Ensemble is a monthly issue from The Overfit, aggregating the most relevant updates in the Data Science & AI industry for the given month, always with some personal takes on “why is this happening?” and “what will happen next?”

AI News

While January was fascinating with all the noise from ChatGPT, February showed that the AI wave continues.

• ChatGPT is integrated with Bing, Edge, and Windows 11. The partnership OpenAI has with Microsoft shows — ChatGPT is already available for multiple products.

• Google announced Bard. In short, it’s the ChatGPT of Google. It’s based on another architecture — LaMDA. It’s not yet available to the public, so it’s hard to compare.

• Snapchat launched an AI Chatbot powered by Open AI. It’s very similar to ChatGPT, but they have tuned it a bit, trying to avoid sensitive topics (violence, swearing, etc.)

• Meta launched Llama. It’s available in different sizes — in billions of parameters — (7B, 13B, 33B, and 65B). The 13B version outperforms GPT-3 (ChatGPT “basis”) in most benchmarks and is way more efficient — GPT3 has 175B parameters (Llama is more than 10x smaller)

Generally, we should expect more products based on this type of AI — generative text, chatting, etc. — soon. There will undoubtedly be a lot to talk about in the coming months! ;)

Layoffs

In the last ensemble article, I discussed why layoffs are happening. As I mentioned, layoffs were still expected in February.

This time, the most significant layoffs for February are:

• Ericsson: 8500

• Dell: 6650

• Yahoo: 1600

• Byju’s: 1500

• Twilio: 1500

• Zoom: 1300

And many other “smaller layoffs” from notable companies such as GoDaddy, eBay, Thoughtworks, Wix, etc. You look at them individually at layoffs.fyi.

I still expect some layoffs during March (as some planning for the year finishes during February), but I guess the “layoff velocity” will decrease after April.

Notes on the market

Even though we still have uncertainty in the market and layoffs should continue throughout the year, this year will most likely be an excellent opportunity for AI-based startups to get some traction.

For instance, Tome — a “Generative Storytelling” app — is leading a $43M investment round. And OpenAI has a dedicated fund specifically for OpenAI-powered startups.

Because of all the buzz, many startups will either surge or pivot to embrace AI solutions. Let’s see which ones will stick around!

In this article, I selected some topics extremely relevant to AI professionals. I will explain what they are and recommend where you should start for each of them.

1. Interpretability

Models generate predictions. However, those predictions by themselves might not be enough:

• You might be developing an application that requires telling WHY the model predicted X and not Y. [e.g., explain why the model thinks the person has a particular disease; or explain the reason for a loan denial]

• Additionally, when iterating and trying to improve your model, it is an excellent practice to go over some predictions and understand WHY the model is performing poorly in some instances.

Interpretability techniques can help in both cases — they inform “the why behind your model’s predictions.” 

A fundamental interpretability technique is the “Feature Importance” in decision trees based models. It shows what features matter the most to your model. However, it does not provide enough clarity to explain specific predictions. 

For that, you can use techniques that provide “local interpretability.” The most well-known are LIME and SHAP. The remarkable thing is that they are model-agnostic — in other words, they work for any ML model. I recommend diving deeper and understanding how they generate local interpretability instead of using them blindly.

There are, of course, many others. Some of them are specific — Such as Layer-wise Relevance Propagation (LRP) for Neural Networks—but I advise focusing on LIME and SHAP, as they already cover the most common use cases.

2. AI Fairness

All models learn some bias. In a certain way, bias is necessary for making predictions (you learn them through your data and exploit them to forecast the future).

However, not all bias is wanted. For instance, a model that favors a particular ethnicity over others is unacceptable — e.g., only approving loans for white people. In other cases, using ethnicity can help provide a more accurate prediction for them — such as in a medical application.

Learning to identify and mitigate unwanted bias is vital for those applications, and AI Fairness is the field that takes care of this. I have published an article that goes through a more detailed explanation of how to identify and mitigate those biases — you can read it here.

Anyway, here is a fantastic table referencing some of the AI Fairness techniques:

3. AI Privacy

Models learn from data to make their predictions. Sometimes, that data can be sensitive. Some malicious people can try to use your model to reverse engineer some information about the dataset you used to train it on, for instance:

• Membership Inference: Identify whether a person was in the dataset.

• Model Inversal: Construct the original dataset (entirely or partially).

• Statistical Inferences: Estimate feature averages, distributions, etc.

These three are only some examples of what attackers might want to do with your model. Again, this is especially true for sensitive data (people info, credit card info, medical records, etc.)

You can use AI Privacy techniques to make your model more robust to these attacks. I recommend studying the following topics to start:

• Data Cleaning: If you can, remove or encrypt sensitive data.

• API Hardening: Making it hard for attackers to make a bunch of inferences with your model.

• Differential Privacy: Change your data and model to guarantee certain privacy levels, typically in exchange for performance. Try to understand how much privacy you need for your application.

• Federated Learning: A way to train models in a distributed framework without directly sharing data, helping with other data privacy concerns.

4. AutoML

AutoML is the field that tries to automate parts of the Machine Learning pipeline to improve the model training process, making it faster and potentially better.

• Feature Engineering: Creating features is typically an extensive process. Auto Feature Engineering tools try to create and evaluate them for you. (Tooltip: featuretools)

• Pre-processing + model Selection + Hyperparameter Optimization: Instead of manually setting parameters, these techniques provide optimized searching solutions to find the best combination for your problem. (tooltip: auto-sklearn)

5. Causal Inference

Sometimes, instead of focusing on predictions, we might want to choose the right “treatment” in an application. For instance:

• Choosing Model A vs. Model B

• Decide how much discount to give to a customer

• Define a customer credit limit

If you are entirely new at this, I recommend starting with the basics of A/B testing, which should help you design practical experiments.

Afterward, my golden recommendation is the Causal Inference for the Brave and True, which goes in-depth into causality with convenient python examples. You will there learn about estimating the “Conditional Average Treatment Effect“ (CATE) using models (namely, X-learners, S-learners, and T-learners)

6. Fast Training

Especially for those working with larger models that take a long time to train, looking for ways to speed up the training is vital to achieving results more quickly.

