Neural Machine Translation
Delivered December 3, 2019. Contributor: ChristineD
Goals
Determine information on Neural Machine Translation components. The information will be used for an article.
Early Findings
Neural Machine Translation
n (NMT)
Nodes in NMT
- Machine translation involves capturing text in one language and producing text in another language.
- When neural networks are employed for this process, this is referred to as neural machine translation.
-
The encoder-decoder configuration is a popular recurrent neural network (RNN) component.
- The setup is composed of an encoder network and a decoder network.
- The encoder portion processes the input text and the decoder part provides the translated text as its output.
- For massive translation requirements like the ones that the Google neural machine translation (GNMT) system is handling, it is necessary to distribute the complete workload into multiple processors and compute nodes.
- The system is also optimized by extending the bucketing logic for combining similar length sentences into several nodes to attain load
- There are several available platforms for data science programming.
- One example is the "Deep Learning Keras Integration."
- The platform can be used to handle neural machine translation processes.
- Due to its graphical interface setup, the computing units in the platform are "small colorful blocks called nodes."
- Arranging nodes in a pipeline in a specific sequence will produce a data processing application.
- The pipeline is referred to as the "workflow."
- The platform is made up of a software base and several extensions and integration from the developer community.
- These extensions and integration programs can improve the base software features and can produce complex AI algorithms such as deep learning related applications.
- As an example, the "Deep Learning" extension combines several functionalities from the Keras libraries.
- These libraries then consolidate the functionalities from TensorFlow in Python programming.
- Within the Keras integration, there are several nodes on-hand to develop specific network layers.
- There are also several nodes that are on-hand to "train networks in Keras, Tensorflow, and Python."
- The benefit of using the Keras integration include the massive lowering of the amount of code to generate.
- There are several Keras library functions that have been integrated into the nodes.
- Most of these functional nodes provide a graphic dialog window.
- Some nodes also lets the combination of more Keras/TensorFlow libraries through Python code.
- There are other nodes that are available in the Keras platform.
- A massive amount of nodes execute neural layers such as "input and dropout layers in Core, LSTM layers in Recurrent, and Embedding layers in the Embedding sub-category."
- Furthermore, the "Learner, Reader and Writer nodes" can be used to "respectively train, retrieve and store a network."
- The main nodes are the "DL Python Network Executor and DL Python Network Editor."
- These two nodes correspondingly enable "custom execution and custom editing" of a "Python compatible Deep Learning network via Python script, including Jupyter notebook."
- These two nodes "effectively bridge Keras nodes with all other not yet integrated Keras/TensorFlow library functions."
Cost of Computing Nodes
Argon Phase 3
Non-GPU Compute Nodes - 80 Compute Slots, Intel Xeon Gold 6230@2.1GHz (Turbo up to 3.9GHz), DDR4-2933, 1TB SSD, 100Gbps InfiniBand EDR, 10Gb Ethernet
GPU Capable Compute Nodes (Up to 8 GPUs per Node) - 80 Compute Slots, Intel Xeon Gold 6230@2.1GHz (Turbo up to 3.9GHz), DDR4-2933, 1TB SSD, 100Gbps InfiniBand EDR, 10Gb Ethernet
- Low Memory - 96GB - $7,423
- Standard Memory - 192GB - $7,768
- Mid-Memory - 384GB - $8,553
- High Memory - 768GB - $10,583
- Very High Memory - 1.5TB - $14,298
- GPU Cards - The configurations above do not include GPU cards.
GPU Cards
Proposed next steps:
You need to be the project owner to select a next step.
Our initial research provided several available insights and details on neural machine translation systems. Given that there are available resources on this, we propose continuing the research to provide additional insights and data points on the following topics:
(1) Provide 5-7 high-level insights that describe a computing node as it relates to Neural Machine Translation processes. The process should revolve around the tactic of adding a computing node to a server to increase throughput.
(2) Provide additional data on how much a computing node typically costs. Provide also the cost of a GPU.
(3) Provide 5-7 insights that explain how the Graphic Processing Unit (GPU) works, in relation to NMT if possible. We will include an explanation of how GPU inclusion affects translation speed as adding a GPU to a server is regarded as a common acceleration tactic.
We also recommend proceeding with additional research to provide 2-3 key players in the neural machine translation space. The major ones can be defined by their revenue sizes. For each player, we will provide the name, solutions being offered, why it is a key player, and the value proposition.