The Fast Forward Labs blog covers applied machine learning research, including probabilistic programming, model auditing, text summarization and recommender systems.
Commenters recommend its posts as good overviews, such as text summarization progress and document fingerprinting with steganography, and note proof-of-concept code.
10 thoughts on data visualization best practices and tools:
1) For interactive visualizations of data on 3D globes, I use a mix of C++, Python (for data cleaning), and Unreal Engine (with a plugin called Cesium). An example of this is at https://youtu.be/9i-tQ8Sr80o.
2) If I am trying to put together a 3D globe that has less quality but that can be accessed by the web, I use Mapbox GL JS, D3.js, and React. An example of this is at https://www.whiteowleducation.com/blog/2022/10/14/real-estate-data-exploration/.
3) I have seen others use Three.js for developing 3D data visualizations on the web. An example of this in a data science context is at https://blog.fastforwardlabs.com/2019/04/29/visualizing-active-learning.html.
4) If you are trying to do 3D population density maps in R, there are a lot in the community that say you should use https://www.rayshader.com/ with R.
5) If you are really trying to push the limits of data visualization, follow https://twitter.com/Arti_AR_video . He is doing data vis in AR. Robert Scoble had a good tweet the other day (https://twitter.com/Scobleizer/status/1620498790653501440?) showing Arti with 3D bar charts sitting on a table.
6) If you are doing data vis for urban planning, odds are they are already using ArcGIS, and odds are you will be using something like that.
7) If you are trying to do data vis that relates to architecture, I would actually suggest starting with Twinmotion (which is part of the Unreal Engine ecosystem).
8) If…
The probabilities are fun (and ability to manipulate/mask/logit bias them), but small nit pick: Beam search isn't popular because it results in less human like text and takes more processing time, as a concept it sounds great in theory, but it's generally worked worse in practise compared to the next token with temperature, and results in text that is not human like, the gap is larger with larger models too, it doesn't get better:
https://blog.fastforwardlabs.com/images/2019/05/Screen_Shot_2019_05_08_at_3_06_36_PM-1557342561886.png
The way most GPT models are set up, they don't edit, but there was a lot of work (made redundant by OpenAI) on using gap filling tasks (masking) to iteratively improve or construct or otherwise edit and formulate logic using the encoder-decoder models of yesteryear - in many ways, I think it would have been more exciting if that approach powered the likes of 'think step-by-step' results.
This blog post (from colleagues) covers the ideas in a little more detail http://blog.fastforwardlabs.com/2017/06/23/fingerprinting-documents-with-steganography.html and comes with proof-of-concept code https://github.com/fastforwardlabs/steganos.
This is a good overview: https://blog.fastforwardlabs.com/2018/05/02/progress-in-text-summarization.html