On data, causality, and AI

Pierre-Simon Laplace in Théorie Analytique Des Probabilités

We need more decision / data scientists to ask “What is the probability of sun rising tomorrow?” yet we don’t seem to put a strong emphasis on probability theory for several reasons.

Probability is not thoroughly covered (ideally as a standalone course) in most data science / business analytics programs. In predictive analytics, common packages/libraries for ensemble methods focus on classifications, almost hiding the probability calculations (which are distorted in some anyway). Most frequentist reporting are limited to point estimates and errors, again hiding the underlying probabilistic assumptions. Et Cetera.

In search of a short reading to share with my students, I’ve come across a recent book (updated in May, 2022) by Dirk P. Kroese, Zdravko Botev, Thomas Taimre, Radislav (Slava) Vaisman Ph.D. The book is open access and the appendices serve as a nice refresher. Appendix C is a little primer on probability for example.

Source

In this latest collaboration with Duygu Dagli, we shed some light on it and soon other topics. We plan to bridge conceptual and applied takes on statistical modeling and machine learning methods and the business problems they may help solve. We’ll likely touch on causal inference, predictive analytics, and how firms can organize to be more data centric. Check it out and stay tuned for the upcoming pieces on experiments and the analysis of experimental data.

Source

Jeremy Siegel almost loses it here for good reason. The Federal Reserve Board called it transitory inflation when prices were skyrocketing last Fall. When the same Fed (except a couple members) now argues that inflation is not yet slowing at a reasonable pace despite the price contractions recent data shows, questions arise. Fed gives the impression that it is now overcorrecting what it overcorrected earlier by loosening the monetary policy too much and for too long.

In case after case, our data modeling and inference practices are tested against lags in data. Lack of a high predictive power, we resort to overcorrections. Overcorrecting is doing more than enough (vs. not doing enough) and sounds better than coming short. But then, the pendulum swings back a little harder.

What do we learn from such swings? Well, one rather obvious takeaway is to put more emphasis on correctly understanding and modeling lags in time series. Another one is to be content with coming short occasionally, especially when the cost of overcorrecting is much higher than the cost of coming short.

Source

The commercial refrigerator/shelf looking device in the picture is an ikejime machine. Ikejime is considered the fastest and most humane method of killing fish. The method also leads to the best taste of fish because fish are killed instantly before their bodies go into distress and produce lactic acid and ammonia into their muscles. Ikejime involves the insertion of a spike quickly and directly into the hindbrain, causing immediate brain death.

That is, if fishermen know where the hindbrain exactly is for each species, and can insert the spike quickly and precisely within minutes of catching a fish, and have time to do so repeatedly. Well, that’s what robots are for.

Shinkei Systems’ machine is a combination of hardware and an edge detection algorithm, the engine behind object recognition in convolutional neural networks. Challenges abound. The machine operates on a fishing boat that tilts around even at zero speed. Apparently, “even in the same species, even with the same contour, the brain can be in a different location” as well.

Working with fishermen in Maine, New Hampshire and Cape Cod, Shinkei Systems seems to have been accomplishing the task on fresh-caught fish at a rate of one every 10-15 seconds. Moving forward, accuracy should increase and time to complete the task should decrease, leading to further opportunities.

Source

“We wanted to make sure people with proximity to the data, their country, the language they speak, had a hand in choosing what language came into the model’s training,” says Jernite.

#BLOOM, the first truly open-science, open-access, and multilingual (46 languages) large language model with 176B parameters (slightly larger than GPT-3) will soon be released as a complete pretrained model. Behind the project is #BigScience, a wide-scale collaboration of over 1,000 researchers.

The project is quite impressive overall, both for the extent of collaboration and outcome. It’s also an engineering delight to watch. The model has been trained using 384 A100 GPUs (with 80 GB of memory each) since March 11, 2022.

BigScience provides updates on training every day (having hit its initial target earlier than planned, the model is currently being trained for “a few more days”). See the links in the comments to follow the updates and download the model. The full model will be released on #HuggingFace (also a partner of the project).

This is a significant step forward for at least two reasons: the way the training data was collected and the core values behind the initiative. BigScience seems to have prioritized data quality by hand crafting the training data. In a world of models that favor kitchen sink approaches (because they can!), this is a progress. More obviously, BLOOM paves the way for a true democratization by removing the strings that have been attached to the use of such models by #OpenScience, #Google, and #Facebook (apply for API access, accredited researcher only etc.).

