One architecture to complete all tasks - transformer architecture is unifying the AI Jianghu on its own

Catalog

An architecture to accomplish all tasks —Transformer The architecture is being unified on its own AI Rivers and lakes

---

An architecture to accomplish all tasks —Transformer The architecture is being unified on its own AI Rivers and lakes

Language model , Images 、 Video has been Transformer The architecture refreshes the model scale and performance benchmark at the same time . I still want to talk about Transformer All kinds of variants of have been brilliant in this year , At the same time NLP and CV The field frequently brushes the list .

In recent years, ,transformer Architecture gradually extends its influence to various new fields . first ,Transformers It is developed for natural language processing , Now it is becoming a Swiss Army knife for in-depth learning . 2021 year , They are used to find drugs 、 Recognize voice, painting and other tasks .

transformers Has proven to be good at visual tasks 、 Predicting earthquakes and classifying and generating proteins . In the past year , Researchers have pushed them into broad new fields .

TransGAN：TransGAN It's a generative confrontation network , It is a combination of transformer To ensure that each generated pixel is consistent with its previously generated pixel . This work has achieved the most advanced results in measuring the similarity between the generated image and the training data .

TimeSformer：Facebook Of TimeSformer This architecture is used to identify actions in video clips . It explains the sequence of video frames , Instead of the usual sequence of words in the text . Its performance is better than convolutional neural network , You can analyze longer clips in a shorter time , And use less power .

GPT-2：Facebook、Google And researchers at the University of California, Berkeley trained on the text GPT-2, Then it freezes its self attention and feedforward layer . They can fine tune in a variety of areas , Including mathematics 、 Logic problems and computer vision .

AlphaFold 2：DeepMind Released AlphaFold 2 Open source version of , It USES transformer Find the protein according to the amino acid sequence 3D shape . The model has aroused the interest of the medical community , Because it has the potential to promote drug discovery and reveal biological insights .

Vision Transformer(ViT) as well as Video ViT：

Transformer On 2017 Made its debut in , And quickly changed the language modeling . Its self attention mechanism tracks the relationship between each element in the sequence and each other element , Not only suitable for analyzing word sequences , It is also suitable for analyzing pixels 、 Video frame 、 Amino acids, 、 Seismic wave sequence . be based on transformer The large language model of has become an example of the emerging basic model variety —— A model of pre training on a large unlabeled corpus , Special tasks can be fine tuned for a limited number of markup examples .transformer The fact that they can work well in various fields , It may indicate the basis beyond language transformer The basic model of .

    The history of deep learning has witnessed some rapidly popular ideas ：ReLU Activation function 、Adam Optimizer 、 Attention mechanism and current transformer. Developments over the past year have shown that , This architecture is still working .

Reference article ：https：//read.deeplearning.ai/the-batch/issue-123/