Vanishing Gradients

Hugo speaks with Jim Savage, the Director of Data Science at Schmidt Futures, about the need for data science in executive training and decision, what data scientists can learn from economists, the perils of "data for good", and why you should always be integrating your loss function over your posterior.
Jim and Hugo talk about what data science is and isn’t capable of, what can actually deliver value, and what people really enjoy doing: the intersection in this Venn diagram is where we need to focus energy and it may not be quite what you think it is!
They then dive into Jim's thoughts on what he dubs Executive Data Science. You may be aware of the slicing of the data science and machine learning spaces into descriptive analytics, predictive analytics, and prescriptive analytics but, being the thought surgeon that he is, Jim proposes a different slicing into
(1) tool building OR data science as a product,
(2) tools to automate and augment parts of us, and
(3) what Jim calls Executive Data Science.
Jim and Hugo also talk about decision theory, the woeful state of causal inference techniques in contemporary data science, and what techniques it would behoove us all to import from econometrics and economics, more generally. If that’s not enough, they talk about the importance of thinking through the data generating process and things that can go wrong if you don’t. In terms of allowing your data work to inform your decision making, thery also discuss Jim’s maxim “ALWAYS BE INTEGRATING YOUR LOSS FUNCTION OVER YOUR POSTERIOR”
Last but definitively not least, as Jim has worked in the data for good space for much of his career, they talk about what this actually means, with particular reference to fast.ai founder & QUT professor of practice Rachel Thomas’ blog post called “Doing Data Science for Social Good, Responsibly” (https://www.fast.ai/2021/11/23/data-for-good/). Rachel’s post takes as its starting point the following words of Sarah Hooker, a researcher at Google Brain:
"Data for good" is an imprecise term that says little about who we serve, the tools used, or the goals. Being more precise can help us be more accountable & have a greater positive impact.
And Jim and I discuss his work in the light of these foundational considerations.
Links
Jim on twitter (https://twitter.com/abiylfoyp/)
What Is Causal Inference?An Introduction for Data Scientists (https://www.oreilly.com/radar/what-is-causal-inference/) by Hugo Bowne-Anderson and Mike Loukides
Jim's must-watch Data Council talk on Productizing Structural Models (https://www.datacouncil.ai/talks/productizing-structural-models)
[Mastering Metrics}(https://www.masteringmetrics.com/) by Angrist and Pischke
Mostly Harmless Econometrics: An Empiricist's Companion (https://press.princeton.edu/books/paperback/9780691120355/mostly-harmless-econometrics) by Angrist and Pischke
The Book of Why (https://en.wikipedia.org/wiki/The_Book_of_Why) by Judea Pearl
Decision-Making in a Time of Crisis (https://www.oreilly.com/radar/decision-making-in-a-time-of-crisis/) by Hugo Bowne-Anderson
Doing Data Science for Social Good, Responsibly (https://www.fast.ai/2021/11/23/data-for-good/) by Rachel Thomas

Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

Hugo speaks with Heather Nolis, Principal Machine Learning engineer at T-mobile, about what data science, machine learning, and AI look like at T-mobile, along with Heather’s path from a software development intern there to principal ML engineer running a team of 15.
They talk about: how to build a DS culture from scratch and what executive-level support looks like, as well as how to demonstrate machine learning value early on from a shark tank style pitch night to the initial investment through to the POC and building out the function; all the great work they do with R and the Tidyverse in production; what it’s like to be a lesbian in tech, and about what it was like to discover she was autistic and how that impacted her work; how to measure and demonstrate success and ROI for the org; some massive data science fails!; how to deal with execs wanting you to use the latest GPT-X – in a fragmented tooling landscape; how to use the simplest technology to deliver the most value.
Finally, the team just hired their first FT ethicist and they speak about how ethics can be embedded in a team and across an institution.
Links
Put R in prod (https://putrinprod.com/): Tools and guides to put R models into production
Enterprise Web Services with Neural Networks Using R and TensorFlow (https://medium.com/tmobile-tech/enterprise-web-services-with-neural-networks-using-r-and-tensorflow-a09c1b100c11)
Heather on twitter (https://twitter.com/heatherklus)
T-Mobile is hiring! (https://www.t-mobile.com/careers)
Hugo's upcoming fireside chat and AMA with Hilary Parker about how to actually produce sustainable business value using machine learning and product management for ML! (https://www.eventbrite.com/e/select-ml-project-where-value-is-not-null-tickets-284000161127?aff=hba)

Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

Rachael Tatman is a senior developer advocate for Rasa, where she’s helping developers build and deploy ML chatbots using their open source framework.
Rachael has a PhD in Linguistics from the University of Washington where her research was on computational sociolinguistics, or how our social identity affects the way we use language in computational contexts. Previously she was a data scientist at Kaggle and she’s still a Kaggle Grandmaster.
In this conversation, Rachael and I talk about the history of NLP and conversational AI//chatbots and we dive into the fascinating tension between rule-based techniques and ML and deep learning – we also talk about how to incorporate machine and human intelligence together by thinking through questions such as “should a response to a human ever be automated?” Spoiler alert: the answer is a resounding NO WAY!
In this journey, something that becomes apparent is that many of the trends, concepts, questions, and answers, although framed for NLP and chatbots, are applicable to much of data science, more generally.
We also discuss the data scientist’s responsibility to end-users and stakeholders using, among other things, the lens of considering those whose data you’re working with to be data donors.
We then consider what globalized language technology looks like and can look like, what we can learn from the history of science here, particularly given that so much training data and models are in English when it accounts for so little of language spoken globally.
Links
Rachael's website (https://www.rctatman.com/)
Rasa (https://rasa.com/)
Speech and Language Processing (https://web.stanford.edu/~jurafsky/slp3/)
by Dan Jurafsky and James H. Martin
Masakhane (https://twitter.com/MasakhaneNLP), putting African languages on the #NLP map since 2019
The Distributed AI Research Institute (https://www.dair-institute.org/), a space for independent, community-rooted AI research, free from Big Tech’s pervasive influence
The Algorithmic Justice League (https://www.ajl.org/), unmasking AI harms and biases
Black in AI (https://blackinai.github.io/#/), increasing the presence and inclusion of Black people in the field of AI by creating space for sharing ideas, fostering collaborations, mentorship and advocacy
Hugo's blog post on his new job and why it's exciting for him to double down on helping scientists do better science (https://outerbounds.com/blog/hba-excited-to-join-metaflow-and-outerbounds/)

Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

Jeremy Howard is a data scientist, researcher, developer, educator, and entrepreneur. Jeremy is a founding researcher at fast.ai, a research institute dedicated to making deep learning more accessible. He is also a Distinguished Research Scientist at the University of San Francisco, the chair of WAMRI, and is Chief Scientist at platform.ai.
In this conversation, we’ll be talking about the history of data science, machine learning, and AI, where we’ve come from and where we’re going, how new techniques can be applied to real-world problems, whether it be deep learning to medicine or porting techniques from computer vision to NLP. We’ll also talk about what’s present and what’s missing in the ML skills revolution, what software engineering skills data scientists need to learn, how to cope in a space of such fragmented tooling, and paths for emerging out of the shadow of FAANG. If that’s not enough, we’ll jump into how spreading DS skills around the globe involves serious investments in education, building software, communities, and research, along with diving into the social challenges that the information age and the AI revolution (so to speak) bring with it.
But to get to all of this, you’ll need to listen to a few minutes of us chatting about chocolate biscuits in Australia!
Links
* fast.ai · making neural nets uncool again
* nbdev: create delightful python projects using Jupyter Notebooks (https://github.com/fastai/nbdev)
* The fastai book, published as Jupyter Notebooks (https://github.com/fastai/fastbook)
* Deep Learning for Coders with fastai and PyTorch (https://www.oreilly.com/library/view/deep-learning-for/9781492045519/)
* The wonderful and terrifying implications of computers that can learn (https://www.youtube.com/watch?v=t4kyRyKyOpo) -- Jeremy' awesome TED talk!
* Manna (https://marshallbrain.com/manna) by Marshall Brain
* Ghost Work (https://ghostwork.info/) by Mary L. Gray and Siddharth Suri
* Uberland (https://www.ucpress.edu/book/9780520324800/uberland) by Alex Rosenblat

Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe

In this brief introduction, Hugo introduces the rationale behind launching a new data science podcast and gets excited about his upcoming guests: Jeremy Howard, Rachael Tatman, and Heather Nolis!
Original music, bleeps, and blops by local Sydney legend PlaneFace (https://planeface.bandcamp.com/album/fishing-from-an-asteroid)!

Get full access to Vanishing Gradients at hugobowne.substack.com/subscribe