For that, we can train a model way quicker through different techniques, such as:

• Learning Rate Finders: For instance, using a Cyclical Learning Rate (CLR) to arrive at convergence with fewer iterations (image below)

• Mixed-Precision: Instead of using full-precision of the GPU operations, reduce it fully or partially (mixed) to half-precision (fp16), making the calculations faster and resulting in less GPU memory usage. Of course, a reduced precision may impact your final model performance.

• Model Compression: Reducing the model, for instance, by cutting down weights (model pruning), can also help with speed. Also, some research shows that larger, sparse models are often better than smaller, dense ones.

7. Unstructured Data

Most people start in the field by tackling structured data (tables). However, dealing with unstructured data can provide some exciting modeling opportunities for you. 

In all of them, you can work with a more classical approach (such as using libraries to extract features from them) or use Deep Learning. For Deep Learning, I recommend sticking to PyTorch or Tensorflow (in doubt? Pick PyTorch). 

Some additional libs I recommend by data type:

• Images: Scikit-image, OpenCV, and TorchVision

• Audio: Librosa, TorchAudio and SpeechBrain

• Text: SpaCy, Sklearn, and HuggingFace Transformers.

• Graphs: PyG and DGL

Finally

While there are many other exciting topics in Machine Learning, these 8 are my top recommendations based on how useful they can be for your projects and careers.

Hopefully, you find some of them interesting or useful! :)

• Services: 5 Types of Freelancing, How to choose one, and Niches

• Pricing: Framework for pricing your work.

• Getting Clients: 3 approaches to get you started

• Creating your Proposal: exclusive proposal deck template.

• Scaling your business: What to do when you have enough work?

Service Types

I divide it into 5 categories:

Read more

There I was, early in my career, working on a freelance machine learning project for a bank. I was posed with the challenge of creating their loan approval system: for a beginner, this project should sound pretty simple, right? It is a classification model that tries to identify which customers are “good” (e.g., lucrative) and which aren’t.

Given that the company had a simple policy for customer application approval based on credit bureau scores and data on which customers paid their loans or defaulted, I had enough data to train an initial model. But then… I started reflecting.

You know, loans can change lives. Sometimes loans are made to pay off an unexpected health expense, other times when someone’s primary income gets compromised, or even to start a business. Therefore, when you deny a loan, you are potentially closing up a life-changing opportunity for them. At the same time, it’s business. You cannot simply approve all loans; otherwise, the bank would rapidly go bankrupt due to default.

Can I use any features? Well, not really. If we have, say, a gender or race feature, the model will directly learn to be more discriminative towards either group, which would ultimately make it benefit certain groups over others.

I tried excluding those features, but when I analyzed the “feature importance” given by my model, I found that the most relevant feature was one related to the person's region. And upon further investigation, I discovered that pretty much all areas with higher black people's presence ended up being denied more.

Why does it matter? You see, the model still favors one group over the other. It might even have data to back that up (let’s say black people have a significantly higher default rate), but this poses mainly three problems:

• i) Group Discrimination: If the other characteristics are the same for a white and a black person, the model could deny the black person just because they live in a black-dominated neighborhood, which is unfair [and it is also a crime in most countries - as it configures racism];

• ii) Feedback Loop Problem: By denying more black people, we end up creating a vicious cycle: we have fewer data about them (because we only know “good” customers after they paid a loan in total), so when retraining, we won’t learn more about them (you can read more about degenerate feedback loops here).

• iii) Opportunity Loop: Similarly, by approving fewer loans to a group, fewer people in that group will be able to change their lives with the loans they sought. This also intensifies the disparity of opportunities, worsening the situation in the long run.

So, I established: I would try not to use features that directly or indirectly identify a protected group (gender, race, age, etc.). But… There are correlations everywhere.

For instance, we could have salary disparity among the groups, and a lower salary could be connected to females, whereas higher wages could be tied to males. If our model learns to reject most low-salary people, we are falling again into the same trap.

Measuring Bias

Even though I identified the “bias problem” above by inspecting my model and thinking about the business domain, the fact that I couldn’t tell “how unfair” my model was started to bother me.

I started with a straightforward way of measuring it: Opportunity Inequality. I used my model to predict the test data and then checked the loan approval rate for each protected group. The difference between the “highest approval rate” and the “lowest approval rate” became this “Opportunity Inequality” measure.

Later in my career, I discovered a better name for this one — statistical parity difference — and other metrics as well, such as:

• Predictive Parity: Difference % of correct predictions for each group

• Statistical Parity: Difference in the rate of favorable outcomes

• Disparate Impact: Privileged Outcome Ratio / Unprivileged Outcome Ratio

• Theil Index: Entropy of Benefit for all individuals

Achieving Fairness

After seeing that my initial model statistical parity was terrible, I started thinking about ways to reduce it.

I started tweaking my features: I conducted a feature selection to get to a model with a good enough statistical parity. However, this process created a pretty inaccurate model, as I was left with few features that would not serve the business purpose.

Then, changing the model: I tried to change the model and play with hyperparameters to achieve something with less variance so it would have less of a problem fixating on the more problematic attributes. Not enough.

Finally, the output: You can see the fairness metrics impacts by tweaking the threshold for each protected group. I tried the two following approaches:

• Demographic Parity: Trying to get the same amount of loans approved in each group. However, this led to a very unprofitable model, as the groups in my case had a very different “rate of paying the loans back.”

• Opportunity Parity: To find an “Equal Opportunity Rate” for each group, I found custom thresholds that maintain the same percentage of approved loans to each group, considering only those that can pay the loan (True Positive Rate). This was the best approach, achieving accuracy, profit, and equality, and it also had the benefit of being very easy to implement and maintain.

Going Beyond

There are other exciting approaches to promoting AI Fairness I learned after this project. I am not going into them super in-depth, but I will provide tools to implement them (searching their name on google should do the trick to dive a bit deeper)

Practical Tips:

• Simple first: Even though terms like “Adversarial Debiasing” and “Meta Fair Classifier” are very appealing, I encourage Data Scientists to experiment first with post-processing techniques. It’s easier to understand and tune, you can use it on any trained model, and it’s easier to maintain - if the AI Fairness Metrics worsen with time, you can tweak the thresholds instead of going through a model retraining process.

