Payal Mitra

Where would non profits (NPOs/NGOs) with modest to little resources benefit from Machine Learning (particularly NLP) solutions the most? A surface level answer could be 'where would they not?'. How can one find and volunteer ML skills towards such projects?

This is where my shallow (but not first) search fails. Here is a 2018 (!) McKinsey report on AI-for-social-good confirming that there should be enough problems to go around. Beyond think-pieces and AI-for-social-good bootcamps (mostly for university students), I don't easily find advertised projects tied to nonprofit missions that allow variable-commitment. Note, I am not including university initiatives or social wings of large companies (Google/Microsoft/IBM, etc).

Here are my scribbles on why there is a gap in reliable resources or 'market'places for ML volunteer work for professionals (Oops, yet another think piece).

1. Wouldn't it be convenient to have a crowd sourced platform for such projects (an Amazon Mechanical Turk for NPO tasks? Not Kaggle). However, given the domain is cash/resource strapped, it isn't a surprise this does not exist.

2. Was #1 not existing simply due to the lack of resources and friction to enable the marketplace? Or could it be a supply side issue? i.e., not enough professionals wanting to/being able to reliably and recurringly commit time to volunteering.

3. Is there a demand side issue? NPOs might be wary of contracting volunteers to support their ML needs given their unsteady and uncertain stream. Further, the organisation might be fire-fighting and not be able to dedicate time to study alternatives, or may not have the ML expertise to know what is feasible.

4. What about maintenance? Beyond a place for buyers-sellers of a need to meet, we also miss a framework that handles the aftercare of any ML solution developed by volunteers. Who takes responsibility? Many NPOs do not have the luxury of in-house tech support. What is a sustainable approach/framework to BAU?

One could of course individually contact the organisations whose missions they believe in (ChatGPT also suggests this). However, I am guilty of not following through. Or, we could build such a crowdsourced platform? I hope to have more than questions.

"CRISPR gene editing technology is like an operating system that scientists can build biotech apps on" - an idea from 'The Code Breaker' by W. Isaacson that blew my mind.

CRISPR had an explosive impact on the biotech world as it offered a foundational tool that scientists/practitioners could use to bring their use-cases, or 'apps' to life. Be it to edit a gene to act as a diagnostic tool, to act as a mechanism for vaccines, or to edit out undesirable diseased genes. 'Bio-hackers' even had home-kits for gene editing!

The more ubiquitous and pervasive example that touches most humans - is how computer operating systems (OS) such as Unix-based Mac OSX, Windows, or the open-source Ubuntu OS distributions have made possible the computing devices and software that now act as an extension of our lives. (As an aside: Ubuntu philosophy)

The point I am getting to is that an Operating System acts as a platform, that unlocks the potential of human ingenuity by allowing the creation of applications (apps) on it.

I have a bittersweet relationship with newsletters.

It is easy for the rate of collection of newsletter subscriptions/receipt of newsletter emails to far exceed the rate of absorptive reading. It reminds me of something someone said: "Buying books and reading books are two different activities/hobbies".

If you are at all interested in keeping abreast of economic/business journalism in India, or on deep think pieces on society and economics globally, try these posts:

1. Finshots ('The Ken' is a close contender)

Fascinating topics and well-researched delivery of financial business & economic news from India. Highly educative - on startup valuations, government financing, financial innovation, failures, etc. Somehow, it is free! I would definitely convert if it were paid.

'Prompt' has become an everyday word on account of the popularity of ChatGPT. However the idea of prompting any language model (particularly decoder, or encoder-decoder models) to get a response back in natural language has been around for a few years. Here are a few ideas from research that can help you perform few-shot learning.

This musing is mostly born from the challenges I faced while trying to load LLMs locally (for inference and fine tuning) that are 1/ 1000th of the parameter size of GPT-4.

The dilemma around how transparent, democratised, scalable, and sustainable the hugely popular LLM capabilities (such as ChatGPT) are to practitioners across the world, reminded me faintly of the Blockchain Trilemma and Blocksize war.

Well, this is more of an analogy rather than a theoretical similarity between the dilemmas these two very different technologies face when they attempt to scale. The Blockchain Trilemma states that in any blockchain you can only fully solve 2 of the 3 problems of security, scale and decentralisation (however, there are now workarounds (yet to stablise) to solving this, such as Sharding, Roll-ups)

Granting myself a creative license to stretch the analogy. If we view the LLM deployment issues faced today from a blockchain trilemma lens, then we face -

I've recently been reminding myself about why I wanted to write in the first place. A daily writing habit for the self is the goal - but, publishing and expecting an external reward is counter-productive at times.

