Paper Review: Tensorflow, Large-Scale Automobile Learning On Heterogeneous Distributed Systems
The newspaper is available here.
TensorFlow is Google's novel framework for implementing machine learning algorithms using dataflow graphs. Nodes/vertices inwards the graph correspond operations (i.e., mathematical operations, machine learning functions), as well as the edges correspond the tensors, (i.e., multidimensional information arrays, vectors/matrices) communicated betwixt the nodes. Special edges, called command dependencies, tin dismiss also be inwards the graph to announce that the source node must complete executing before the finish node starts executing. Nodes are assigned to computational devices as well as execute asynchronously as well as inwards parallel in 1 trial all the tensors on their incoming edges becomes available.
It seems similar the dataflow model is getting a lot of attending of late as well as is emerging every bit a useful abstraction for large-scale distributed systems programming. I had reviewed Naiad dataflow framework earlier. Adopting the dataflow model provides flexiblity to TensorFlow, as well as every bit a result, TensorFlow framework tin dismiss last used to limited a broad multifariousness of algorithms, including preparation as well as inference algorithms for deep neural network models.
Wait, what? Why does TensorFlow demand to run on wimpy phones?!
The newspaper says the betoken is but portability: "Having a unmarried organisation that tin dismiss bridge such a broad arrive at of platforms significantly simplifies the real-world usage of machine learning system, every bit nosotros accept constitute that having form systems for large-scale preparation as well as small-scale deployment leads to pregnant maintenance burdens as well as leaky abstractions. TensorFlow computations are expressed every bit stateful dataflow graphs."
I sympathise this, yep portability is of import for development. But I don't purchase this every bit the explanation. Again, why does TensorFlow, such a powerhorse framework, demand to last shoehorned to run on a unmarried wimpy phone?!
I recall Google has designed as well as developed TensorFlow as a Maui-style integrated code-offloading framework for machine learning. What is Maui yous ask? (Damn, I don't accept a Maui summary inwards my blog?)
Maui is a organisation for offloading of smartphone code execution onto backend servers at method-granularity. The organisation relies on the mightiness of managed code surroundings (.NET CLR) to last run on dissimilar platforms. By introducing this automatic offloading framework, Maui enables applications that move past times memory/computation limits to run on smartphones inwards a battery- & bandwidth-efficient manner.
TensorFlow enables cloud backend back upwards for machine learning to the private/device-level machine learning going on inwards your smartphone. It doesn't brand feel for a power-hungry entire TensorFlow programme to run on your wimpy smartphone. Your smartphone volition last running solely certainly TensorFlow nodes and modules, the residuum of the TensorFlow graph volition last running on the Google cloud backend. Such a setup is also neat for preserving privacy of your telephone acre nonetheless enabling machine learned insights on your Android.
I volition utter nearly applications of this, but rootage allow me cite this other evolution nearly TensorFlow that supports my guess.
Why did Google opensource this projection relatively early on rather than keeping it proprietary for longer? This is their answer: "We believe that machine learning is a commutation element to the innovative products as well as technologies of the future. Research inwards this surface area is global as well as growing fast, but lacks criterion tools. By sharing what nosotros believe to last 1 of the best machine learning toolboxes inwards the world, nosotros promise to create an opened upwards criterion for exchanging enquiry ideas as well as putting machine learning inwards products."
This supports my guess. TensorFlow's emphasis on heterogeneity is non but for portability. Google is thinking of TensorFlow every bit an ecosystem. They desire developers to adopt TensorFlow, thence TensorFlow is used for developing machine learning modules inwards Android phones as well as tablets. And then, Google volition support/enrich (and honor ways to exercise goodness from) these modules past times providing backends that run TensorFlow. This is a dainty strategy for Google, a machine learning company, to percolate to the machine learning inwards the Internet of Things domain inwards general, as well as the mobile apps marketplace position inwards particular. Google tin dismiss last the monopoly of Deep learning As A Service (DAAS) provider leveraging the TensorFlow platform.
How tin dismiss Google exercise goodness from such integration? Take a await at this applications list: "TensorFlow has been used inwards Google for deploying many machine learning systems into production: including spoken language recognition, reckoner vision, robotics, information retrieval, natural linguistic communication processing, geographic information extraction, as well as computational drug discovery."
With a mobile-integrated TensorFlow machine-learning system, Google tin dismiss render amend personal assistant on your smartphone. Watch out Siri, amend spoken language recognition, calendar/activity integration, aspect upwards recognition, as well as reckoner vision is coming. Robotics applications tin dismiss enable Google to penetrate self-driving auto OS, as well as drone OS markets. And afterward that tin dismiss come upwards to a greater extent than transformative globe-spanning physical the world sensing & collaboration applications.
With this I residuum my case. (I got carried away, I but intended to exercise a newspaper review.)
In this talk, Jeff Dean gives a real dainty motivation as well as introduction for Tensorflow.
Here is an annotated summary of the TensorFlow paper.
This post service explains why TensorFlow framework is proficient tidings for deep learning.
TensorFlow is Google's novel framework for implementing machine learning algorithms using dataflow graphs. Nodes/vertices inwards the graph correspond operations (i.e., mathematical operations, machine learning functions), as well as the edges correspond the tensors, (i.e., multidimensional information arrays, vectors/matrices) communicated betwixt the nodes. Special edges, called command dependencies, tin dismiss also be inwards the graph to announce that the source node must complete executing before the finish node starts executing. Nodes are assigned to computational devices as well as execute asynchronously as well as inwards parallel in 1 trial all the tensors on their incoming edges becomes available.
