapersWithLibraries - a index of ML libraries made by the AI community. Currently, PapersWithLibraries indexes 3 libraries: TIMM, a state of the art image classification library containing over 300+ pre-trained models; TorchVision, a popular image processing library which contains many image related datasets, 60+ pre-trained models, and pipelines for image transformations; and Detectron2, a state of the art object detection library which contains over 80 pre-trained models.

BIG-Bench
Google announce BIG-Bench, a collobartive benchmark to quantitively evaluate large language models (LMs). Recently, LMs such as OpenAI's GPT-3 have gotten so good that existing benchmarks are no longer suitable for measuring their performance. To this end, BIG-Bench is attempting to collect tasks that could be used to evaluate these extremely large LMs. All submitters of accepted tasks will be featured as co-authors on the paper introducing the benchmark. The paper and benchmark will be released during the Workshop on Enormous Language Models: Perspectives and Benchmarks at ICLR 2021. Find out more information about BIG-Bench here.

The world's largest Turing Test
AiCore announce an experiment: the world's largest Turing test. Intended to test whether OpenAI's GPT-3 can fool a human into thinking that their talking to a human, the experiement will take place on the 25th of March. Sign-ups are open to the general public, so if you're interested in participating in what could be a huge technological milestone, or you're just curious to find out more information about the experiement, check out their webpage