MACHINE LEARNING - AI Dynamics

Creation of a training set for Noise2Music with LaMDA

By

–

09 February 2023 21h41

The researchers created a training set for Noise2Music by using two models to label a collection of 6.8M music source files. They used a large language model (LaMDA in this case) to come up with sentences that describe music in a general way.

→ View original post on X — @aibreakfast

9 February 2023

Future of music creation via text prompts using diffusion models

By

@aibreakfast

–

09 February 2023 21h41

In the not-too-distant future, anyone will be able to create any type of music they want via text prompts.

Here's a look into recent research on diffusion models for generating high quality music audio from text prompts:

(more examples below) pic.twitter.com/ywgJRVoxLx
— AI Breakfast (@AiBreakfast) 9 février 2023

In the not-too-distant future, anyone will be able to create any type of music they want via text prompts. Here's a look into recent research on diffusion models for generating high quality music audio from text prompts: (more examples below)

→ View original post on X — @aibreakfast

9 February 2023

Adversarial Robustness: Fundamental Problem Underpinning AI Safety

By

@thegautamkamath

–

09 February 2023 20h53

I sadly missed @zicokolter
's talk, but I really vibed w/ one of his punchlines he shared w/ me today: adversarial robustness may be a basic and toy problem, but we still haven't solved it. Inability to do this indicates gaps in our knowledge, which underlie more complex settings. https://
x.com/NicolasPaperno
/NicolasPapernot/status/1623324869000667137
…

→ View original post on X — @thegautamkamath

9 February 2023

Coders Using AI as 10x Stack Overflow with RLHF

By

@alexandr_wang

–

09 February 2023 20h29

ok I will now admit that coders are using it everyday as 10x stackoverflow good thing they did all that coding RLHF!

→ View original post on X — @alexandr_wang

9 February 2023

Non-Turing Computation in Living Systems: Long-term Research

By

@rodneyabrooks

–

09 February 2023 19h00

Love this paper; augmenting computation with non-Turing processes; something that most likely occurs in all living systems. This is long term research; don't think about it in terms of immediate applications; it is exploring wild new spaces. @denizzokt

→ View original post on X — @rodneyabrooks

9 February 2023

Predicting Large Model Performance Across Unseen Tasks

By

@_jasonwei

–

09 February 2023 18h56

(3 cont.) Can you predict the performance of one model from another?
Can you predict the performance of a 128B model on an unseen task, given models up to some smaller threshold size and some performances of 128B models on other tasks?

→ View original post on X — @_jasonwei

9 February 2023

BIG-Bench metrics deserve deeper analysis and study

By

@_jasonwei

–

09 February 2023 18h56

(3) Even just looking at BIG-Bench metrics is quite understudied IMO. There are hundreds of tasks in BIG-Bench, and each task has dozens of models evaluated, each with many evaluation metrics. There are task logs for some models. This raises natural questions:

→ View original post on X — @_jasonwei

9 February 2023

Unlocking emergent abilities in smaller language models like Flan-T5

By

@_jasonwei

–

09 February 2023 18h56

(2) One key question of emergence is how to unlock such abilities at smaller scales, and I don't think there has been much work on flan-T5, which often beats PaLM 62B as shown in this paper: https://
arxiv.org/abs/2210.11416

→ View original post on X — @_jasonwei

9 February 2023

MIT Creates Liquid Neural Networks Outperforming Convolutional Networks

By

@mit_csail

–

09 February 2023 18h20

MIT scientists created a new class of speedy “liquid” neural networks that can change underlying algorithms on the fly, sometimes outperforming convolutional neural nets: https://
bit.ly/3K0lLxv Image v/
@QuantaMagazine

→ View original post on X — @mit_csail

9 February 2023