Selected cutting-edge technical articles of Bi Ren Academy of science and technology

 

2022 It's more than half , In the past six months , Multimodal models are booming ,Text-to-Image The model shows amazing effects ,NeRF Relevant research is still active , More computing power 、 More advanced architecture 、 More flexible software and hardware collaborative design 、 And more efficient parallel algorithms have also made giant models break records .

The following is a selection of articles in some key fields by the Bi Ren Institute of science and Technology , I hope I can learn and make progress with you .

 

Frontier algorithm

DALL·E 2 and Imagen And other excellent texts - Image generation model “ A literary creation ” The image effect is amazing , The core of this capability is the diffusion probability model . When we first came into contact with the diffusion probability model last year, we thought it was a promising research direction , And made a detailed analysis of this .

• New ideas of deep generation network ： Diffusion probability model

The emerging neural rendering technology brings new opportunities to computer graphics , It has been widely concerned by academia and industry . In recent years, computer vision and graphics summit , We can all see a variety of refreshing neural rendering applications . Neural rendering is a new graphics rendering method based on deep learning , It is to realize the current hot concept “ Meta universe ” The key elements of . At the same time, neural rendering is also the frontier direction of deep learning , It has spawned many new deep learning methods and computing structures .

 

• Recent advances and algorithms in neural rendering （ One ）： Semantic image synthesis

• Recent advances and algorithms in neural rendering （ Two ）：NeRF And its evolution

• Feature depth calculation feature in neural rendering

• Visual nerve field ：NeRF A new perspective of research

• NeRF Solution of camera parameters in method

• Kubric： Efficiently synthesize visual data sets

Recommendation systems and voice models have been widely used , We analyze the implementation of these models in detail from the algorithm and architecture level .

• “Hello, world!”, It's not so easy to say

• “Hello, world!”, It's not so easy to say （ To continue ）

• I know you better than you do — The deep combination of neural network and recommendation system

• Everything can be embedded --embedding stay GPU In the implementation of

• Recommended scene training acceleration ： Large scale hybrid distributed system

• Near memory processor design for recommendation system

 

Meta AI Open source OPT-175B, The distributed training framework is becoming more and more perfect ,GPU Computing power is increasing , The scale of the model is constantly refreshed . To better support “ Giant ” model training , On the one hand, we focus on how to provide more computing power , On the other hand, think about how to make more efficient use of the existing system .

• Skyreach pillar ： An efficient two-dimensional method for training super large models

• Faster than fast ： Implementation of hybrid accuracy calculation acceleration

• Crowdsourcing training ： Another distributed heterogeneous depth model training method

• Meet the communication challenges in large-scale training ： Synchronization and topology

• Meet the communication challenges in large-scale training ： Compression and scheduling

 

With GPU The improvement of ability , The software and hardware ecosystem supporting deep learning has developed rapidly . It has become a trend to solve scientific computing problems through deep learning , Among them, the method of using deep learning to solve differential equations is also gradually emerging . Of particular interest is a method called physically guided Neural Networks , It has injected new vitality into the field of scientific computing .

 

• The acceleration problem of low numerical accuracy in the field of Scientific Computing

• Graphical neural network solution of partial differential equations

• Analysis of physical guidance neural network method

• Fluid mechanics and physics guide Neural Networks

• Get dynamic information from data ： Dynamic mode decomposition (DMD) With the transcendental combination of Physics

We We will also continue to pay attention to the core components of cutting-edge technology , Arrangement Novel methods and ideas .

• Must back propagation be used in training neural networks ？

• 「 Thermodynamic calculation 」： from Landauer Boundary to the ultimate computer

• Transformer Hegemony , Is dot product self attention irreplaceable ？

• be based on Object Query New ideas of machine vision : DETR And development

• GAN Applied research in causal reasoning
• The latest progress of causal discovery and its practice in complex systems

• Roll plus “ volume ”： Implementation of special convolution

• High order optimizer ： A sharp weapon for deep learning acceleration

 

Advanced architecture

Whether it is the algorithm we are already familiar with , Or a new algorithm , How to use the tools in hand to design software and hardware architecture , Execute these algorithms efficiently , It's something we keep thinking about , It can be said to be the daily work of architects . Of course , The answer to this question may also be an important decision , Big innovation , Even the technical route changes . In Bi Ren Research Institute , We will start from the design space “ Expand ” and “ Compress ” These two different directions to think about this problem .

 

• How Google builds the world's fastest AI Supercomputer system ？（ On ）

• How Google builds the world's fastest AI Supercomputer system ？（ Next ）

• System optimization for large-scale graph computation （1）

• System optimization for large-scale graph computation （2）

• System optimization for large-scale graph computation （3）

• System on chip design case analysis - Xbox host

• Figure in depth investigation of neural network accelerator （ On ）

• Figure in depth investigation of neural network accelerator （ in ）

• Figure in depth investigation of neural network accelerator （ Next ）

• Application of machine learning method in digital chip design

• Hardware acceleration of sparse tensor operator

 

Collaborative design

since AI Chip has become a hot research topic , Many about AI Academic articles on chip architecture exploration continue to emerge , Everyone has different views on AI Chip architecture analysis and performance optimization .MLIR It is the open source compiler framework launched by Google team , Quite remarkable , The flexible compiler architecture enhances its potential in many fields . By customizing IR The connection between the two , You can explore and MLIR Build a bridge between , During compilation , Automatically explore hardware architecture and optimize software compilation , Even generate hardware code , Realize software and hardware collaborative design .

• MLIR Discussion on the implementation of homomorphic encryption by multi-layer compilation framework

• be based on IREE HLO Project view MLIR Compiler implementation process and advantages

• MLIR Compiler technology meets hardware design challenges ？—— On CIRCT project

• be based on MLIR Realization GEMM Compile optimization

• Application of polyhedron compilation technology in software and hardware collaborative design

• Challenges and implementation analysis of dynamic model compilation optimization

• MLIR Thoughts on software and hardware collaborative design under compilation framework

• Borrow from others ：Pluto help MLIR Compiler polyhedron optimization