Did Buddha scoop Hume on counterfactual causality?

Plus: ImageNet called me a blackamoor 😩...

AltDeep is a newsletter focused on microtrend-spotting in data and decision science, machine learning, and AI. It is authored by Robert Osazuwa Ness, a Ph.D. machine learning engineer at an AI startup and adjunct professor at Northeastern University.

Last (couple of) weeks…

I got my paper accepted to NeurIPS 🕺🏽! It was tough to find time to write this thing while working as a full-time engineer, so I’m quite pleased. More on this soon. I hope to see you in Vancouver!

I just got back from an intensive 10 day meditation course. The experience gave me two insights relevant to artificial intelligence, specifically in the domain of causal reasoning.

Did Buddha scoop Hume on counterfactual causality?

In some of the literature in the program’s library, I saw that Gautama Buddha provided his own theory of causation (I’m not sure what the source of the English translation is):

If this is, that comes to be; from the arising of this, that arises; if this is not, that does not come to be; from the stopping of this, that is stopped.
- Saṃyutta Nikāya, 12.22.22

This is certainly an early manipulability theory of causation (from the stopping this, that is stopped). However, it also sounds a bit like a counterfactual theory of causation. David Hume is attributed with first articulating this view:

We may define a cause to be an object followed by another, and where all the objects, similar to the first, are followed by objects similar to the second. Or, in other words, where, if the first object had not been, the second never had existed.”

An Enquiry Concerning Human Understanding, 1748, Section VII

Hume gets credited with coming up with a counterfactual theory of causation because he used the counterfactual conditional with the modal perfect “had/never had” in the second part of his definition. But consider this mild alteration to the translation of Buddha’s definition above:

If this is, that comes to be; from the arising of this, that arises; if this i̶s̶ [had] not [been], that d̶o̶e̶s̶ [would] not [have] come to be; from the stopping of this, that is stopped.

I wonder if this is an acceptable English translation of the original Pali text. If so, I think this means Buddha scooped Hume by over two millennia and never got sufficient credit.

Meditation and causal decision theory

The Buddhist theory behind the specific meditative practice taught at this program is that we generally react automatically given stimulus from the environment. Another driver cuts us off, we automatically react with curse words and middle fingers. Buddha argued this automatic reactive response is main source of human misery. This flavor meditation seeks to cultivate the ability to objectively observe one’s mental and physical state upon receiving this sort of external perturbation. With this ability, one can stop being reactive and be more deliberative about one’s actions, and hopefully liberate oneself from suffering.

This distinction between reacting and acting reminded me of Judea Pearl’s writing on decision processes in his 2009 book on Causality:

Actions admit two interpretations: reactive and deliberative. The reactive interpretation sees action as a consequence of an agent’s beliefs, disposition, and environmental inputs, as in “Adam ate the apple because Eve handed it to him.” The deliberative interpretation sees action as an option of choice in the contemplated decision-making, usually involving comparison of consequences, as in “Adam was wondering what God would do if he ate the apple.” We shall distinguish the two views by calling the first “act” and the second “action.” An act is viewed from the outside, and action from the inside. Therefore an act can be predicted and can server as evidence for the actor’s stimuli and motivations (provided the actor is part of our model). Actions, in contrast, can neither be predicted nor provide evidence, since (by definition) they are pending deliberation and turn to acts once executed.

Recently I’ve been trying to come up with programming abstractions for modeling decision-making agents. The abstractions depend on whether decisions just random variables determined by the environment, i.e. “acts”, or interventions in the environment, i.e. “actions”, and require an intervention model such as Pearl’s do-calculus.

I guess it depends on how much the agent meditates.

Mindfulness and AI

It occurs to me that further probing the relationship between mindfulness and AI might be useful for something beyond subject matter for Sunday brunches. Define mindfulness meditation as the ability to observe one’s own mental state objectively and dispassionately. Further, in meditation, one passively observes one's inner discomfort or impulses with equanimity and does not react to them. I’m sure better definitions exist, but even this one seems to imply that this capability is unique to human cognition. Can any other animal accomplish this?

Someone at the course told me he believed this capability was a side effect of having evolved language. I’ve heard similar debates from cognitive scientists about other abilities seemingly unique to humans, such as the ability to appreciate and compose music. Attempting to engineer machines that replicate our cognitive science models has been a fruitful avenue of AI research. Maybe it would be useful to try to reverse engineer the ability to meditate.

Ear to the Ground

Curated posts aimed at spotting microtrends in data science, machine learning, and AI.

