Edition#2: Causal reasoning and awesome AI Art
Understanding why, duo thinking system, generative perception
Welcome to the 2nd edition of Cross Sections 🎊- a newsletter covering personal reflections and curated content on data science, data viz & communication.
🖇 Causal inference 101
Surely you have heard of ‘correlation doesn’t imply causation’. But this dangling quotation left us wondering: exactly what implies causation?
Another way to ask this question is - how do you know it’s not caused by other factors? Or, to what extent is the intervention effective? There comes techniques originated from econometrics to address such. For example, instrumental variables is a technique that finds another variable (called the instrument) that is related to the outcome yet does not directly cause the outcome. And to get the effect of the treatment on the outcome, one would find the correlation of the instrument on the outcome and subtract out the effect of the instrument on the treatment.
Such consideration is not just mind games - they have practical applications when randomized experimentation is not feasible. Sometimes an act is already carried out and one needs to estimate how things would be different had it not been carried out. As a quick read on how the various causal inference methods work, 5 Tricks When AB Testing Is Off The Table offers a clear illustration.
Some also believe it’s possibly harder to estimate yet easier to generalize than ML, as the features in ML might be the proxies instead of real causes.
For an easy-to-understand overview of essentials concepts explained via case studies, the go-to book would be Mastering Metrics by Joshua Angrist and Jörn-Steffen Pischke.
For a more comprehensive reading list depending on your various preference, there’s Which causal inference book you should read: A flow chart
📚 Recent Reads
This is a section I share about the books, papers, or blogs on data visualization, data science, or communication.
Thinking fast and slow by Daniel Kahneman
This book goes deep on human cognition and decisions. It proposes a framework of 2 systems of thinking - a low-fidelity fast process (system 1) & a high-fidelity slow process (system 2) and elaborates how logic is supported by intuition.
More info: the video interview of the author via lexfridman.com
Think fast and slow in AI by Booch, G. et.al.
This newly published paper hinges on the book above, and suggests translating these humanly cognitive capabilities to the current narrow AI, which tends to specialize in some limited types of tasks like image recognition. It surveys the concept of I-consciousness (I for Information - akin to system 2) Â and M-consciousness (M for mysterious - akin to system 1), and explore how the 2 systems manifest themselves in ML and AI. Â A series of research questions were proposed to develop additional capabilities in AI systems such as causation, generalization, ethical reasoning, and explainability.
🔦 Kaleidoscope
This is an ad-hoc section on curious finds.
This week I’ll highlight some AI art pieces. People have always been searching for new inspirations. The ever-evolving algorithms provide such a venue - artists have been blending neural networks with different mediums of presentation - drawing with robots, creating motion graphics from fragments of history, or placing the digital artifacts into the physical world - like these unreal birds on the real marshes.
There is a lot of novelty and constant new streams of works, as the intermediate steps of these explorations could just be as valuable as the output in terms of creative values. There is also a fair bit of duplication in the field - as many people seem to gravitate towards a limited number of algorithms.
Over time we also see certain work with a creative tint such as old photo recolorization getting productionized, such as in the case of Adobe Neural Filters.
One piece of artwork that resonated with me is the Perception Engine created by New Zealand based artist Tom White a couple of years ago. By constructing abstract representations that cause the classifier to recognize as particular objects, it meditates how the computer articulates the world.
In addition, for a gallery of different ML art filterable by techniques, mediums, and artists, ML Art showcase a wide variety of works.
If want to read up on the different artists, I would go to AI Artists.
For some creative mix mashing of different images (i.e. broccoli, volcano, and Chesapeake Bay Retriever), ArtBreeder could be fun to play with.
For an in-depth read, this article on Generative Portraiture could be a good option/
Hope to see you in the next edition!