AI Conference Summary

Recently I attended the O'Reilly AI Conference in London. Since I am not a good multi-tasker and do never take notes, here is my quick wrap-up after-the fact from what I memorize. Any corrections are appreciated. I will only cover talks I did attend myself.

Talks

Applied machine learning at Facebook: An infrastructure perspective

Yangqing Jia of Facebook spoke about the machine learning infrastructure at Facebook. A very well presented talk that touched organization, software and hardware at for ML at Facebook. They use a software infrastructure called FBLearner consisting of several components:

• FBLearner Feature Store as a "marketplace for features". Where for it is unclear to me whether it stores the actual features or the rules to generate them. In the paper they write "The Feature Store is essentially a catalog of several feature generators that can be used both for training and real-time prediction, and it serves as a marketplace that multiple teams can use to share and discover features" - so it seems that those feature vectors are not materialized. Both has advantages and disadvantages. From our experience rule-based feature generation on-the-fly can be really expensive for re-training. Materializing features in a generic way seems also difficult for me.
• FBLearner Flow for training. You can define a workflow consisting of operators, FBL Flow taking care of dependencies. Flow also has tooling for experiment management, something I stumble upon quite often recently.
• FBLearner Predictor for inference.

Facebook has a large number of different ML workloads: Newsfeed, Sigma, Ads, Search, Lumos (Image analysis), Facer and Translation. All those workloads have different characteristics with regard to frequency of model updates or training times. The FBLearner infrastructure takes care of those issues.

You can find all the details in the paper the talk was based on. (The software is Facebook-internal)

Federated learning

This talk by Cloudera Fast Forward Labs, Ryan Micallef gave an introduction to federated learning. The idea is to keep the data at the device (Smartphone, Manufacturing Site), train models locally and then just transfer the model parameters. FF Labs implemented a prototype based on the public C-MAPSS dataset. The algorithm itself presents itself quite simple:

It averages the parameters of the distributed models and pushes them back to the devices. (For me this seems very similar to what a parameter server does during distributed training using TensorFlow, comments?)

I think that federated learning holds a lot of promise with regard to privacy. However I didn't see many real world implementations outside the Apples and Googles yet.

Tensor2Tensor (sponsored by Google)

Tensor2Tensor, or T2T for short, is a library of deep learning models and datasets designed to make deep learning more accessible and accelerate ML research. T2T is actively used and maintained by researchers and engineers within the Google Brain team and a community of users" (from the website). This talk was a fast walk-through covering main use cases and API of the library. They showed the different integrated models and datasets with short code-examples and concluded for each task with the statement that T2T achieves state-of-the-art performance or at least close to it. Anyhow, we will evaluate T2T shortly for automatic summarization. Unfortunately, as I learned, the newly (or soon to be) released BERT models will not integrate with T2T.

Frontiers of TensorFlow: Mathematics and music (sponsored by Google)

This were two quite different talks in one session. The first was about TensorFlow Probablility which is a library for probabilistic reasoning and statistical analysis in TensorFlow. The talk introduced code examples for different modeling tasks some of which you can probably find at the blog post by the speaker, Josh Dillon: https://medium.com/tensorflow/introducing-tensorflow-probability-dca4c304e245

The second part was about multiple art/music-related projects at Google Brain, especially Project Magenta. For example hierarchical variational autoencoders for music synthesis as shown in this video. Great talk, however more an overview than in-depth.

How CLEVER is your neural network? Robustness evaluation against adversarial examples

I did like this one a lot. It first gave a nice overview and examples of adverserial attacks and why it is difficult to harden models against those. Then he introduced ways to evaluate robustness and explained their new robustness score. Right now, adverserial attacks and the defense against them are an arms-race which should currently hinder application of deep learning models in critical tasks. You can find a blog post about CLEVER and the paper on the web.

Slides.

Deprecating the state machine: Building conversational AI with the Rasa stack

This talk by rasa.ai Cofounder Alan Nichol gave an overview on how the Rasa stack can support you with creating "level 3 conversational apps". He presented quite a few use and business cases but did not go technical There are versions of this talk containing deeper stuff, e.g. here.

Personalizing the user experience and playlist consumption on Spotify

In this great talk by @Mounia Lalmas Director of Research @ Spotify she presented current work at Spotify to improve personalization by co-clustering playlists and users.

Slides on her homepage: https://mounia-lalmas.blog/2018/10/11/personalizing-the-user-experience-and-playlist-consumption-on-spotify/

Accelerating innovation through analogy mining

I was looking forward to this talk, since we are currently also working on patent mining. Dafna Shahaf presented their work, the corresponding publication was the best paper at KDD17. Since the paper was reviewed by the "Morning Paper", I am just linking the the review over there.

Performance evaluation of GANs in a semisupervised OCR use case

Talk by our Florian Wilhelm. You can find the slides here:

Natural language processing, understanding, and generation

This talk by Amy Heineike was mostly a walk-through to use-cases of their primer.ai product. I was a little disappointed since she didn't mention any algorithms, papers or code. I also would like to question her postulation that machine reading is "solved" and the interesting stuff is now in NLG. Don't get me wrong, their product looks pretty cool but probably I was the wrong target group for this talk. Anyway, they do have two interesting posts on summarization on their blog, https://primer.ai/blog/TLDR/ and https://primer.ai/blog/seq2seq/

Topic Tables

I found the topic tables at lunch a very nice ice-breaker to engage in conversation. Even though Florian took the picture below, we had a good discussion about AI ethics at the very same table on the second day of the conference.

Final remarks

I very much enjoyed this conference. While I would have enjoyed more deep technical talks about the topics I am involved in, the overall mid-level depth allowed me also to follow stuff I am less familiar with.

October seems to be crammed with AI Conferences. At the same time there was the World Summit AI in Amsterdam and the NVIDIA GTC in Munich. The week after, also in London, mcubed, the UK version of "minds mastering machines" (which was awesome) took place. I would love to hear how the other conferences were perceived.