Khaled Koutini

132 posts

Khaled Koutini

Khaled Koutini

@kkoutini

Researcher @cpjku | Machine learning | Signal processing | Model compression | Adversarial Robustness. https://t.co/Hi9m5OSgK6

Austria Katılım Haziran 2009
703 Takip Edilen161 Takipçiler
dr. jack morris
dr. jack morris@jxmnop·
wrote some faster dataloading code for HuggingFace datasets – sped up datasets.load_from_disk() from 17 minutes to around 30s the problem I was running into is that our virtual disk at school is really slow. HF loads datasets in a single thread and doesn't always memory-map. there shouldn't be a serial dependency here, and there's no reason to fully load the dataset into memory (can just mmap) big datasets are stored in many separate files ("shards") so you can load each shard individually in multiple processes or threads and then concatenate into a dataset object. this way is so much faster! gist.github.com/jxmorris12/eed…
dr. jack morris tweet media
English
8
23
383
39.3K
dr. jack morris
dr. jack morris@jxmnop·
@gblazex @huggingface i was looking this up and there's actually some discussion on eventually integrating this: github.com/huggingface/da… i think there are some complicated issues that make the threading hard to merge for the general case, not sure exactly why so i just post code here instead :-)
English
2
0
13
1.6K
Khaled Koutini retweetledi
Sepp Hochreiter
Sepp Hochreiter@HochreiterSepp·
Super excited about our adversarial models (not examples!) for uncertainty quantification. Model prediction is uncertain if other models with large posterior predict differently. Improved integral approximation by mixture importance sampling based on constraint optimization. Cool
Kajetan Schweighofer@kschweig_

🚀 Excited to share our latest research on quantifying the predictive uncertainty of machine learning models. QUAM searches for adversarial models (not adversarial examples!) to better estimate the epistemic uncertainty, the uncertainty about chosen model parameters. 1/5

English
1
17
62
9.1K
Khaled Koutini retweetledi
Benjamin Elizalde
Benjamin Elizalde@benjaminelizal·
Our CLAP 👏 model learns audio concepts from natural language supervision. CLAP achieves SoTA in Zero-Shot learning for different audio tasks. For example, in ESC50 CLAP achieves 82% acc, better than human classification with 81%. arxiv.org/pdf/2206.04769…
English
2
3
25
0
Khaled Koutini retweetledi
HEAR Benchmark
HEAR Benchmark@hearbenchmark·
HEAR PMLR journal submissions are open until 2022-06-30. neuralaudio.ai/hear2021-pmlr.… Besides that, people have asked if they can run HEAR benchmarks, get on the leaderboard, cite us in the future. Yes! HEAR is here to stay. See our updated website: neuralaudio.ai
English
0
5
17
0
Khaled Koutini
Khaled Koutini@kkoutini·
If you're attending @NeurIPSConf, don't miss today's Competition Track at 19:00 GMT. HEAR 2021 will be presented @neuralaudio. I will give a lightning talk about our latest work "PaSST: Efficient Training of Audio Transformers with Patchout" #NeurIPS2021
English
0
2
7
0
Khaled Koutini retweetledi
Eduardo Fonseca
Eduardo Fonseca@edfonseca_·
New paper! We evaluate two pooling methods to improve shift invariance in CNNs, obtaining SOTA on FSD50K. Methods are based on low-pass filtering & adaptive sampling of feature maps. They increase robustness to time/freq shifts in the input! w/ @andrebola_ arxiv.org/abs/2107.00623
Eduardo Fonseca tweet media
English
1
4
30
0
Khaled Koutini retweetledi
Hamid Eghbalzadeh
Hamid Eghbalzadeh@heghbalz·
And go check out @kkoutini et, al poster on Over-Parameterization and Generalization in Audio Classification! 📜: arxiv.org/abs/2107.08933
Hanie Sedghi@HanieSedghi

Our @icmlconf workshop, Overparameterization: Pitfalls & Opportunities happens this Saturday! tinyurl.com/yarcsh7k Please join us! If you are interested in attending but could not register for ICML, please reach out! We have a limited number of free registrations! #ICML2021

English
1
1
5
0
Khaled Koutini retweetledi
Shreyan Chowdhury
Shreyan Chowdhury@shreyan_ch·
So many exciting papers accepted for @ismir2021! Can't wait for the conference! Adding ours: "On Perceived Emotion in Expressive Piano Performance: Further Experimental Evidence for the Relevance of Mid-level Perceptual Features" w/Gerhard Widmer @cpjku
English
0
4
19
0
Khaled Koutini retweetledi
Verena Praher
Verena Praher@veroamilbe·
I am happy to announce that our latest work "On the Veracity of Local, Model-agnostic Explanations in Audio Classification: Targeted Investigations with Adversarial Examples" (together w/ @katxiip, Arthur Flexer and Gerhard Widmer from @cpjku) has been accepted to @ismir2021 🎉🥳
English
1
5
25
0
Khaled Koutini retweetledi
Markus Schedl
Markus Schedl@m_schedl·
Very happy to announce a new #LIT project to start in fall @cpjku: Mitigating Gender Bias in Job Recommender Systems: A Machine Learning-Law Synergy (TIMELY). We are looking for PhD students in recommender systems, NLP, and non-discrimination. #HCAI #MMS #RecSys #NLP #RecSys2021
English
1
12
20
0
Khaled Koutini retweetledi
Verena Praher
Verena Praher@veroamilbe·
On Wednesday we are going to present our latest work "Tracing Back Music Emotion Predictions to Sound Sources and Intuitive Perceptual Qualities" at the Sound and Music Computing Conference in Session 8. See you there! @smcnetorg
Shreyan Chowdhury@shreyan_ch

In explainable ML, multiple levels of explanations can provide a more nuanced interpretation of predictions. In our latest (w/@veroamilbe), we combine source separation with perceptual features for 2-level music emotion explanations arxiv.org/abs/2106.07787

English
0
2
11
0

Keşfet

@gblazex @jxmnop @huggingface @ISCAInterspeech @heghbalz @NeurIPSConf @neuralaudio @andrebola_