• Pre-processing techniques are neat. For instance, Disparate Impact Removal can change the distribution of a feature (say, salary) to become similar among your protected groups, making them indistinguishable. Therefore the model would not be able to favor one group over the other. However, you should also consider that if you tweak the data representations, your model’s interpretability will decrease, especially if you run this process on many features for different protected groups. Only go with them if this is not a concern for your application.

• Identify the Root Cause: Sometimes, your bias comes from your “label.” For instance, if you have a classifier for “good candidate” vs. “bad candidate” based on approved candidates in interviews, you might carry the interviewers' bias. So, if historically, you favor one group over the others, the model will learn that trend too. In this case, even though the techniques presented here can help, it would be better to address the root cause — the bias the interviewers have.

Tools

Alright. No one wants to implement those Fairness algorithms from scratch. I will list my favorite fairness tools here:

• aif360: Developed by IBM, it has the most used AI Fairness metrics and mitigation algorithms, both available in Python and R.

• Responsible AI Toolbox: developed by Microsoft, it has tools to identify, diagnose and mitigate fairness issues.

• Fairlearn: Another project I used (scikit-fairness) merged with them. Their documentation has excellent explanations, and it also has metrics and mitigation techniques.

• Tensorflow Fairness Indicators: If you are developing deep learning models with TensorFlow, this tool can easily integrate to provide your fairness metrics.

• Interpretability tools: such as SHAP-values and LIME, can show which features are most relevant in your models, which can help in identifying potential fairness problems. I’ll dive deeper into interpretability in a future article.

Inscrever-se agora

Conclusion

I had the opportunity to face this challenge early on and then faced it again in different scenarios. Apart from the specific AI Fairness techniques, the most important thing is to remember that all we want is to ensure that the model does not carry undesirable bias.

Many businesses don’t care about this, but if you are a data scientist, I highly recommend you introduce this topic to your peers and measure the fairness of current models of your company to highlight whether there is a possible problem. Even if the businesses don’t mind, many will at least fear a backlash from regulators in their industry or the general public calling the company “discriminatory.”

Nevertheless, the approaches should depend highly on what you apply them to. For instance, if a Face Identification system struggles to identify one race over the other, you should look at the “Prediction Parity.” Instead of using the techniques I presented here, you could come up with something that works for this problem — such as adding more examples of the group with less accuracy to your training data.

In other cases, “fairness” might not make that much sense. Suppose you developed a disease-detecting model that uses the person's gender and age as features. In this case, there is no problem in using those features, as we are not favoring one group over the other; we are simply adding contextual information that improves the model for everyone (the biology of each impacts the disease-likelihood).

Many AI systems can drastically impact people’s lives. Please do your best to create them responsibly.

If you enjoyed it, subscribe. In future articles, I will cover other aspects of “Responsible AI” (such as privacy and security).

• The Beginning (before ChatGPT)

• The First GPT (understand its predecessors)

• Finally, ChatGPT (what are its intricacies?)

• Limitations and Applications (with examples)

• Google, Meta & Microsoft (are they in fear?)

• The New AI Startup Wave (what’s coming next?)

• Will AI replace all jobs? (the future of work)

• Conclusion (written by ChatGPT)

The Beginning

Back in 2017, some researchers at google were studying how to empower machine learning models with the ability to learn context. Their efforts resulted in an incredible paper named “Attention is All You Need,” in which they use this “attention mechanism” to describe the architecture we know as “Transformers.”

In short, this work enabled neural networks to focus on specific parts of the input data while ‘ignoring’ others. Humans tend to filter information to focus on what interests us (and not to be overwhelmed); this is pretty much what this architecture replicates.

Since then, many AI models have been created based on the Transformers architecture, especially in Natural Language Processing (noticeably for text data). All models from here on mentioned are also based on Transformers (including ChatGPT).

The First GPT

As you can see in the image above, it was a long ride. Before 2018, most models were trained for specific tasks (such as question answering, sentiment classification, etc.).

This changed with the “first GPT” introduced by OpenAI, which proposed a Generative Pre-training Transformer (that’s why “GPT”) process: the model should first learn from “any text” (without a task in mind), to only then get fine-tuned to a specific task.

Now that a model could learn by using “unlabeled data,” we could feed massive datasets during pre-training. This approach resulted in the model needing fewer examples during the fine-tuning step to learn a specific task than predecessors.

Finally, ChatGPT

Even though many other models were introduced after the first GPT, most of the basis remained the same. However, in each iteration, they started to increase the model architecture (which makes it harder to train — more examples, more costly)

• GPT-1 had 117 million params, trained on ~5GB dataset, est. U$500 training

• GPT-2 had 1.2 billion params, trained on ~40GB dataset est. U$43k training

• GPT-3 had 175 billion params, trained on ~45TB dataset, est. U$4.6M training

GPT-3 is the latest “family” of models from OpenAI, and ChatGPT is simply a variant of this GPT-3 fine-tuned on the “dialogue” task (chatting).

Ta-dah! This is the history of ChatGPT. Hopefully, it is clear that, even though the media is blowing up with news about ChatGPT, it resulted from a continual effort to learn how to train larger models more effectively.

All these big models are also known as “Large Language Models” (LLMs).

But hold on! I will get to its applications and limitations.

Limitations and Applications

I started using ChatGPT when it launched because I already had an invite (I previously used GPT-3 and DALL-E). My main tip is to provide specific prompts to receive unique answers. Otherwise, results can look generic.

For instance, instead of “Ideas of Data Science articles,” scoping it down to “Ideas of Data Science Articles for the General Public, focusing on tools, in a funny-style.” provides less-generic results.

Limitations

• Frozen in Time: After the model is trained, it can only refer to the knowledge obtained with the training dataset. Therefore, new information that can appear (e.g., news on the internet) will be alien to it. So it does not provide some utility that other more basic virtual assistants do.

• Dialogue-limited Context: Although you can refer to what was previously mentioned in the chat, it cannot consider other contextual information (previous searches, location, response time, etc.)

• Correctness: Because it’s a language model, it generates text based on the probability of word co-occurrences given the prompts. In the end, it is not trying to provide the correct info, so it would instead give a “convincing but incorrect” answer over a not convincing but accurate answer. This is why correction like the one below happens.