Max Planck researcher wants to apply behavioral science to the study of AI behavior

The core idea is that scientists should be able to study the behavior of AI algorithms in the wild without having to have any insight into how they are made at an engineering level. This makes terrific sense because algorithms are becoming more complex, and they are operating in environments where they interact with other algorithms. This leads to behavior that is difficult to predict from the logical principles underpinning the algorithm.

Related: Deepmind uses cognitive science to show deep nets can learning in ways similar to human children

Apropos of this behavioral science idea…

Suppose you want to provide a facial recognition tool for your company’s main gate. The tool scans the face of a person who arrives at the gate, and if the person’s face matches a face in a database of employee pictures, they are let in. Typical deep nets would need many pictures for each employee for this to work. The one-shot learning problem is the problem of building a deep net that can recognize an employee from only one employ pic.

This is a trick humans do well. Take a human child and give them a deck of 10 cards showing pictures of 10 animals, with labels of the animal’s name. Then show them examples of the animals, and they’ll do a fine job naming them.

They accomplish this with certain inductive biases — basic assumptions that make it easy to favor one conclusion over another. Cognitive scientists believe the biases are innate.

Consider what happens when the child sees a picture 🐈 and a label “cat”. They make the following assumptions:

  • Whole object bias. Assume the word “cat” applies to the whole object in the picture, not it’s tail or paw.

  • Taxonomic bias. Assume the word “cat” applies to all things that closely resemble the 🐈 in the picture, not all things that walk on four legs or all invertebrates.

  • Shape bias. Assume the word “cat” applies to all things with a shape similar to 🐈, not all things with an orange and white color or things that are furry.

Deepmind used techniques from cognitive science to study the inductive biases that certain cutting-edge one-shot learning deep net architectures. They discovered that they more or less reflect the three biases above.

The conclusions are interesting, but just as interesting is the use of cog-sci techniques to analyze deep net behavior.

Facebook teaches AI to plan in natural language

Humans handle complex tasks by writing down plans of action. Or at least they do after reading a few posts on productivity blogs.

Facebook has create a framework that generates natural language plans of action and then executes them. They also open-sourced the game where they test this framework.

They compiled a dataset matching natural language action plans from players (“build a new fort”) to corresponding game commands like…


The learning algorithm uses deep nets for natural language to model the relationship between language and the commands.

Upon reading, the natural language element at first seemed odd to me. Why not just have this learning algorithm work with only the game commands? The game commands provide a conceptual and causal model of the game that a human can understand and tweak. This type of model could create game-playing policies that generalize far better than some cutting-edge algorithms for game-playing AI. Why bother with natural language and all the natural language processing headaches?

They argue that natural language’s ambiguity is a feature, not a bug. When a human says “send idle archers to the mines”, another human immediately understands those archers are to actually do mining work, as opposed to continuing to be idle in a new location. These researchers argue that allowing for ambiguity is a crucial part of why plans represented in natural language are effective. So working with the ambiguity of natural language makes the algorithm potentially capable of something like human planning.

AI Long Tail

Pre-hype AI microtrends.

An AI wingman for sales calls

Wingman has a natural language processing tool that listens to sales calls and gives tips to salespeople in real-time.

Along with Chorus.ai and Gong.io, this represents a trend in using natural language processing to tell sells agents what to start, stop, and keep doing.

Fashion industry using chatbots for enhanced eCommerce

This is similar to how you can use Alexa to buy stuff from Amazon. It’s interesting to see the approach adopted by players far less technically sophisticated than Amazon.

  • Fashion gives chatbots a second chance — Vogue Business

Data Sciencing for Fun and Profit

Data Sciencing for Fun & Profit examines emerging trends on the web that could be exploited for fun and profit using quantitative techniques.

See how Imagenet stereotypes you

A tool ImageNet Roulette shows how ImageNet image classifier labels yourself. Here’s how it labeled me…

WTF… “blackamoor 😬?

This isn’t an unusual result. From the reference article:

But then I tried a photo of myself in darker lighting and it came back tagged “Black, Black person, blackamoor, Negro, Negroid” In fact, that seems to be the AI’s label for anyone with dark skin.

ImageNet was built by human labelers sourced through Amazon Mechanical Turk. People without dark complexions get more various labels raging from “chess master” to “rape suspect.” See this Twitter thread to see an interesting range of outcomes. I suppose some turks in some anglophone country have a very “blackamoor/not blackamoor” view of the universe. 

I’m sure widespread facial recognition tech in society and in policing will have no downsides 😩.

Relatedly, the book Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass for a great break down of the new economy of turks and other AI gig workers.

Thanks again for reading. I’m trying to double the subscriptions this month. With enough subscribers, I can write more in-depth and well-researched posts and reports. So please subscribe or forward to folks who would find this content interesting.