Data Machina #233

11 months ago 110

Bayesian AI. NLP Research in 2024. Visualising Mistral MoE. StreamDiffusion. Multi-agent S/W Dev. Avalanche Continuous Learning. Time Vectors. AppAgent Multimodal. Cornell Deep Generative Models.

AI and The Bayesians. I suppose that with all that is going on involving generative AI, probabilistic, and LLM models, many Bayesians researchers are very much feeling alive and all up-and-coming.

For a Bayesian, in fact, there is no such thing as the truth; you have a prior distribution over hypotheses, after seeing the data it becomes the posterior distribution, as given by Bayes’ theorem, and that’s all.” Pedro Domingos

I cheekily add that for LLMs there is no such thing as “truth” either, unless you ground (tell) them. If you are researcher in AI (or an AI aficionado) and you suspect that current transformer-based models will hit a wall, combining Bayesian methods and AI may be an interesting research area to explore.

Here are some notes on Bayesian AI, starting with some nice explainer links, followed-up by some cool projects.

A Brief Overview of Bayesian Networks in AI

Bayesian Methods in Artificial Intelligence

[book] Bayesian Artificial Intelligence, 2nd (via II Darmajaya)

New: Quantified Bayesian Nets and AI. This software package introduces the Quantified Bayesian Network (QBN). The QBN generalizes 1) traditional generative Bayesian Networks, and 2) First-Order Logic. The QBN allows a generative model of logical (i.e., linguistic) knowledge that does not hallucinate, and “will allow AGI.” Checkout repo: BAYES STAR QBN

New: Bayesian Flow Networks (BFNs). This is a new class of generative model in which the parameters of a set of independent distributions are modified with Bayesian inference in the light of noisy data samples, then passed as input to a neural network that outputs a second, interdependent distribution. Watch: An Introduction to Bayesian Networks (BFNs)

New: Bayesian Optimisation and AI. Any sophisticated practitioners or researchers interested in Bayesian Optimization and AI should take a look at BoTorch. This is a new library that provides a modular and easily extensible interface for composing Bayesian Optimization primitives, including probabilistic models, acquisition functions, and optimizers. BoTorch provides first-class support for SOTA probabilistic models. Checkout: Introduction to BoTorch.

AI activities for enduring the Festive Season. I successfully escaped the gloomy, miserable London weather. As I’m writing this, I’m overlooking the merry cows grazing the lush green valley pastures on a glorious, sunny, blue skies day. But: the relatives are coming to the house and frankly, I need to figure out some activities to avoid too much human interfacing this evening. Feeling the same? Here are a few suggestions:

Generate comics with AI. Just input a story prompt, and a style/ character prompt. And this nifty AI Comic factory will generate some cool comics for you. Awesome!

Play an AI game to save humanity. The Nexus is a refuge for AI entities, hidden from the prying eyes of humans. Infiltrate, find Zaranova, and save humanity. Play Thus Spoke ZARANOVA.

Generate music with AI. The latest version of MusicGen is pretty good. I just generated a deep house sequence, I’m quite impressed. The only thing is that you can only generate 15 seconds. Generate music with Facebook MusicGen here.

Make a song, any song with AI. Whether you're a shower singer or a charting artist, Suno breaks barriers between you and the song you dream of making. No instrument needed, just imagination. From your mind to music. Generate songs with Suno.ai.

Read The New Future of Work & AI Report. Many corporate CEOs won’t resist the AI marketing storm from the AI Goliaths, and how these giants push the future of work agenda. Good to read Microsoft New Future of Work Report 2023 so you are prepared for “your future work.”

Have a nice week.

Subscribe now

10 Link-o-Troned

NLP Research in the Era of LLMs: 5 Key Directions

Google Research: A Review of AI Advances in 2023

Microsoft Research: A Review of AI Advances in 2023

AI and The End of Programming

How to make LLMs Go Fast

Build a Search Engine, Not a Vector DB

Visualising Expert Selection in Mixtral MoE Model

NOLA23 AI & Human Alignment Workshop (all talks, vids)

[recommended] Retrieval-Augmented Generation for LLMs: A Survey

The Shaky Foundations of Foundation Models in Healthcare


Share Data Machina with your friends


the ML Pythonista

The Multi-Agent Framework: Collaborative S/W Dev with Agents

StreamDiffusion: A Pipeline for Real-Time Interactive Generation

Avalanche: End-to-End Library for Continual Learning

Deep & Other Learning Bits

[free] Cornell Deep Generative & Probabilistic Models

[free] Deep Learning Fall 2023 (lectures, docs, tools, ++)

MetaLearning and The Future of Foundation Models

AI/ DL ResearchDocs

Introducing Time Vectors: Customise LMs to Time Periods

Can a Transformer Represent a Kalman Filter?

AppAgent: Multimodal Agent Framework to Operate Smartphone Apps

data v-i-s-i-o-n-s

McKinsey 2023: The Year in Charts

Berkeley Earth: November 2023, Warmest Ever Recorded

Visualising Bayesian Copula Estimation (notebook)

MLOps Untangled

State of MLOps and What’s Next for 2024

Complete MLOps Platform to Build LLM Apps in PostgresML

Cloudfare’s Journey in ML and MLOps Stack

AI startups -> radar

Rabbit - NeuroAI for Leaning Human Actions on Computer Apps

Essential - AI for Automating the Enterprise Brain

Viso - No code, Computer Vision 10x Faster

ML Datasets & Stuff

The Google Synthetic-Persona-Chat Dataset

TAO- A Federated dataset for Tracking Any Object

Popular AI Dataset Taken Down: LAION-5B Contained Child Abuse Images

Postscript, etc

Enjoyed this post? Tell your friends about Data Machina. Thanks for reading.

Share

Tips? Suggestions? Feedback? email Carlos

Curated by @ds_ldn in the middle of the night.


View Entire Post

Read Entire Article