  

My favorite Applications

• Brainstorming: ChatGPT can apply general frameworks to provide insightful examples if you get specific about what you want. (e.g., list ideas for a youtube video title regarding X)

• Summarization: Either to summarize a topic or a concept, ChatGPT can give good answers to things such as “Explain Quantum Physics like I am 5 years old”. Still, you should be careful as sometimes the explanations are incorrect (as explained above) — double-check them.

• Boilerplate: It can provide an excellent boilerplate code. Generally, it still has many issues spitting out incomplete or incorrect code, so I would not recommend a “copy and paste” approach. But it can save time by giving you the overall structure.

Google, Meta & Microsoft

Let me correct the public perception that only OpenAI is behind this and that every other company is lost. First, much of the research basis comes from other AI Labs, and secondly, they have even released similar products:

• LaMDA by Google

• Blenderbot by Meta

• And Microsoft has a long-term partnership with OpenAI.

Notably, Big Techs are part of this revolution. Google is the one that started the foundation for Transformers back in 2017. Meta is the one that created PyTorch, the framework OpenAI and others use to build these LLMs. So, we shouldn’t disregard their contributions.

At the same time, it is undeniable that ChatGPT's strategy worked to grab a ton of media attention. In contrast, LaMBDA and Blenderbot were considered “more boring” when released, which is possibly correlated with those companies trying to be conservative when publishing a new AI that people could use in malicious ways (Google and Meta already have sort of image issues regarding data privacy and so on, so they take this seriously)

If we look at ChatGPT, when it launched, it was straightforward to try to “list potential websites to exploit” and “write a code to exploit a vulnerability, X.” Now they are already correcting this issue, trying to avoid some of those malicious prompts.

ChatGPT made everyone realize that this tech is valuable and should be part of our future. I am excited to see what these companies will release this year, and I will cover each of them as they release in this Newsletter.

Inscrever-se agora

The New AI Startup Wave

OpenAI launched the “OpenAI Startup Fund,” in which they have allocated $100M to invest in AI-based startups, supporting them with their released tech. So far, they have invested in the following companies:

• Descript: Using AI for video editing — as simple as editing a text document.

• Harvey: AI for Legal Workflows (research, drafting, analysis, comms, etc.)

• Mem: Self-organizing workspace, connecting notes, meetings, and more.

• Speak: AI tutor to teach languages (pronunciation, grammar, vocabulary, etc.)

Of course, many other AI-focused startups are out there, and I expect many more to surge this year and the following.

Even for companies not focused on AI, I can see many pivoting to make it a focus, and it will look like a race of “who will be the AI leader in each field?” (e.g., video editing, music generation, art, etc.)

Will AI replace all jobs?

As I’ve demonstrated in this article, AI still has a long way to go to provide near-perfect output that would justify replacing someone. ChatGPT is heavily dependent on prompt design (how you formulate the question) and currently can only output text, not actions (to run a generated code, for instance, you would need to set the environment up yourself). Therefore, it is far away from being an end-to-end tool.

Nevertheless, some of those issues will be solved with time. Will it then replace our jobs? I don’t think so. But it can for sure change the way we work. For instance, most straightforward parts could be automatically generated instead of developers writing code from scratch. Then the developer would only need to tune the more complex functions. Another example would be a financial analyst using these technologies to create code that automates part of his workflow.

The possibilities are endless, but I believe in a world where it will enhance our capabilities, not make them useless.

Below is my response to: “could ChatGPT and Midjourney replace the average UI/UX designer?”

And I think that is true for most professionals. We will replace less creative and repetitive tasks with AI capabilities, whether writing boilerplate code, scheduling a haircut, or driving a car.

Conclusion

If you want to play with ChatGPT, you can access chat.openai.com to create an account. The free version works very well, and the paid includes better availability, faster responses, and access to new features first. I would only recommend subscribing after using the free version for a while and identifying it as a vital part of your workflow.

Finally, I asked ChatGPT to write my outro:

"In conclusion, ChatGPT is a technological marvel that will have you asking questions left and right, just to see what clever responses it will come up with next. So go ahead, strike up a conversation with ChatGPT and watch as it wows you with its wit and wisdom. Who needs human friends when you have a language model, right? 🤣 And if you want to stay up to date with the latest developments in AI and other tech advancements, don't forget to subscribe to The Overfit Newsletter! 📧"

Ensemble is a monthly issue from The Overfit, aggregating the most relevant updates in the Data Science & AI industry for the given month, always with some personal takes on “why is this happening?” and “what will happen next?”

Layoffs

Last year, the tech industry had some notable layoffs. Many people thought the layoff waves would stop already, but January came hard with many big layoffs:

Why are layoffs happening? My personal view:

• Tough past year: 2022 was the worst year for the U.S. stock market since the 2008 crisis. You can see in the image below a visualization of 2022 market changes. Except for the energy and healthcare sectors, every other industry took a big hit.

  

• Uncertainty: Historically, high inflation + FED increasing interest rates result in recessions, which is why many economists believe one will happen in 2023. With that, companies and investors try to be more conservative, which makes getting funding more challenging (for growth projects — both in a startup and in larger companies), reducing growth expectations for the following year. On top of all that, we have political issues that we cannot predict when they will end, such as the Ukraine-Russia conflict.

• Predictions: Given the tough past year, uncertainty, and less accessible funding, companies are compelled to make conservative estimations for 2023 by their shareholders. This requires companies to adjust their headcounts to maintain desired profit-to-cost / profit-to-staff ratios, avoiding a more significant impact in the case of an economic downturn, which finally results in layoffs, hiring freezes, or at least employee growth reduction.

• Herd mentality: Last but not least, professor Jeffrey Pfeffer attributes the recent layoffs as a “social contagion” (source). As companies see others doing it, they feel compelled to follow. In my view, it combines this “mentality” with everything mentioned above, as the “uncertainty” for the next year has some logic and emotion behind it — behavioral economics in action.

Microsoft Example:

• Why Microsoft is considered odd: Microsoft is one company that surprised many with its layoff. The company is currently at an all-time profit high, and its revenue distribution is better than other big tech companies (such as Meta, with 97% of revenue coming from ads), with products such as Azure (31.3%), Office (23.7%), Windows (13.8%), etc.

• Priority Hypothesis: In the layoff e-mail, Satya Nadella (Microsoft CEO) said they would “continue to hire in key strategic areas.” This indicates that they might shift priorities during 2023, shutting down some initiatives that will no longer require employees while focusing on others.