Source

While attending the Symposium on Data Science & Statistics to present our study in Improving Algorithms for Big Data session two weeks ago, I learned about useful new methods (and met the great people behind them).

One of my favorites is Sparsity Ranked Lasso (SRL) by Ryan Peterson. The paper mainly focuses on lasso but the idea is also extended to other regularization approaches such as elastic net.

Takeaway: Use SRL over ordinary lasso especially if your model has interaction terms and polynomials. On average, SRL’s predictive performance is better than lasso in the 112 datasets from the Penn Machine Learning Benchmark database. Ryan goes on to show also that SRL overperforms a Random Forest (RF) in a case study both in accuracy and efficiency. Even if SRL performs on par with a RF, why not use SRL as it is both interpretable and explainable!

The part I loved about SRL is the simple yet important challenge, which the authors call “covariate equipoise”: the prior belief that all covariates are equally likely to enter into a model. Basically, a model’s simplicity is usually defined by its parsimony: the number of parameters. This is no matter whether a parameter is an interaction (or a polynomial form) of the other terms in the model. This is problematic for obvious reasons and SRL solves it by treating covariate groups differently based on their type.

And yes, there is a package for that: sparseR. Link to the R package and nicely written paper are in the comments.

R package – Paper

A seemingly persistent effect of the pandemic on #Uber is a 50% decrease in the mobility’s share of revenue (a decrease from a 80% share to less than a 40% share of rides in total revenue). Based on revenue, Uber is now more a delivery company than a mobility company.

This is data centricity extrapolated: a shift from carrying people to carrying objects while solving pretty much the same data-driven optimization problem. The article is from last year but the effect persists as of Q1 2022: Only 37% of the revenue is from carrying people.

Source

Now that object detection is almost a solved problem, work on the next frontier, text-to-image generation, began to thrive. #Google Research’s most recent work on generative models, #Imagen, uses text embeddings from a large language model called #T5 (similar to #GPT3 and #OPT175B) to encode text for image synthesis.

Interestingly, the study finds that increasing the size of the language model improves performance more than increasing the size of the image diffusion model. Imagen achieves exceptional similarity between real and synthetic images (measured by the distance metric FID, Imagen achieves a score of 7.27 on the COCO dataset). Human raters confirm the performance of the model.

The paper is nicely written with a much-needed ethics discussion at the end, and full of colorful images. Apparently, Imagen does not perform as well when generating images that portray humans.

Synthetic data generation and image restoration are two common use cases of #GANs. I will post a link to one such study on medical images in the comments. #Arts and #crafts is obvious. I can also think of use cases for #fashion and potentially #personalization of products in #retail. What are some other business use cases?

Source

Interesting read and perspective on modeling the learning in hippocampus and potentially applying the model structure to the design and development of algorithms. Clone-structured cognitive graph (CSCG) uses Markov chains and dynamic Markov compression. So CSCGs form a probabilistic sequence model.

Source

Models remain interpretable to the extent that the components of the original function are retained. The authors claim 55% to 92% improvement in accuracy in short to medium-term forecasts, which is impressive if generalizable. Model training time increases 4-fold but prediction time improves 14-fold. Developed on PyTorch so it can be parallelized and deployed on GPUs, potentially to reduce training time. Ported to R but using a Python environment.

Looks promising especially for “AI on the edge” type mobile applications.

Source

Meta AI’s release of Open Pretrained Transformer (OPT-175B), which is on par with OpenAI’s GPT-3 at 175 billion parameters/weights, emphasizes responsible compute and claims one-seventh the computational cost in terms of carbon footprint. Pretrained model weights are free to download (link in the comments). This is good news for open collaboration and better news for the environment.

Source

Birds seem to beat humans by the sheer variety of training sample they provide when it comes to modeling bipedalism. Such high variety, effective learning offers a takeaway for analytics problems: Focus at least as much on data as on modeling it.

Source

You may have heard of #ESG (Environmental, Social, and Governance) investing. It’s also called “socially responsible investing” when ethics is added to the picture. Public companies are assigned an ESG score, which is a quantification of the social impact. What social impact though? You would probably expect ESG ratings to quantify the societal impact of (not on) a company, right? Well, you’ll be disappointed. “Socially responsible investing” is a misnomer when associated with the ESG ratings, at least those reported by #MSCI, a leading provider of the ESG ratings globally.

