NLP Architect by Intel AI Lab | Hacker News Reader

NLP Architect by Intel AI Lab

nlp_architect.nervanasys.com

77 pointstsaprailis11 comments

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.

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.

So does this not work if you don’t have fancy new intel hardware?

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 :)

Thanks, I'll update my post to clarify, and also add the Tensorflow page. Great work by you and your team!

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.

By the way, the last author of the TCN paper (Vladlen Koltun) works at Intel Labs (Intelligent Systems Lab).

How does this compare to word2vec or fasttext?

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.

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?

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.

Yet another interface on top of Pytorch/TF/Gensim.

Reply on news.ycombinator.com