• Overestimation Hypothesis: Let’s look at employee growth at Microsoft: 9.92% in 2019, 13.19% in 2020, 11.04% in 2021, and 22.1% in 2022 (note the spike). Whereas, in gross profit growth, Microsoft had: 16.89% in 2019, 19.52% in 2020, and 17.06% in 2021. This indicates that they expected more growth during 2022 and decided to make more conservative estimates for the next year, given the market uncertainty and potential business priority changes, as mentioned before.

(even though there might be a priority change, the overestimation hypothesis is less farfetched for me, given the solidity of Microsoft’s business — closing one small project or another would not make such a significant difference in HC)

Despite all this, out of the laid-off people, "72 percent have found new jobs within three months. Even more surprising, a little over half of them have landed roles that pay more than what they were earning in the jobs they lost" (source)

Will this continue? I expect layoffs to continue, especially during February (as it matches the budgeting cycle for some companies). We will likely only recover from this scenario after the market uncertainty declines significantly and companies can get more confident in their growth.

ChatGPT

What is it? ChatGPT is a large language model that is trained to generate human-like text. It can be used for a variety of natural language processing tasks such as language translation, text summarization, and question answering. It is trained using a variant of the GPT (Generative Pre-training Transformer) architecture and is fine-tuned for specific tasks using additional training data.

(ChatGPT generated the above description)

This technology, released by OpenAI, caught a lot of attention — one million users within a week of its launch — because it provides excellent responses to many prompts, and it made the public pressure Google to show advances to their search engine in similar ways.

Still, there are many limitations to ChatGPT. For instance, it will agree when given a misleading prompt:

In the example below, I asked it to generate code to monitor concept drift, and it generated some code that “makes sense”; However, i) it does not follow “best practices” for detecting concept drift; ii) some critical parts are left out, such as the “handle_drift” function; iii) It is not functional, as the threshold variable was never declared (is it global?).

Still, with more guidance and some tweaks, it can generate useful code (at least to serve as a quick start), especially for more straightforward, well-defined tasks.

Finally, OpenAI announced the transition from 100% free to a freemium business model, in which you pay $42 a month for higher availability, response speed, and access to new features.

I will continue to use ChapGPT on some applications (the free version, most likely), and I am excited about the subsequent releases from OpenAI and possible competitors. This is a promising year for AI-based applications, especially in Natural Language Processing and Computer Vision.

In a future post, I will go through the details of how ChatGPT works, its predecessors, and expectations for the next versions. Click below if you want to read that and did not subscribe yet! :)

Inscrever-se agora

Research Picks

• CICERO: “An agent that can play at the level of humans in a game as strategically complex as Diplomacy is a true breakthrough for cooperative AI.” (Meta AI)

• Teaching Algorithmic Reasoning via In-context Learning (arxiv) [actually from last year]

• Deciphering Clinical Abbreviations with Privacy Protecting ML (Google AI)

• Instruct Pix2Pix: Learning to follow image editing instructions (git / arxiv)

Event

There is an exciting event on January 26th (tomorrow if you received when I sent this issue) with the following topics:

• Data-centric AI (a chat with Andrew Ng)

• Operationalizing Machine Learning at a Large Financial Institution

• Large Language Models (LLMs) to Production

• Get Your Company Started with AI: The Right Way

That’s our wrap-up for January 2023!

Let’s assume you are training a Machine Learning model for a bank to decide whether each customer is “a good customer” (default is unlikely) or “a bad customer” (default is likely). We can call this our “approval model.”

You used data from the company's customers and finally created a model with reasonable precision. By your calculations, it should be highly profitable (as estimated by the customer’s default reduction). Therefore, the “approval model” goes to production, and now every customer will either be approved or denied following the model prediction.

After a couple of months, you have gathered more data (new customers and more default info about old customers) and decided to retrain the model, trying to increase the model’s precision. And so you did: your retrained model and achieved a higher precision.

You might think: Well, it’s time to go to production by replacing the old model, right? However, there is a hidden problem that goes by many professionals.

Look at the diagram that represents this flow:

The problem here is: Because we cannot generate the “label” (whether a customer defaulted or not) for denied customers, we generate selection bias in our model. This poses two issues, mainly:

• There might be many good customers in the “denied” population, but our model cannot identify them because we never approve them.

• The problem aggravates with every retraining — because within the approved population, the only mistake the model makes is to “approve a ‘bad’ customer.” Hence, the incentive is to create a more strict model every time.

Hopefully, this exemplified well what a degenerate feedback loop is, but here I follow with an excellent quote:

“A degenerate feedback loop is created when a system’s outputs are used to create or process the same system’s inputs, which, in turn, influence the system’s future outputs” — by Chip Huyen

Here are some more examples of the Degenerate Feedback Loop:

• Fraud / Crime Analysis: Model that sends to the analysis suspicious financial behaviors, and we don’t get the info about the “not suspicious.”

• Credit Lines: Model that decides how much credit we should give a customer. We don’t have the information on what would happen with customers if they received a lower or higher line than recommended.

• Search Engines / Recommendations: In recommender systems, if you present the user with a list, there is a natural “position bias,” in which users typically already click the first results. (we don’t get the info about what would happen if the order was different)

• Geographical Predictions: If a model predicts crime-rate for different locations, and we use that to send more cops to certain places, we might get more crime reports in a region just because there are more cops there. We get the data about areas with no/few cops.

Solution #1: Sampling

Well, if we do not have the data for part of the population, we can create ways of gathering it. The most common approach is to use a random sampling strategy.

Using the example above of “customers approval/denial,” we can set a small percentage of customers to approve regardless of the model's prediction. This can also be useful to compare the performance of the model in comparison to a randomized approach.

Mathematically, this is a very sound solution; however, when applying this logic in real businesses, we must be extra careful. In this example, significant losses might occur if we bring too many “risky” customers. Therefore, we should try to set a “reasonable” sampling ratio considering our business requirements and, if necessary, apply other policies to mitigate the risk you are creating. As a Data Scientist, I incentivize you to advocate for this process in your team, as it will bring long-term benefits.

More tips…

• Weighting: When sampling and retraining your model, you should consider your sample's representativeness. Depending on the modeling technique you are using, it might be a good idea to assign greater weights or oversample/undersample to balance out the populations to reflect the production distribution.

