AAAI-MAKE Trip Report

March 21–23, 2022 @ Stanford University, Palo Alto, California, USA

Sponsored by the Association for the Advancement of Artificial Intelligence

AAAI-MAKE

Prelude

After 2.5 years of limited social interaction I participated and presented a paper in the AAAI-MAKE Spring Symposium series. It was hosted in the History Building of Stanford University. For those who have never been to Stanford before (myself included), it is characterized by an impressive, inspiring and beautiful campus. In terms of facilities, space and things to do it comes second to no other university/campus I’ve ever been. A close competitor would be the University of British Columbia (UBC) in Vancouver Canada. But let’s get to the interesting stuff of this report…

If more people show interest in the process of visa (ESTA), booking flights/hotels etc for a remote conference, I can do another right-up on how to go about it.

Now to the fun part!!

General info

The event was characterized by a variety of research topics with the general theme of being hybrid. Hybrid as in: reflecting on how can we combine ML approaches with KE (Knowledge Engineering) in order to build hybrid intelligence systems. It’s worth noting that the event itself was hybrid (as in participants who couldn’t make it in person, had the ability to join and present via Zoom). Which imho is nice and inclusive.

There was in total 9 Symposia with a variety of topics. I list them here with links to their respective websites and I would recommend folks to check them out. Some of them were so interesting I wish I could be in 2 places at the same time.

The full list:

I joined + followed the - Machine Learning and Knowledge Engineering for Hybrid Intelligence. So most of this report will be about that.

Day 1

The first day started with a keynote by Natasha Noy, who is leading the Dataset Search team at Google. For those who are unaware of Google Dataset Search, I would urge you to check it out. It’s basically a search engine for well…. datasets. Datasets from governments, from the industry or for research and anything that can be crawled out of the web. (@Xu)

One of the main points across was that, despite the absolute usefulness of declarative semantics, it appears that human-generated knowledge suffers from low quality. Which overall kinda contradicts the hybrid paradigm where ML pipelines are the usual suspect for low-quality while the human input is glorified as high quality.

The example Natasha gave to support her claim was from an experiment they performed where they took 600M webpages that had the schema.org tag for “Dataset”. For those who haven’t heard of schema.org, it’s an effort by Google, Microsoft, Yahoo and Yandex to attach semantics (in the form of mark-up) to webpages with the use of vocabularies. The incentive for websites? If you do it right, Google will pick it up and index it properly. For more info schema.org faq

Now back to the experiment….
Out of 600M pages that had the markup “Dataset”, a crazy 84% of them were false positives. Can the reader guess why?… Well, it’s cause humans also apparently understand what a Dataset is (semantically) differently.

Some related figures from her presentation:

Training and testing

Definition of Datasets

natasha

What does that experiment tell us?

  • I guess it’s not easy to crawl the web for datasets… when we cannot agree on what a Dataset is.
  • Schema.org labels are useful, but not when used in the wrong way by humans.

Session 1 in bullet points

..some pictures

LTNs

LTNs

LTNs

LTNs

Post session 1 thoughts:
There seems to be a common understanding that Hybrid Intelligence approaches are not only applicable but also preferable for multiple problems, especially problems characterized by lack of complete information. From Neural-Symbolic approaches for ontological subclass reasoning to Farming and Augmented Reality.

Session 2 in bullet points

General vibe: “Human in the loop”

Session 3 in bullet points

  • SFMGNet: A Physics-based Neural Network To Predict Pedestrian Trajectories
    • The only other paper that used Physics inspired knowledge to enhance a prediction task.
    • Useful for our colleagues at the CIVIC AI Lab (maybe?)
  • Adjoined Networks: A training paradigm with applications to network compression (@Michael, @Daniel - This has obvious applications to CNNs… I wonder what would happen if we tried to apply it to GNNs instead)
    • Interesting idea to compress deep neural networks for the purpose of utilising them in production and edge devices.
    • Based on the idea of teacher and student NNs.
    • They introduce an Adjoined Network that trains a smaller compressed network alongside the original network.
    • They show that they can simultaneously compress (with the student network) and regularize (the teacher network) any architecture.

LTNs

Day 1 after thoughts:

  • The other symposia were also extremely interesting for AI researchers, they touched upon philosophy - the example of morality vs ethics - but also human-robot collaboration in teams and trust.
  • Creating design patterns for ML was a well perceived idea by the community, as in the past the community has put together best-practices and design patters for software or knowledge engineering.
  • Lots of discussion around how do we go neuro-symbolic with Logic Tensor Networks making an appearance on Day 1 as one of the ways. (@Emile)

ml design patterns

Day 2

Day 2 started strong.

The first session of the day had everything… From combining contrastive visual and language embeddings for visual relationship detection (@Tae), to code-mixed language (Code-mixed language is a form of language wherein syntactic elements from one language are inserted into another language in such a way that the semantics of the resultant language remains the same)

Some highlights:

** Post Day 2 **
We had the plenary session where all the participants of every symposium joined. It was really cool and interesting to see the implicit connections between topics as well as the diversity of researchers + disciplines involved.

Day 3

Short paper session: characterized by early-works but quite interesting as ideas.

Disclaimer
As I was the last presenter - the real headliner of the thing :P , I couldn’t pay a lot of attention to the last days papers. Also, not a lot of notes from that day.

Last but not least, yours trully representing the VU :D

Dimitrios presenting

Random Notes

  • Post event the organisers held a pretty horizontal meeting where they asked the participants how to make the event better. I found this brilliant, since it really helps in aligning what the researchers want with what the event can be about.
    • For future iterations of MAKE: including the design patterns / best practices to combine KR/E and ML as part of the submission. -> find patterns of hybrid systems (bottom-up boxology)
  • Interesting 3rd day comment was that combining KE with ML actually sped up the process of reaching a good POC, ready to be evaluated.” ( personally I would think the overhead of combining them correctly would slow things down but well…. That’s why it was interesting.
  • I seriously liked California, I think I will pursue a position there for an internship for like 6 months - 1 year. A bit pricey but oh well…
    • The food was fantastic (american portions + asian cuisine <3 )
  • A shoutout to Takashi Kido - he gave by far the funniest talk during the plenary session :D :D
  • It was the first time I met Greek colleagues from the NTUA in the US - if I was a classifier my job would be easy ( 3/4 had long hair + beard lol )

Also some random pictures of California:

Dimitrios Golden State tee

Sel Dimitris Stanford

The end.