Tensors are a popular programming interface for developing artificial intelligence(AI)algorithms.Layout refers to the order of placing tensor data in the memory and will affect performance by affecting data locality;t...Tensors are a popular programming interface for developing artificial intelligence(AI)algorithms.Layout refers to the order of placing tensor data in the memory and will affect performance by affecting data locality;therefore the deep neural network library has a convention on the layout.Since AI applications can use arbitrary layouts,and existing AI systems do not provide programming abstractions to shield the layout conventions of libraries,operator developers need to write a lot of layout-related code,which reduces the efficiency of integrating new libraries or developing new operators.Furthermore,the developer assigns the layout conversion operation to the internal operator to deal with the uncertainty of the input layout,thus losing the opportunity for layout optimization.Based on the idea of polymorphism,we propose a layout-agnostic virtual tensor programming interface,namely the VTensor framework,which enables developers to write new operators without caring about the underlying physical layout of tensors.In addition,the VTensor framework performs global layout inference at runtime to transparently resolve the required layout of virtual tensors,and runtime layout-oriented optimizations to globally minimize the number of layout transformation operations.Experimental results demonstrate that with VTensor,developers can avoid writing layout-dependent code.Compared with TensorFlow,for the 16 operations used in 12 popular networks,VTensor can reduce the lines of code(LOC)of writing a new operation by 47.82%on average,and improve the overall performance by 18.65%on average.展开更多
This article explains how two AI systems have been incorporated into the everyday operations of two Singapore public healthcare nation‐wide screening programs.The first example is embedded within the setting of a nat...This article explains how two AI systems have been incorporated into the everyday operations of two Singapore public healthcare nation‐wide screening programs.The first example is embedded within the setting of a national level population health screening program for diabetes related eye diseases,targeting the rapidly increasing number of adults in the country with diabetes.In the second example,the AI assisted screening is done shortly after a person is admitted to one of the public hospitals to identify which inpatients—especially which elderly patients with complex conditions—have a high risk of being readmitted as an inpatient multiple times in the months following discharge.Ways in which healthcare needs and the clinical operations context influenced the approach to designing or deploying the AI systems are highlighted,illustrating the multiplicity of factors that shape the requirements for successful large‐scale deployments of AI systems that are deeply embedded within clinical workflows.In the first example,the choice was made to use the system in a semi‐automated(vs.fully automated)mode as this was assessed to be more cost‐effective,though still offering substantial productivity improvement.In the second example,machine learning algorithm design and model execution trade-offs were made that prioritized key aspects of patient engagement and inclusion over higher levels of predictive accuracy.The article concludes with several lessons learned related to deploying AI systems within healthcare settings,and also lists several other AI efforts already in deployment and in the pipeline for Singapore's public healthcare system.展开更多
基金supported by the National Key Research and Development Program of China under Grant No.2021zD0110101the National Natural Science Foundation of China under Grant Nos.62090024,61872043,and 61802368the Australian Research Council grant under Grant Nos.DP180104069 and DP210102409。
文摘Tensors are a popular programming interface for developing artificial intelligence(AI)algorithms.Layout refers to the order of placing tensor data in the memory and will affect performance by affecting data locality;therefore the deep neural network library has a convention on the layout.Since AI applications can use arbitrary layouts,and existing AI systems do not provide programming abstractions to shield the layout conventions of libraries,operator developers need to write a lot of layout-related code,which reduces the efficiency of integrating new libraries or developing new operators.Furthermore,the developer assigns the layout conversion operation to the internal operator to deal with the uncertainty of the input layout,thus losing the opportunity for layout optimization.Based on the idea of polymorphism,we propose a layout-agnostic virtual tensor programming interface,namely the VTensor framework,which enables developers to write new operators without caring about the underlying physical layout of tensors.In addition,the VTensor framework performs global layout inference at runtime to transparently resolve the required layout of virtual tensors,and runtime layout-oriented optimizations to globally minimize the number of layout transformation operations.Experimental results demonstrate that with VTensor,developers can avoid writing layout-dependent code.Compared with TensorFlow,for the 16 operations used in 12 popular networks,VTensor can reduce the lines of code(LOC)of writing a new operation by 47.82%on average,and improve the overall performance by 18.65%on average.
文摘This article explains how two AI systems have been incorporated into the everyday operations of two Singapore public healthcare nation‐wide screening programs.The first example is embedded within the setting of a national level population health screening program for diabetes related eye diseases,targeting the rapidly increasing number of adults in the country with diabetes.In the second example,the AI assisted screening is done shortly after a person is admitted to one of the public hospitals to identify which inpatients—especially which elderly patients with complex conditions—have a high risk of being readmitted as an inpatient multiple times in the months following discharge.Ways in which healthcare needs and the clinical operations context influenced the approach to designing or deploying the AI systems are highlighted,illustrating the multiplicity of factors that shape the requirements for successful large‐scale deployments of AI systems that are deeply embedded within clinical workflows.In the first example,the choice was made to use the system in a semi‐automated(vs.fully automated)mode as this was assessed to be more cost‐effective,though still offering substantial productivity improvement.In the second example,machine learning algorithm design and model execution trade-offs were made that prioritized key aspects of patient engagement and inclusion over higher levels of predictive accuracy.The article concludes with several lessons learned related to deploying AI systems within healthcare settings,and also lists several other AI efforts already in deployment and in the pipeline for Singapore's public healthcare system.
