iiinova — where your digital self keeps evolving

Christos Plachouras @cplachouras.bsky.social · 19/08/2026

We launched a way to search for sounds on Freesound with your voice 🎤✨ - just imitate what you're thinking of! It's free and open-source; check it out: thatsoundslike.me . It uses a tiny AI model we made that runs on your browser, not a server 😉

011

Reposted by Christos Plachouras

Dan Stowell @danstowell.mastodon.social.ap.brid.gy · 13/06/2026

The first trillionaire is such a sad moment in the history of humanity - that after so many eras, so many stories, institutions, dreams, our species has arrived at such an *idiotic* mis-allocation of resources. Our entire species. Everything we have. This is where we are.

2710

Christos Plachouras @cplachouras.bsky.social · 02/07/2025

If you are at #IJCNN in Rome, see you tomorrow (Thu July 3) at 4:30pm in the Position Papers session, Leonardo da Vinci hall!

010

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

More in the paper and code: github.com/chrispla/syn.... Big thanks to my collaborators @juj-guinot.bsky.social, George Fazekas, Elio Quinton, @emmanouilb.bsky.social, and Johan Pauwels! I’ll be at IJCNN 2025 in Rome in a month to present this - see you there!

010

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

We argue that downstream task evaluation cannot easily uncover these behaviors, and that equivariance, invariance, and disentanglement are critical components that enable a variety of real-world applications like retrieval, generation, and style transfer.

100

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

Additionally, model and data scale has different effects for different models on the equivariance, invariance, and disentanglement of concepts in their representations!

100

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

Importantly, different pretraining paradigms and architectures exhibit different behavior in terms of these aspects, suggesting the underlying mechanisms for their downstream performance are different.

100

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

However, we show that representations differ in other important aspects, such as their equivariance (how predictably they change under input transformations), invariance (how stable they remain under them), and disentanglement (how separated different concepts are within them).

100

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

There’s widespread suspicion that as model and data scale increases, model representations converge (e.g., the platonic representation hypothesis (arxiv.org/abs/2405.07987), in part supported by different architectures and pretraining paradigms leading to similar downstream performance.

100

Christos Plachouras @cplachouras.bsky.social · 13/05/2025

Excited to share our new paper: arxiv.org/abs/2505.06224! We introduce a unified framework for evaluating model representations beyond downstream tasks, and use it to uncover some interesting insights about the structure of representations that challenge conventional wisdom 🔍🧵

153

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

finally, we found that regardless of pretraining data scale, the inherent robustness of representations to relevant data perturbations is still concerning (MusiCNN 🟠, VGG 🔵, AST 🟢, CLMR 🔴, TMAE 🟣, MFCC 🩶)

100

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

we also saw that handcrafted features are actually still outperforming learned representations in some tasks (MFCCs 🩶, Chroma 🤎)

110

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

this could be somewhat attributable to larger-capacity downstream models (2-layer MLP) being able to recover more relevant information from a “bad” representation, hinting that more data could be contributing to relevant information becoming more accessible (MusiCNN 🟠, VGG 🔵)

100

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

we found that representations from some non-trained models perform surprisingly well in tagging, and that the tagging performance gap for different pretraining scales isn’t very large (MusiCNN 🟠, VGG 🔵, AST 🟢, CLMR 🔴, TMAE 🟣, MFCC 🩶)

100

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

we created subsets of MTAT ranging from 5 minutes to 8,000 minutes of audio and pretrained MusiCNN, a VGG, AST, CLMR, and a transformer-based MAE. we then extracted representations and trained SLPs and MLPs on music tagging, instrument recognition, and key detection

100

Christos Plachouras @cplachouras.bsky.social · 10/04/2025

how much music data do we actually need to pretrain effective music representation learning models? we decided to systematically investigate this in our latest #ICASSP2025 paper with @emmanouilb.bsky.social and Johan Pauwels

172

Reposted by Christos Plachouras

MTG-UPF @mtg-upf.bsky.social · 10/03/2025

In the context of the UPF-BMAT Chair on AI and Music, we offer 2 PhD positions to cover topics related to the development of AI models to support the music sector, emphasizing ethical approaches. 🔹 Application deadline: April 21st 2025 🔹 Starting date: October 2025 www.upf.edu/web/mtg/home...

upf.edu

PhD positions on AI and Music

PhD positions on AI and Music 10.03.2025 RRSS WhatsApp RRSS Bluesky RRSS Twitter RRSS Facebook RRSS email Copiar URL PhD positions on AI and Music Application deadline: April 21st 2025.... Research...

022

Reposted by Christos Plachouras

QMUL School of Electronic Engineering and Computer Science @qmuleecs.bsky.social · 04/12/2024

We've created our first Starter Pack, collecting accounts from the brilliant researchers and staff who are working at the School of Electronic Engineering and Computer Science. We'll keep adding to this list as our Bluesky community grows: go.bsky.app/45VsoiD #EduSky #AI #ComputerSkyence

go.bsky.app

QMUL EECS

Join the conversation

031

Reposted by Christos Plachouras

keunwoochoi.bsky.social @keunwoochoi.bsky.social · 24/11/2024

#ismir2025 already has an website! isn't it crazy? ismir2025.ismir.net

ismir2025.ismir.net

ISMIR 2025

1145

Reposted by Christos Plachouras

DCASE Challenge @dcase-challenge.bsky.social · 22/11/2024

We're here too now! 🥳

086