• Reinforcement Learning: You can increment the sampling strategy with more advanced techniques, such as the ones based on Reinforcement Learning, as this can easily be framed as an “exploration-exploitation” process. Still, if the randomized approach is not that costly for your application, getting that extra data with the “randomness” property should be pretty valuable.

Inscrever-se agora

Solution #2: Counterfactual

We can try another approach when we have a feedback loop that only affects some of the outputs.

Let’s assume you have a model that predicts whether the person would click on that recommendation. Here we have the feedback loop problem because people naturally click more on the first options (known as position preference).

A straightforward approach to overcome this is that you may gather the data of your model (features and label) and create an additional feature to each instance indicating if your label (click or no-click) happened when it was the first in the order or not. After training our model on that data, we can, for a given instance, force the flag 0 (not first). This way, we remove some of the bias of the cases always in the first position. See this flow below:

This is known as “The Counterfactual Model,” as we want to know what would have happened if a different course of action had been taken (if our prediction was not for a “first position” item, would it still be this relevant?)

Of course, this process can apply to other scenarios. Credit limit policies can benefit significantly from this, as we may want to know, “if we had a different credit limit, would this person still not default?”

If you want to dive deeper into how a mature company like Google does it, you can watch this video, which explains their study “Recommending What Video to Watch Next: A Multitask Ranking System.”

Conclusion

Despite following the modeling methodology step by step, data scientists can easily fall into the degenerate feedback loop trap, which will most likely generate long-term problems. However, by considering the entire application flow early on, we can anticipate it and implement the proper mitigation solution.

I hope you found this helpful. Good luck with your feedback loops!

Inscrever-se agora

• How to properly define goals;

• Goal-setting for your personal and professional life;

• Goal-setting for Data Science & AI teams;

• What to do with the final results.

Let’s go!

Defining Goals

SMART is a framework that defines a series of properties a goal should have. I will use a personal example as I go through them. Let’s say you are considering “improve my health” as a goal this year. We should transform this goal into something more:

• Specific: Instead of “improve my health,” narrow it down to one component of your health. For instance, you might want to be “less sedentary.”, which is more specific.

• Measurable: “Less sedentary” is hard to measure. We should add something that would let us check our progress. We could change the goal to “exercise 300 days this year”.

• Achievable: Exercising 300 days a year might not be feasible for your current schedule. It is okay to be ambitious, but it is not okay to be ridiculous. We could tune it to something more achievable like “Exercise 3 times per week”.

• Relevant: Is this goal aligned with your values and long-term objectives? If you want to become a more fit person long-term, this goal makes sense!

• Timely: With the “Exercise 3 times per week”, it is unclear how long we should sustain it. Set an end date, such as “Exercise at least 3 times per week throughout 2023” or “Exercise 156 times throughout 2023.”

Comment: I prefer “156 times throughout 2023” over “3 times per week throughout 2023” because i) unexpected events might happen, and if you struggle in a week you could still compensate in others; ii) intermediate progress is more accurately tracked if you count by single days rather than "weeks that you did at least 3 times” — helping with the measurable property as well.

Finally, after defining the goal, develop a strategy for achieving it. In the case of “Exercise 156 times throughout 2023”, you already know that you will need to exercise, on average, 3 times per week. So, you could time-block the hours in your agenda (or schedule sessions with a personal trainer). If you have any upcoming trips, decide whether you will compensate in other weeks or how/where you will exercise during your trip.

Goal Setting for You

I classify each personal goal as either a “habit-defining goal” or a “project goal.” In the example earlier, “Exercise 156 times throughout 2023”, I would call it a “Habit-defining goal” because it focuses more on trying to keep a practice going throughout the year than on completing something.

Opposed to that, a “project goal” could be “Publishing my Book by December 2023”. It still respects all SMART goals (some might say it is not measurable, but it is, with “published” or “not published” criteria), but it focuses more on the Output (results) rather than the Input (effort).

For project goals, it is crucial to set intermediate checkpoints to create a roadmap:

1. By the end of January - Create a draft for each chapter;

2. Write one chapter per month, finishing the whole book by September

3. By the end of October - Finish the book review and layout;

4. By the end of November - Finalize details for publication and promotion;

5. By the end of December - Have it published on Amazon KDP;

Of course, you could keep breaking down (from year to months to weeks to days to hours), but try to stop at a level that gives you enough clarity to take action. In this example, I would do a quick planning session each month to schedule the activities I need to complete that month.

Organizing your personal goals

First, I separate my goals into 5 areas:

• Health: Exercises, Diet, Mental Health, and so on;

• Intellectual: Books, Learning Projects, Languages, etc.;

• Professional: Career Growth, Side Projects, and more;

• Finances: Savings, Expenses, and Investments;

• Other: Leisure, social, or anything I cannot fit into the above.

Start with the longer-term goals for each category because you can check whether everything is cohesive when defining the shorter-term goals.

For instance, you might have a long-term goal such as “Retire with 1 million dollars by 50 years old”. Suppose you are 40 years old and currently have $500k. In that case, you can then calculate how much money you need to save per year and then either create financial goals like “Save X this year” or professional goals related to creating more revenue streams (“publish a paid mobile app”) or increasing current ones (“get promoted in my job”).

Also, it’s okay to disregard this connection between short-term and long-term goals because sometimes we want to do something just because it is fun. It is more important to check whether a new goal will impede others. For example, if you have two goals, “visit 10 countries this year” and “save X money this year,” you need to understand if all the trips might get in the way of your saving money goal.

(Example) My personal goals

Here are my goals for 2023 as an example:

Goal-setting for Data Science & AI teams

How to Organize Goals - OKRs

Instead of setting a bunch of goals in a list, the most common approach in the industry is to use a framework named OKR (Objectives and Key Results). It is pretty easy to apply:

• Start with an Objective (e.g., I want to improve our fraud detection solution)

• Define Key Results, typically 3 to 5, following SMART principles. Examples:

  • Improve by 50% the solution financial return (from 100k to 150k/month)

  • Run a modeling experiment data from a new vendor

  • Implement an automatic retraining pipeline

  • Reduce scoring time from 1s to 0.5s

• Track your progress. This could be done via periodic checkpoints.