It seems similar the dataflow model is getting a lot of attending of late as well as is emerging every bit a useful abstraction for large-scale distributed systems programming. I had reviewed Naiad dataflow framework earlier. Adopting the dataflow model provides flexiblity to TensorFlow, as well as every bit a result, TensorFlow framework tin dismiss last used to limited a broad multifariousness of algorithms, including preparation as well as inference algorithms for deep neural network models.
TensorFlow's heterogeneous device support
The newspaper makes a large bargain nearly TensorFlow's heterogenous device support, it is fifty-fifty correct at that spot inwards the newspaper title. The newspaper says: "A computation expressed using TensorFlow tin dismiss last executed amongst niggling or no modify on a broad multifariousness of heterogeneous systems, ranging from mobile devices such every bit phones as well as tablets upwards to large-scale distributed systems of hundreds of machines as well as thousands of computational devices such every bit GPU cards."Wait, what? Why does TensorFlow demand to run on wimpy phones?!
The newspaper says the betoken is but portability: "Having a unmarried organisation that tin dismiss bridge such a broad arrive at of platforms significantly simplifies the real-world usage of machine learning system, every bit nosotros accept constitute that having form systems for large-scale preparation as well as small-scale deployment leads to pregnant maintenance burdens as well as leaky abstractions. TensorFlow computations are expressed every bit stateful dataflow graphs."
I sympathise this, yep portability is of import for development. But I don't purchase this every bit the explanation. Again, why does TensorFlow, such a powerhorse framework, demand to last shoehorned to run on a unmarried wimpy phone?!
I recall Google has designed as well as developed TensorFlow as a Maui-style integrated code-offloading framework for machine learning. What is Maui yous ask? (Damn, I don't accept a Maui summary inwards my blog?)
Maui is a organisation for offloading of smartphone code execution onto backend servers at method-granularity. The organisation relies on the mightiness of managed code surroundings (.NET CLR) to last run on dissimilar platforms. By introducing this automatic offloading framework, Maui enables applications that move past times memory/computation limits to run on smartphones inwards a battery- & bandwidth-efficient manner.
TensorFlow enables cloud backend back upwards for machine learning to the private/device-level machine learning going on inwards your smartphone. It doesn't brand feel for a power-hungry entire TensorFlow programme to run on your wimpy smartphone. Your smartphone volition last running solely certainly TensorFlow nodes and modules, the residuum of the TensorFlow graph volition last running on the Google cloud backend. Such a setup is also neat for preserving privacy of your telephone acre nonetheless enabling machine learned insights on your Android.
I volition utter nearly applications of this, but rootage allow me cite this other evolution nearly TensorFlow that supports my guess.
Google Opensourced the TensorFlow API inwards Nov 2015
Google opensourced the TensorFlow API as well as a limited reference implementation (the implementation runs on a unmarried device only) nether the Apache 2.0 license inwards Nov 2015. This implementation is available at www.tensorflow.org, as well as has attracted a lot of attention.Why did Google opensource this projection relatively early on rather than keeping it proprietary for longer? This is their answer: "We believe that machine learning is a commutation element to the innovative products as well as technologies of the future. Research inwards this surface area is global as well as growing fast, but lacks criterion tools. By sharing what nosotros believe to last 1 of the best machine learning toolboxes inwards the world, nosotros promise to create an opened upwards criterion for exchanging enquiry ideas as well as putting machine learning inwards products."
This supports my guess. TensorFlow's emphasis on heterogeneity is non but for portability. Google is thinking of TensorFlow every bit an ecosystem. They desire developers to adopt TensorFlow, thence TensorFlow is used for developing machine learning modules inwards Android phones as well as tablets. And then, Google volition support/enrich (and honor ways to exercise goodness from) these modules past times providing backends that run TensorFlow. This is a dainty strategy for Google, a machine learning company, to percolate to the machine learning inwards the Internet of Things domain inwards general, as well as the mobile apps marketplace position inwards particular. Google tin dismiss last the monopoly of Deep learning As A Service (DAAS) provider leveraging the TensorFlow platform.
How tin dismiss Google exercise goodness from such integration? Take a await at this applications list: "TensorFlow has been used inwards Google for deploying many machine learning systems into production: including spoken language recognition, reckoner vision, robotics, information retrieval, natural linguistic communication processing, geographic information extraction, as well as computational drug discovery."
With a mobile-integrated TensorFlow machine-learning system, Google tin dismiss render amend personal assistant on your smartphone. Watch out Siri, amend spoken language recognition, calendar/activity integration, aspect upwards recognition, as well as reckoner vision is coming. Robotics applications tin dismiss enable Google to penetrate self-driving auto OS, as well as drone OS markets. And afterward that tin dismiss come upwards to a greater extent than transformative globe-spanning physical the world sensing & collaboration applications.
With this I residuum my case. (I got carried away, I but intended to exercise a newspaper review.)
Related links
In the side past times side decade nosotros volition come across advances inwards machine learning coupled amongst advances inwards Internet Of Things.In this talk, Jeff Dean gives a real dainty motivation as well as introduction for Tensorflow.
Here is an annotated summary of the TensorFlow paper.
This post service explains why TensorFlow framework is proficient tidings for deep learning.
0 Response to "Paper Review: Tensorflow, Large-Scale Automobile Learning On Heterogeneous Distributed Systems"
Post a Comment