?? My ceiling isn’t the best, but it’s up there

it's been a mixed bag of a week. The weather is great and I made fantastic progress, but the state of the world is still shocking. Anyways, let's look at some machine learning.

The Latest Fashion

• Pen and Paper Exercises in Machine Learning just dropped. It’s fairly basic with a?focus on math instead of applied ML?or models, but insightful in that regard and great for preparation for technical interviews.
• Google Brain have published a paper on their?reinforcement learning efforts to aide in designing chips?for the latest generation of TPU chips.
• Still wonder what these transformers are all about? Here’s an?opinionated guide to transformers, from basics to implementation and theory.

Hot off the Press

I produced a video about legal ways to obtain?$90 research papers entirely for free.

Two blog posts went up recommending?books for data science?and?books for AI ethics.

I wrote a tweet that was seen?by 90,000 people?that showed off the?dirty-cat library. This library encodes high-cardinality and messy categories in data automagically! I shared this library?8 months ago?in this newsletter.

Also, I contributed to this blog post on?the future of hybrid meetings.

This is "Light to the Party". All links and extra content can be found in the full issue from last week. Want the latest in your inbox??Join 555+ other curious minds.

Machine Learning Insights

Last week I asked you, “What is the difference between generative and discriminative models?” and here’s the answer:

The simplest way to think about discriminative models vs generative models is that discriminative models draw a decision boundary within the data and generative models try to learn the data distribution. The generative model basically should learn the system behind a signal that is generated from the data.

In statistical terms, given the data?x?and the labels?y, the discriminative model learns?p(y|x), so the probability of?y?given?x. Whereas, the generative model learns the joint distribution of?p(x,y). That means that the generative model learns the overall distribution of?x?and?y, and the discriminative model learns a conditional distribution. We can also use generative models for a classification task using Bayes’ Rule, but these models are much more versatile then the specific application of discriminative models. The generative model for a classification task is basically used in a sense of “given my assumption of the generative system, which is the most likely class”.

Discriminative models include linear regression models, Random Forests, and SVMs, but also many neural networks that are used for classification (and much of scikit-learn honestly). Generative models include Gaussian Mixture Models, Hidden Markov Models, Naive Bayes, and neural networks like Variational Auto-Encoders (VAEs).

Generative Adversarial Networks, a special combination of two neural networks, utilize?both a generative network and a discriminative network?to create realistic fake data.

Question of the Week

• What is feature engineering and how is it important in machine learning?

Post them on Twitter and Tag me. I'd love to see what you come up with. Then I can include them in the next issue!

This is "Light to the Party". All links and extra content can be found in the full issue from last week. Want the latest in your inbox??Join 555+ other curious minds.

Tidbits from the Web

• Tom Scott overcame his?genuine fear of rollercoasters?in a truly vulnerable video.
• This blind kid makes a?basketball free throw?simply based on sound! I was utterly impressed.
• You may have followed the “Google made a sentient AI”-discussion. Here’s a BS?Bingo for AI Sentience card?for your convenience.

You just read issue #85 of Light To The Party. You can also browse the?full archives?of this newsletter.