  • How much did your team achieve already?

  • Do we need to change something? (priority, strategy)

(Example Goals) When first creating a model…

My suggestion is to start with a well-defined business goal. Say a team is developing a machine learning model to detect fraudulent credit card transactions to block the suspicious ones automatically. In this case, we should start with “what is the minimum acceptable performance for a solution that still brings value to the company?”.

Let’s dive a little deeper into this example. We need to understand the value of each thing.

• True Positives (TP) - Correctly Identified Frauds: This would stop some frauds from happening, which gives mainly two benefits: i) negates losses the company would have to pay for; ii) prevents attrition with a customer that would have an awful experience.

• False Positives (FP) - Incorrectly Appointed Frauds: Blocking a credit card incorrectly could generate customer attrition, impacting satisfaction and reflecting how long they would stay as customers. We could go further and, through experiments/studies, understand how much the churn rate of customers increases when they have incorrectly blocked transactions.

With that, one recommended approach is to find a minimum “True Positive to False Positive” ratio (also known as “precision”) by estimating the actual value of a true positive and a false positive.

One simplistic way to do this is to get the “average savings” in identifying fraud and the “average losses” by incorrectly blocking a transaction.

After some studies, we end up with TPs and FPs worth $4000 and -$100, respectively. With this information, we can easily estimate the expected financial return for a given model and threshold.

Finally, it would not make sense for most businesses to implement a fraud detection system that would bring only $10 in revenue. We should define how much money the solution should bring to the table to be considered an option. For our hypothetical business, it could be “at least” $100,000 positive return per month.

In this example, we have a specific way of measuring the effectiveness of a given model, and we also have a minimum performance that we need to achieve ($100k), so we could have a general goal as:

Fraud Goal Example: Design and Deploy an AI-enabled solution to automatically identify fraud transactions, bringing at least $100k positive return considering a $4000 value for each identified fraud, and -$100 value for each incorrectly identified fraud, by the end of this quarter.

This would be a worthy goal for the team as it clarifies what they should work towards, the requirements, how to measure it, and the deadline.

(Example Goals) For Mature Solutions

All right! Now we have successfully designed and deployed the fraud detection solution. It might be an excellent moment to create goals for the next quarter so we can improve them and bring more value.

We still should look at business requirements to orient “how much improvement is enough to make an impact.” Sometimes, the improvement expectations are way higher than what is possible to do with currently available data. In those cases, we can revisit this project in other cycles or focus on further actions to make more/better data available.

If we already have an excellent model, it might be more complicated to improve than a newer model created using just a few features. For more mature solutions, instead of using “improve the model by X%,” it would be more interesting to focus on “process goals” instead, such as:

• Create and validate 3 hypotheses for improving model accuracy;

• Reduce the time we take to validate a new feature by X%;

• Experiment with automatic retraining strategies and publish results;

This is my experience; I know not all managers agree, but I feel it’s better to push improving the process that will ultimately help with the end goal than trying to set an arbitrary number of “push the recall rate from 98% to 99%”, considering all the uncertainty that can be behind machine learning projects, which could bring unnecessary anxiety to the team.

Other Practical Tips for Team Goal-Setting:

• Single-metrics: Avoid using easy-to-fool metrics, such as precision, by themselves. Instead of “Achieve 90% precision”, prefer “Achieve 90% maintaining our current recall of 85%”.

• Value the process: Being result-oriented is excellent, but data science is a field that requires testing and validation. Naturally, the hypothesis might come out false in the end. Refrain from only valuing results and consider the process as a whole. (e.g., creating key results for completing experiments)

• Involve your team: Engaging everyone in creating a goal is crucial to improve the sense of belonging but also helps make goals more achievable.

• Frequency: Most companies create their OKRs quarterly, biannually, or yearly. Generally, high-uncertainty companies (such as startups) might benefit from shorter OKR cycles, as things can change rapidly, whereas more mature companies may prefer extended periods.

What to do with the final results?

One mistake many companies make is to assign the team goals results as the performance of the team members. I’m not too fond of this approach, especially for data science & AI teams, because not achieving a given goal might result from the problem being too hard (and we might not have known it beforehand). This creates a problem of people trying to move to teams with “easier OKRs.”

At the same time, we should not completely disregard it. If a team had 5 goals and had poor results in all of them, this could be a strong indicator that the team is underperforming or overestimating tremendously. Either way, you got a problem.

My recommendation is to handle performance reviews separately from OKRs. However, considering their current level, we should try to see whether the person was providing their best effort towards the team OKRs. Therefore, the OKRs would impact the performance indirectly, making this a less complicated process.

Yes. It is also very context-specific. In other words, what can be achieved in a specified period will vary drastically by company and factors such as data availability, infrastructure, team seniority, team size, etc. But experience will be valuable towards better goal-setting after all.

The important thing is that you finish the OKR cycle by reviewing them with your team, listing what you have learned trying to make it better in the next cycle.

Conclusion

Even though data-science projects are filled with uncertainty, defining goals is possible and brings many benefits:

• Determines a clear direction for the team (people can understand whether what they are doing contributes to the team goals and adjust)

• Promotes a shared sense of responsibility when done correctly.

• Enforces a data-oriented culture (we don’t manage what we don’t measure)

• Improves project syncing, as we can see all goals and main deadlines.

“If one does not know to which port one is sailing, no wind is favorable.” — Seneca, stoic philosopher

That should be enough to justify implementing them.

Good luck with your goals! Consider subscribing if you found it helpful!

Inscrever-se agora

In this post, I'll list the most important books I read in my Data Science career while giving you more context that should be helpful for you to decide whether to read it.

For Beginners

A beginner, in my definition, is someone who either wants to start in this career or just started and still needs to work on their data science & AI knowledge breadth and depth.

Practical Statistics for Data Scientists [Buy it Here]

This book from O’Reilly is a great starting point in Machine Learning because it covers the essential techniques and concepts in the field. Don’t be fooled by its name; it also teaches other techniques that are not considered “statistical” by some (such as decision-tree-based models and unsupervised learning techniques)

If you decide to buy this, you might get overwhelmed with so many topics to learn. I recommend the following approach:

1. Read and practice the “Exploratory Data Analysis” (EDA) part, in which you will learn how to perform EDA and the core data cleaning and preprocessing techniques for ML, as well as what the entire project cycle looks like.

