Show newer

Finally, the online network is trained using gradient descent, while target network's weights are updated by averaging them (exponential moving average) with online's weights.

Show thread

This way they are taught to achieve consistent embeddings of observations across different ways of introducing noise.

Show thread

First of all, they specify a fast-learning "online" network, and a slow-learning, "target" one.
For a given sample the online network is trying to predict target network's embedding.
The catch?
They are using different augmentations!

Show thread

Can data augmentation benefit from a separation of "online" and "target" network?
BYOL's answer is yes!
Furthermore, they suggest that using negative examples might be obsolete, as they achieved new SOTA without them.

How does it work? 👇
1/4

Pyhopper is my new best friend for hyperparam optimization, I think we will share a few adventures together.

Pyhopper is my new best friend for hyperparam optimization, I think we will share a few adventures together.

pyhopper.io/

Oh well
---
RT @tunguz
You could spend 5 years in grad school working on a new NN for tabular data.

Or you could spend 5 seconds to type

import xgboost as xgb

It‘s your choice.

But only one of those choices is guaranteed to work.
twitter.com/tunguz/status/1579

RT @CAMDA_conf
Tomorrow Tue 9am EDT, CAMDA2022 Trophy Winner Eunyoung Kim @GIST_science talks about Interpretable prediction of drug-drug interactions using drug-induced gene expressions (DeSIDE-DDI) at the @iscb Academy Webinar👉 iscb.org/cms_addon/academy/eve camda.info @US_FDA @xdopazo

RT @pawel_labaj
Just a few more days (30.09) to apply for PostDoc in my lab @MCB_UJ @JagiellonskiUni to explore microbial synergies in @metasub, , and studies. Work with @TomaszLabMCB @bioinfkowalski @systemnaBio @daga_blaszczyk @wwydmanski and @mason_lab tinyurl.com/postDocPPLabaj twitter.com/pawel_labaj/status

RT @AsiaK1918
@maksymeristavi Every time I see „referendum” written without quotation marks I’m dying inside. Even worse when combined with „The West considers the referendums to be a „sham””. Disgusting, dishonest media legitimize this cosplay through bad framing.

RT @AlinejadMasih
Islamic Republic killed this woman to enforce hijab.

After days in a coma, source said “Mahsa Amini, 22, died today”.

She was beaten up by morality police because of wearing “bad hijab”.

Iranian women are outraged. Forced hijab is the main pillar of religious dictatorship.

RT @amanpour
Protests are sweeping Iran & women are burning their hijabs after the death last week of Mahsa Amini, following her arrest by the “morality police”. Human rights groups say at least 8 have been killed. Last night, I planned to ask President Raisi about all this and much more. 1/7

Ahh, finally a break from machine learning.

Behave well without me, my models

Do you have a specific method of organising conclusions from your ML experiments?
I'm using @Cometml for automated tracking, but I'm still keeping any higher-level thoughts in notebook-like apps.

RT @pawel_labaj
Explore microbial synergies in @metasub, , and studies. Apply for PostDoc in my lab @MCB_UJ @JagiellonskiUni and work with @TomaszLabMCB @mason_lab @Daniela_Bezdan @chmoei @akkah21 @klas_yu @systemnaBio @RafalMostowy Deadline Sep 30!!! tinyurl.com/postDocPPLabaj

Another day, another week of experiments lost due to "this function is too simple to mess it up, I don't need to test it"

Don't miss out on this, an awesome group is looking for a PhD and a postdoc!
---
RT @pawel_labaj
Still over a week to apply for PhD @MCB_UJ @JagiellonskiUni and join my lab! (Sep 9th) science.phd.uj.edu.pl/en_GB/re In project where we exploit of the microbial intra-community synergies also a postDoc position open (Sep 31st) tinyurl.com/PPL…
twitter.com/pawel_labaj/status

RT @TomekNarkun
Dane z rynków najmu z Kanady i USA. Kryzys mieszkalnictwa wkroczył w rynek najmu. Ceny wzrosły gwałtownie i zabierają rekordową cześć wynagrodzenia. Młodzi już nawet nie myślą o kupnie, martwią się teraz czy cokolwiek wynajmą. Ta sytuacja ma miejsce w wielu rozwiniętych krajach.

I strongly believe that only half of citations of "Microbiome Datasets Are Compositional And This Is Not Optional" are due to it's novel observation and high quality of text.

The other half surely comes from how cool the title is.

frontiersin.org/articles/10.33

Show older
Qoto Mastodon

QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.