MSCI basically quantifies the impact of environmental, social, and governance risks on a company’s operations (not the other way around!). In other words, if we rely on ESG ratings while making investment decisions, we may not be doing any social good. We are essentially ensuring that our investments are protected from the environmental, social, and other risks such as climate change. After all, why would we care about the carbon footprint of our investments on the environment as long as profits are good?

MSCI’s plot offers some takeaways on how to generate #data and model it. Apparently, measuring reverse #causation and packaging it to look like the #cause and #effect are in the right place can be quite profitable. To be fair, MSCI is explicit about its data generation and modeling process residing in the #darkside.

Source

A valuable insight on the proof-of-concept to production gap in computer vision that underlines again the importance of context:

“It turns out,” Ng said, “that when we collect data from Stanford Hospital, then we train and test on data from the same hospital, indeed, we can publish papers showing [the algorithms] are comparable to human radiologists in spotting certain conditions.”

But, he said, “It turns out [that when] you take that same model, that same AI system, to an older hospital down the street, with an older machine, and the technician uses a slightly different imaging protocol, that data drifts to cause the performance of AI system to degrade significantly. In contrast, any human radiologist can walk down the street to the older hospital and do just fine.”

Source

An obvious but often neglected fact is the overemphasized value of accuracy as a performance metric. In a two-class problem where 99% of the cases are of 0 (Not a spam email), achieving an accuracy of 99% is as easy as classifying all emails as safe. Sensitivity, specificity, and other metrics exist for a reason.

The story of Waymo, Google’s self-driving car, resembles the value of solving the remaining 1% of the problem where conventional machine learning gets stuck due to the limitations of training data. If 1% of the error turns into a make or break point, one needs to get creative. On a long tail that extends to infinity, walking faster or running does not probably help as much as a leap of imagination.

I must note that it’s not fair to expect an autonomous car to be “error-free” given we do not expect human drivers to perform error-free at the driver license exams and road tests. The two will just make different errors.

To me, this is interesting not because of the lack of transparency in methodology but the potential reason for the rankings to be wrong.

I want to believe that this is a mistake not fraud, but really? Applying the weights before normalizing the scores? And the Bloomberg Businessweek spokesperson says “the magazine’s methodology was vetted by multiple data scientists.”

I have created a quick scenario as a reminder to my former (and current) students (posted in the comments as LinkedIn doesn’t allow here). In the example, the scores are standardized across the five items (which are randomly generated and assigned weights). In the Businessweek rankings, standardization is supposed to be across institutions so that the weights proportionately affect each institution’s score on the corresponding item. Nevertheless, the source of the error is the same. If the weights are applied before normalizing the data, the scores are adjusted by the weights disproportionately. Ranking changes accordingly.

“Zillow Quits Home-Flipping Business, Cites Inability to Forecast Prices,” WSJ reports.* I try to avoid passing along news stories but it’s not everyday I receive a predictive analytics story as breaking news.

I wonder whether the reason is really “an inability to forecast the prices” or “relying too much on an ability to forecast the prices” for a “$20 billion a year” venture as it was debuted.

Zillow announced plans for this data-driven venture in 2018 by citing consumers who “expect magic to happen with a simple push of a button.” In a statement yesterday, Zillow seems to have realized magic is not happening: “But it would be naïve to predict that unpredictable events won’t happen in the future.”

Maybe it is never a good idea to develop a whole business model that grossly underestimates the changes in error (both reducible and irreducible) due to potential bifurcations in market forces.

Source

Alex Webb of Bloomberg offers a linguistic explanation for why technology ceased to be meaningful:

“English lacked an equivalent to the French technique and German Technik. The English word “technique” hadn’t caught up with the innovations of the Industrial Revolution, and it still applied solely to the way in which an artist or artisan performed a skill.”

He contrasts technique as in “artistic technique” in English with technique as in “Lufthansa Technik” in German and argues that technology emerged in the early 20th century for the lack of a better alternative.

Whether the reason is linguistic, sheer overhype, or semantic satiation, we may be better off dropping the “tech company” reference at this point unless it is elaborated further. For the companies that are more tech than your average tech, a good alternative may be “deep tech.”