2. Proceed to the Machine Learning models. There are many in the book, and I suggest picking one or two: the ones I find most important are Linear Regression and Logistic Regression, along with Decision Tree-based models. Also, I recommend that you learn the basic preprocessing techniques.

3. The clustering part (chapter 9) focuses on K-Means and DBSCAN. Learning those will increase your knowledge breadth.

4. Now that you have completed the essential topics, I would study different chapters depending on what you are most interested in, so you can fill some knowledge gaps and start building your “expertise” in something.

(extra tip) Please, focus on understanding how the algorithms work and not only how to apply them, as it is essential for decision-making and interviews.

Deep Learning with PyTorch [Buy it Here]

I used to recommend learning Deep Learning with Keras or Tensorflow. Even though it is still viable (and Keras is generally known for being more straightforward than other frameworks), nowadays, I strongly recommend learning PyTorch.

This is because, in recent years, PyTorch has been used increasingly (see figure below), making it easier to find implementations of the newest architectures in PyTorch. Additionally, related frameworks (such as fast.ai) can be used along PyTorch and lets you quickly implement many techniques.

Nonetheless, I only recommend this book if you are determined to work with Deep Learning primarily; most “Data Scientists” I know to focus on other machine learning techniques targeting structured data.

After finishing this book, to further dive into deep learning, I recommend studying through research papers and Kaggle. It would be best to specialize in a topic within Deep Learning; Computer Vision and Natural Language Processing are the two most common and recommended choices.

Industry

To those looking for knowledge to apply to ML/DS projects.

Machine Learning Systems Design [Buy It Here]

This book by Chip Huyen is incredible. As someone that did not focus so heavily on productizing machine learning models (my focus initially was only on modeling and experimentation), reading this book opened my eye to techniques that, before, I would not have had much clarity on how they worked under the hood and helped me in my transition to managing Machine Learning Engineers along with Data Scientists.

Quick tips:

• This book assumes you already have a basic understanding of ML, which should be covered by the books I recommended for beginners.

• For more senior professionals, it might be worth reading by jumping directly to the topics that interest you the most. That’s how I did it, focusing on issues currently being discussed in the company.

• This a great option if you want to become a Machine Learning Engineer after studying the basics.

Machine Learning Yearning [Available Here]

This free book by Andrew Ng covers how to handle different situations that happen in Machine Learning projects: what to do when you struggle to create a good model? Which metrics should I use to guide my team? And so on.

Quick notes:

• This is not technically heavy, and it focuses on giving broader pieces of advice. So do not expect an in-depth explanation of techniques.

• In my experience, even non-technical managers (that sometimes manage or interact with data scientists / MLEs) find the concepts presented by the book easy to understand.

Deep Diving

If you are already comfortable with the basics, it’s time to specialize further in a topic. Here, I separated two books on hot topics that could serve that purpose.

Deep Learning Book [Buy it Here]

Written by top names in the field, this book is an excellent dive into deep learning, focusing primarily on understanding the techniques. Apart from the Amazon version, there is also a free online version.

Tips:

• It has a refresher on Linear Algebra. Even though it is not 100% necessary to understand the remaining of the book, I recommend studying it so you can gain more in-depth knowledge from this book.

• This book does not contain recent advancements in deep learning (such as transformers, stable diffusion models, etc.), but it surely builds the foundation for you to understand those.

Causal Inference [Buy it Here]

Causal Inference is a topic on the rise in Data Science and AI because the estimation of impacts and effects of changes appear in many practical situations, such as in defining pricing strategies.

Quick tips:

• The author also has a blog version of this book available for free here, a good option if you are in doubt about whether to buy the O’Reilly version.

• Mathematical notation is used, but the author does an excellent job explaining it. Anyway, it is not a book to speed read.

Managers

The Machine Learning Yearning book I mentioned above is helpful for managers with specific advice for ML projects. The following books will not be specifically about ML but apply to managers in this field.

The Making of a Manager [Buy it Here]

Julie Zhuo (ex-VP at Facebook) brings many of her experiences that help readers understand the path to becoming a senior manager.

• “Your job, as a manager, is to get better outcomes from a group of people working together.” This good quote summarizes this book's focus: people management.

An Elegant Puzzle [Buy it Here]

This book focus on “Engineering Management,” which is closely related to what one does as a Data Science / MLE Manager. It focuses on providing the correct systems to handle different topics, making this book the most practical I’ve read on technical management.

Tips:

• It covers many novel topics for new managers, such as Hiring, Influencing through Culture, and Team Design.

• The topics presented are independent, so you can read them in any order.

Classics

To the ones looking for more in-depth resources.

Artificial Intelligence: A Modern Approach [Buy it Here]

Despite having a Machine Learning section, this book comprises Artificial Intelligence (broader than ML). Please look at the content table to ensure the topics interest you.

Tips:

• In my experience, some people that arrive in data science with non-computer science degrees typically have zero to little knowledge about many of the things presented, such as Search Algorithms (BFS/DFS/etc.) and Constraint Satisfaction Problems, which are very useful in many of the problem settings people face in this field.

• Another good section of this book is Reinforcement Learning (RL). It provides an excellent introduction to this topic. I recommend Sutton's classic book that focuses on RL if you love it.

Pattern Recognition and Machine Learning [Buy it Here]

Typically used as a textbook in Machine Learning classes, this book gives a more in-depth explanation of the algorithms with mathematical rigor you don’t find in most online courses.

Tips:

• Generally speaking, studying this book gives a more solid foundation than my beginner recommendations. However, it is way more prolonged, and this depth might be overkill for someone just starting, as most jobs will not require a super in-depth knowledge of each algorithm. In the end, you choose the practical approach or the academic approach.

• I recommend going through this book linearly, as many of the sections reference past sections, and reuse the same notations throughout the book.

Conclusion

The Data Science & AI field has much to study, which makes it both challenging and intellectually rewarding. Remember that it is acceptable only to know a few topics in-depth and try to focus on continual improvement instead.

At The Overfit, I bring the most important topics to foster this continual improvement for anyone building a career in this field, filtering through all the noise and presenting them in a palatable way. Consider subscribing if this appeals to you. :)

I hope you find some of those books interesting! Good reading!