Re:NLP Architect by Intel AI Lab
by jph on Thursday January 10, 2019 @12:57PM
(#18876154)
This AI toolkit works on popular Intel CPUs, and is a big step forward for the new Intel Nervana Neural Network Processor (NNP-I) hardware chip akin to a GPU.
One surprising research result for this NLP is that a simple convolutional architecture outperforms canonical recurrent networks, often. See: CMU lab, Sequence Modeling Benchmarks and Temporal Convolutional Networks (TCN) https://github.com/locuslab/TCN
If you're interested in Nervana, here are some specifics: the chip is for hardware neural network acceleration, for inference-based workloads. Notable features include fixed-point math, Ice Lake cores, 10-nanometer fabs, on-chip memory management by software directly, and hardware-optimized inter-chip parallelism.
I've worked for Intel, and I'm stoked to see the AI NLP progress.
Re:NLP Architect by Intel AI Lab
by peter1983 on Thursday January 10, 2019 @01:21PM
(#18876411)
Hi. I’m one of the authors of the library.
The models work on every popular CPU by Intel.
Nonetheless, we’re supporting Intel Optimzed Tensorflow when installing and in the future plan to add HW optimizations to the models.
Stay tuned :)
Re:NLP Architect by Intel AI Lab
by stochastic_monk on Thursday January 10, 2019 @01:42PM
(#18876605)
So this is aimed at inference rather than training. Does Intel have any plans to produce chips which can scale training as well, or is that largely going to be outsourced to GPUs for models of considerable size for the time being?
I imagine models need to be deployed more often than built, but I thought that the pain point was usually the latter.
Re:NLP Architect by Intel AI Lab
by yorwba on Thursday January 10, 2019 @05:24PM
(#18878909)
word2vec and fasttext are specialized tools for creating word embeddings, this is a more generalist library. It's more comparable to PyText, AllenNLP or Flair, the main difference appearing to be that the other three use PyTorch, not Tensorflow.
Re:NLP Architect by Intel AI Lab
by ___cs____ on Thursday January 10, 2019 @05:57PM
(#18879253)
with the recent change for TF 2.0. If you would design something similar, will you use TF or Pytorch?
What I am trying to ask here is, Is TF 2.0 is comparable to Pytorch when it comes to ease of use?
Re:NLP Architect by Intel AI Lab
by m0zg on Friday January 11, 2019 @12:52AM
(#18881308)
Probably not. It has "imperative" mode, but it also drags in quite a bit of API baggage that PyTorch just doesn't have. It's not that PyTorch is ideal, but its main advantage is it "feels like NumPy". Google also has a library that "feels like NumPy". In fact it kind of _is_ NumPy with hardware acceleration, but it seems to be in very early stages. I heard from insiders that there are only 2 people working on the project, if that. The name of the project is JAX: https://github.com/google/jax. It's arguably a lower level framework on top of which something like PyTorch could be built.
The Intel AI Lab has an introduction to NLP (https://ai.intel.com/deep-learning-foundations-to-enable-nat...) and optimized Tensorflow (https://ai.intel.com/tensorflow/)
One surprising research result for this NLP is that a simple convolutional architecture outperforms canonical recurrent networks, often. See: CMU lab, Sequence Modeling Benchmarks and Temporal Convolutional Networks (TCN) https://github.com/locuslab/TCN
If you're interested in Nervana, here are some specifics: the chip is for hardware neural network acceleration, for inference-based workloads. Notable features include fixed-point math, Ice Lake cores, 10-nanometer fabs, on-chip memory management by software directly, and hardware-optimized inter-chip parallelism.
I've worked for Intel, and I'm stoked to see the AI NLP progress.
I imagine models need to be deployed more often than built, but I thought that the pain point was usually the latter.