Here is my (somehow tongue-in-the-cheek) comparison of the 3 venues:
SIGGRAPH – this paper presents the best invention since the fire, the wheel, the electricity, the internet, etc.
CHI – 2 pages of technical content + 8 pages of user study.
UIST – somewhere in between, no time for conducting/claiming too much user study or technical novelty.
I dreamed about this logo of an ongoing project among existing Adobe apps.
Does this problem require ML, or a hand-crafted method would be better (to control, debug, understand, implement, etc.)?
Is this better for a paper or just open source?
How did you come up with the architecture, and why?
What I like to see in everything machine-learning paper:
Does this problem require ML, or a hand-crafted method would be better (to control, debug, understand, implement, etc.)?
Is this better for a paper or just open source?
How did you come up with the architecture, and why?
— Li-Yi Wei (@liyiwei) February 15, 2021
How people came up with their ideas!
What I most want to know about but is least described in research papers: pic.twitter.com/PVI8ur97Bf
— Li-Yi Wei (@liyiwei) November 27, 2020
If you are an intern candidate who can (1) propose crazy ideas that nobody else can think off and (2) beat everyone else to accomplish that idea, contact me. 🙂
https://research.adobe.com/internships/
In my personal experience, there are two ways to guide a research project: solving a specific problem with whatever solutions that work the best (based on a variety of criteria such as quality, speed, cost, etc.), and devising a novel idea that can span different problems, domains, and applications.
The problem-oriented approach happens more in engineering (which aims to solve practical problems) while the idea-oriented approach happens more in science (especially more theoretical fields like math which aim to formulate fundamental ideas behind a plethora of phenomena).
Solving a specific problem provides a clear goal and reduces the tendency to derail, while aiming for ideas is more likely to work after one has already worked on related problems so as to condense the experiences into the core forms.
If the productivity of certain types of works can be quantified, it can also be optimized and thus automated.
Instead of performance evaluation, such measures are better used for identifying repetitive tasks potential for automation.
To achieve human-level intelligence, we probably need an AI framework very different from existing rule-based (e.g., Minsky) and data-driven (e.g., current ML) approaches.
During vertigo, the eyes can spin around following the incorrect signals sent from a faulty vestibular system.
The effect can be reduced by VR: track the eye movements, classify whether they are caused by vertigo, and adjust the display accordingly (e.g., move the visual field with the spinning eye gaze).
Some publishers like ACM might ask for the source files to compile the camera ready papers. My understanding is that they need the source to tune the paper format, instead of publishing the source. Thus, it should be OK to share the source directly with the publisher.
If you have concerns about internal annotations such as author discussions not meant for the final paper, just clean up the source via arXiv Latex cleaner or something similar.
Theme: Rubric. Get a free blog at WordPress.com