Thus, I paused and found myself going back to two pieces. The title fuses quotes from two of my favourites, Kurt Vonnegut and Elizabeth Gilbert. It contains their meditations on practicing creativity for the self, and emphasises practicing your creativity without expecting an external reward of any sort.

Sharing my fav pieces on this topic:

1. Kurt Vonnegut's response to a letter from the children at the high school in 2006. Teaser excerpt to get you to read the original:

   • "Here’s an assignment for tonight...Write a six line poem, about anything, but rhymed.... Make it as good as you possibly can. But don’t tell, show it or recite it to anybody. Tear it up into teeny-weeny pieces, and discard them into widely separated trash receptacles. You will find that you have already been gloriously rewarded for your poem. You have experienced becoming, learned a lot more about what’s inside you, and you have made your soul grow."

2. Elizabeth Gilbert's Big Magic.

   I had a great many takeaways, but they mostly fall into 4 buckets:

   • "do not demand creativity to pay the bills", or

   • "trust that writing loves you back, so keep showing up for it".

   • "Take care of it (your creativity), and it will take care of you."

   • "follow your curiosity"

Hope these recommendations find you, and that they are useful.

"Why AI Is Incredibly Smart — and Shockingly Stupid" is a Ted Talk by AI Researcher and Professor, Yejin Choi. I highly recommend you watch this excellent 16 minute clip. It talks about the focus beyond just building ever larger and larger language models.

Should you watch this? Either you

• are an AI researcher in either camp (scale enthusiasts, vs. let's-get-smarter-in-building-smaller-models-with-(distilled)-knowledge).

• are looking to utilise ChatGPT-esque technology in your product. It is the user's prerogative to know the cautionary boundaries, and how to overcome them

• are a student/AI practioner without the resources to be able to play or fine tune the models. The open source community (you're a part of it) will find a way.

• are jumping wagons with each shiny new ML tool and are overwhelmed by the pace

Why I liked it

• Echoes a worry I share re: making these models smaller, sustainable and more democratised? Call for/to action.

• Excellent presentation, with insightful analogies.

• I enjoyed Yejin's take on Common Sense Reasoning being the most important task.

• Spells out cautionary lines for people to when integrating ChatGPT-esque tools

• Personally, it re-emphasised the value of going long on certain notions of how you want to contribute. Think of the bigger picture and not get swayed by latest tech

Though, I do want to say - credit where credit is due. Large Language Models and techniques like Reinforcement Learning with Human Feedback are phenomenal feats, and give us fantastic starting points to distil and fine-tune task-directed models.

y = b1x+b0. If x is 1D, this linear model has 2 parameters. In contrast, recent Deep Neural Network(DNN) model 'GPT-3' powering ChatGPT has 175 Billion parameters. 'GPT-4' of ChatGPT Plus has 1 Trillion. Beyond the explosion in number of parameters, DNNs learn complex non-linear functions and are said to be non-interpretable in how they learn to map input(x) to output(y).

But what does interpretability in DNNs mean? (this linked article attempts definitions). For this post, we consider a model as interpretable if we can perfectly predict the output, and can infer how input data changes affect predicted output. Is the model's decision path hard to find or is it that the theoretical framework and design choices in training the models do not easily account for interpretability?

Below I describe 3 aspects of the underlying learning theory for a large section of DNNs (excluding Bayesian Deep learning, etc) that bake in non-interpretability.

Illustration: Imagine we train a DNN to predict the agricultural crop yield of land plots. The model takes as input the raw sensor data on water levels and weather of the land.

• Deep learning does not learn statistical relationships between dependent target variable (y) and input variables (x). DNNs do not rely on explicit input features, but learn their own 'latent' features and representations of input.
  Illustration: The crop yield prediction model does not learn explicit relations linking water levels to crop yield, but hunts for its own patterns in the sensor data (and learns its own input features) that give best performance in predicting the yield.

• Deep learning does not explicitly account for relationships between input variables.

  Illustration: The DNN does not model how weather (e.g. excess temperature) might affect water levels. It thus can not predict how much the yield would change as a function of interlinked changes between inputs. It only predicts on observed data.

• Most importantly, Deep learning does not assume probabilistic processes behind the data generation of either input or output.

  Illustration: Input variables weather and water levels are collected across different days and have distributions. High temperatures and low water levels might require a different crop yield prediction function, while a period with moderate rainfall follows a different yield trajectory. The likeliness or unlikeliness of the draw of a certain input event (tsunami?) should change the crop yield function. Our model does not consider the data to be generated or related in a probabilistic manner.

