---
title: "ICLR 2020 Notes"
date: 2020-05-05
---

ICLR is one of the most important conferences in machine learning, and
as such, I was very excited to have the opportunity to volunteer and
attend the first fully-virtual edition of the event. The whole content
of the conference has been made [[https://iclr.cc/virtual_2020/index.html][publicly available]], only a few days
after the end of the event!

I would like to thank the [[https://iclr.cc/Conferences/2020/Committees][organizing committee]] for this incredible
event, and the possibility to volunteer to help other
participants[fn:volunteer].

The many volunteers, the online-only nature of the event, and the low
registration fees also allowed for what felt like a very diverse,
inclusive event. Many graduate students and researchers from industry
(like me), who do not generally have the time or the resources to
travel to conferences like this, were able to attend, and make the
exchanges richer.

In this post, I will try to give my impressions on the event, and
share the most interesting events and papers I saw.

[fn:volunteer] To better organize the event, and help people navigate
the various online tools, they brought in 500(!) volunteers, waved our
registration fees, and asked us to do simple load-testing and tech
support. This was a very generous offer, and felt very rewarding for
us, as we could attend the conference, and give back to the
organization a little bit.

* The Format of the Virtual Conference

As a result of global travel restrictions, the conference was made
fully-virtual. It was supposed to take place in Addis Ababa, Ethiopia,
which is great for people who are often the target of restrictive visa
policies in Northern American countries.

The thing I appreciated most about the conference format was its
emphasis on /asynchronous/ communication. Given how little time they
had to plan the conference, they could have made all poster
presentations via video-conference and call it a day. Instead, each
poster had to record a 5-minute video summarising their
research. Alongside each presentation, there was a dedicated
Rocket.Chat channel[fn:rocketchat] where anyone could ask a question
to the authors, or just show their appreciation for the work. This was
a fantastic idea as it allowed any participant to interact with papers
and authors at any time they please, which is especially important in
a setting where people were spread all over the globe.

There were also Zoom session where authors were available for direct,
face-to-face discussions, allowing for more traditional
conversations. But asking questions on the channel had also the
advantage of keeping a track of all questions that were asked by other
people. As such, I quickly acquired the habit of watching the video,
looking at the chat to see the previous discussions (even if they
happened in the middle of the night in my timezone!), and then
skimming the paper or asking questions myself.

All of these excellent ideas were implemented by an [[https://iclr.cc/virtual_2020/papers.html?filter=keywords][amazing website]],
collecting all papers in a searchable, easy-to-use interface, and even
a nice [[https://iclr.cc/virtual_2020/paper_vis.html][visualisation]] of papers as a point cloud!

[fn:rocketchat] [[https://rocket.chat/][Rocket.Chat]] seems to be an [[https://github.com/RocketChat/Rocket.Chat][open-source]] alternative to
Slack. Overall, the experience was great, and I appreciate the efforts
of the organizers to use open source software instead of proprietary
applications. I hope other conferences will do the same, and perhaps
even avoid Zoom, because of recent privacy concerns (maybe try
[[https://jitsi.org/][Jitsi]]?).

* Speakers

Overall, there were 8 speakers (two for each day of the main
conference). They made a 40-minute presentation, and then there was a
Q&A both via the chat and via Zoom. I only saw 4 of them, but I expect
I will be watching the others in the near future.

** Prof. Leslie Kaelbling, [[https://iclr.cc/virtual_2020/speaker_2.html][Doing for Our Robots What Nature Did For Us]]

This talk was fascinating. It is about robotics, and especially how to
design the "software" of our robots. We want to program a robot in a
way that it could work the best it can over all possible domains it
can encounter.



* Workshops

* Some Interesting Papers

** Natural Language Processing

*** Harwath et al., [[https://openreview.net/forum?id=B1elCp4KwH][Learning Hierarchical Discrete Linguistic Units from Visually-Grounded Speech]]

Humans can easily deconstruct all the information available in speech
(meaning, language, emotion, speaker, etc.). However, this is very
hard for machines. This paper explores the capacity of algorithms to
reason about all the aspects of the signal, including visual cues.

Their goal is to use spoken captions of images to train a predictive
model.

- speech signal: contains a lot of information (meaning, language,
  emotion, speaker, environment)
  - difficult to identify these different parts for an algo (although
    easy for a human)
- dominated by supervised ML
  - automated speech recognition (ASR) = P(text | waveform)
  - text-to-speech (TTS) = P(waveform | text, speaker)
  - high sample complexity
  - bad out-of-domain performance
  - limited by annotation capability
- human-like learning
  - ability to jointly reason about all the aspects of the signal
  - rapidly adapt to new speaker or noise conditions
  - learn new words for a single example
  - utilize unlabelled multimodal data
- using visual grounding for self-supervision
  - humans can leverage cross-modal correspondences to learn what
    spoken words represent without requiring any text or symbolic
    input whatsoever
  - hypothesis: similar for computer algorithms?
- goal: use spoken captions of images to train a predictive model
  - learn a hierarchical structure of units
  - learn the corresponding text, but also the transcription of the
    spoken sounds, at a sub-word level
- prefer models that learn a discrete tokenisation of the speech
  - language has an intrinsically symbolic structure
    - convey meaning with discrete words
    - words are in turn composed of a finite set of speech sounds
      (phones)
  - model that can discover discrete representations for word and
    phone-like units
    - more interpretable
    - able to do few-shot learning (learn a new word-like unit in
      terms of known phone-like units)
    - path towards learning compositional structure from continuous
      signals
- model for audio-visual grounding:
  - NN for image
  - NN for raw speech
  - shared embedding space
  - semantic supervision
- preliminary studies
  - lower layers features are correlated with phones
  - higher lauer features are correlated with words
- add a vector quantizing layer in the speech NN
- hierarchy of quantization layers
  - capture phones and words
- do an ABX test to compare performance to speech-only models
- conclusion
  - novel linguistic unit learning paradigm using multimodal data
    without text
  - SOTA performance on learning phonetic and word-level units
  - discovery of discreteness as a good inductive bias for semantic
    task from speech

** Reinforcement Learning

** ML and Neural Network Theory