Despite the fact that Machine Learning draws upon on aspects of Statistics and Statistical Learning Theory (SLT), it has many differences.

Much of the math, statistics, modelling and computer science algorithm concepts that make ML possible are abstracted away in present day ML/DL code frameworks (sklearn, pytorch, pytorch-lightning, etc) that allow you to train+predict using an ML model in a few lines of code. Thus, it is very natural to wonder at the differences and similarities between the concepts, particularly if you are approaching ML without academic training in/introduction to the underlying subjects.

Here I share 3 resources to hopefully clear some of the doubts.

This is not a self-help list for rekindling the joy/purpose in a craft you previously loved. Rather, it is far more tangential, and reads like a peek into the musings and of an articulate and established writer. You may take what you will from it.

The author themselves is a writer by profession, and talks about how blogging satisfies a certain itch for him. Payed subscription letters are not conducive enough an outlet for his creativity, indulgences.

The Recommended Blog: Salt Seeking by Venkatesh Rao. And here's why:

It recently dawned upon me that training and serving of ML models can be viewed as a supply chain across companies/part providers. This is a random musing, but if you are into the ML development, then hang on for the metaphor in the end.

Up until a few years (10?) ago, before the widespread use of cloud compute, many ML models would be trained completely in-house on-prem resources (the history of aws says AWS was launched in 2003 as a bid to manage infrastructure scaling for engineering needs). Nowadays, many in the industry use some form of cloud resources to train (/serve) their ML models salably and reliably. An analogy would be developers renting out factory equipment to perform their production deployments.

Beyond cloud compute and scaling solutions, there are several great third party tools, such as those for tracking of ML experiments, data versioning, ETL, creating pipelines from ML components, automating pipelines, and even auto-hyperparameter tuning!

As everyone rushes to use OpenAI models, build atop its APIs, or use services like ChatGPT in their workstreams, many wonder if the dawn of super abundant (conversational) AI is already here.

As someone working in NLP and Deep learning in a non-MAANG or non-startup-born-in-today's-digital-era, my view is that adoption of Large Language Models (LLMs), even open sourced ones, in existing service/product industries (such as healthcare, publishing, etc), will be slower than you think

The idea of LLMs have been around for a while, most notably since BERT models in 2018, but I've seen resistance to its adoption in some companies. One might argue that the recent ChatGPT wind has woken up management professionals to the possibilities of LLMs (although I have heard too many people refer to LLMs and Generative AI and ChatGPT interchangeably - and that worries me). However the integration of such models into company workflows has been slow. This is due to a number of reasons.

These are behaviour patterns that signal effective leadership, at all rungs. They inspire trust in management and drive successful outcomes. This is from observation and applied experience (I'm no CxO, but I take ownership). #everyoneIsALeaderClub

1.Regular Communications (Thoughtful, informative, and concise)

Don't let the dizzying pace of advancements in the ChatGPTera overwhelm. Be inspired by it, accelerate your learning & build cool stuff from shoulders of communities.

One grossly under-emphasised skill for ML projects for both the Data Scientist and the team, is quick, interactive prototyping of an idea.

Five years ago building web services and apps for solely demo purposes of ML tools was painful and less-accessible, often requiring more unicorn data scientists with hack-y full-stack skills. This has dramatically changed in the past couple of years with beautiful efforts and frameworks to help build quick web apis for your ML models (e.g. FastAPI), and also interactive web app prototypes for your ML model or pipeline (e.g.: Streamlit, Gradio) in next to no time.

Why is this important?

• setting projects up for success with better communication of ideas to stakeholders

• helps in extracting feedback early in project from stakeholders (decision makers + users/customers)

• building end to end prototypes tests ideas and holes, and can make iterations stronger and faster

• better communication b/w cross-disciplinary team members - UX, data scientists. engineers, product managers

• can reduce time to production greatly

  I have lost count of the number of times a certain ML project either does not move past the first experimental models, or moves too slowly. Much too often we only see lifeless tables of precision recall values, or slideware telling the story of the Data Science endeavour/intervention. Undoubtedly these tables store the most value and are the output of great skill+time to build and test ML models. However, would it not be transformative if we could bundle that sentiment classification model as a web service and expose it over an api to demonstrate its value to stakeholders? Or if we could have prospective users clicking on the cool Generative AI (ChatGPT-esque) tool to get feedback from users? I have found prototyping tools to be the most amazing research aids to ensure that what we build is what users need and want. It inspires the team and keeps the idea flywheel whirring.

I'm adding to the text jungle on the 'AI taking over jobs' debate, because my mulling continues. Here are two aspects that I find myself filing my musings under: