Early Stopping for Deep Image Prior

https://openreview.net/forum?id=231ZzrLC8X

#imaging #overfitting #prior

Logistic-Normal Likelihoods for Heteroscedastic Label Noise

Erik Englesson, Amir Mehrpanah, Hossein Azizpour

Action editor: Bo Han.

https://openreview.net/forum?id=7wA65zL3B3

#label #classification #overfitting

Label Noise-Robust Learning using a Confidence-Based Sieving Strategy

https://openreview.net/forum?id=3taIQG4C7H

#label #labels #overfitting

Learning Augmentation Distributions using Transformed Risk Minimization

Evangelos Chatzipantazis, Stefanos Pertigkiozoglou, Kostas Daniilidis, Edgar Dobriban

Action editor: Andriy Mnih.

https://openreview.net/forum?id=LRYtNj8Xw0

#augmentation #augmentations #overfitting

Catastrophic overfitting can be induced with discriminative non-robust features

Guillermo Ortiz-Jimenez, Pau de Jorge, Amartya Sanyal et al.

Action editor: Jakub Tomczak.

https://openreview.net/forum?id=10hCbu70Sr

#overfitting #adversarial #robust

Logistic-Normal Likelihoods for Heteroscedastic Label Noise

https://openreview.net/forum?id=7wA65zL3B3

#label #classification #overfitting

4/
A general issue concerns seductive #research black-box tools (or, equivalently, trending methods "inspired" by published works one doesn't really understand): easy to incur #overfitting, which implies modelling not only the "signal" being studied in too few data, but also (or mostly) their useless noise.

Recursive: if we fall into the trap (no proper #validation), our readers may be led to believe that these shortcuts have a chance to work, perpetuating anti-culture.

#ComputationalModelling

4/
A general issue concerns seductive #research black-box tools (or, equivalently, trending methods "inspired" by published works one doesn't really understand): easy to incur #overfitting, which implies modelling not only the "signal" being studied in too few data, but also (or mostly) their useless noise.

Recursive: if we fall into the trap, our readers may be led to believe these self-deceiving shortcuts to have a chance of actually working, perpetuating anti-culture

#ComputationalModelling

4/

Using fashionable #research black-box tools (or, equivalently, trending methods "inspired" by published works one doesn't really understand), it is far too easy to incur #overfitting, which implies modelling not only the "signal" being studied in too few data, but also (or mostly) their useless noise.

Another harm is that future researchers may be led to believe that these self-deceiving shortcuts may have a real chance of actually working, spreading anti-culture.

#ComputationalModelling

4/

Using fashionable, plug&play black-box tools (or, equivalently, trending methods "inspired" by published works one doesn't really understand), it is far too easy to incur #overfitting, which implies modelling not only the "signal" being studied in too few data but also (or mostly) their useless noise.

Another harm is that future researchers may be led to believe that these self-deceiving shortcuts may have a real chance of actually working, spreading anti-culture.

#ComputationalModelling

4/

Using fashionable, plug&play black-box tools (or, equivalently, trending methods "inspired" by published works one doesn't really understand), it is far too easy to incur #overfitting, which implies modelling not only the "signal" being studied in these few data but also (or mostly) their useless noise.

Another harm is that other researchers may be led to believe that these self-deceiving shortcuts may have a real chance of actually working, spreading anti-culture.

#ComputationalModelling

4/

Using fashionable, ready-to-use black-box tools (or, equivalently, trending methods "inspired" by published works one hasn't really understand), it is far too easy to incur #overfitting, which implies modelling not only the "signal" being studied in these few data but also (or mostly) their useless noise.

Another harm is that other researchers may be led to believe that these self-deceiving shortcuts may have a real chance of actually working, spreading anti-culture

#ComputationalModelling

4/

Using fashionable, ready-to-use black-box tools (or, equivalently, trending methods "inspired" by published works one hasn't really understand), it is far too easy to incur #overfitting, which implies modelling not only the "signal" being studied in these few data but also (or mostly) their useless noise.

The real harm is that other researchers may be led to believe that these self-deceiving shortcuts may have a real chance of actually working.

#CognitiveBias #ComputationalModelling

4/

So easy to incur #overfitting, or modelling not only the studied "signal" in these few data but also (or mostly) their useless noise

3/
A more specific problem (special case of the point #Feynman made so clear) deals with #CrossValidation (where data for selecting/tuning a model are also iteratively used to test it, with allegedly "clever" methods to avoid fooling oneself) and other #MachineLearning mathematical tricks where many dimensions/parameters are tuned by using much less data.

So easy to fall into #overfitting, or modelling not only the studied "signal" in these few data but also (or mostly) their useless noise.

3/
A more specific problem (special case of the point #Feynman made so clear) deals with #CrossValidation (where data for selecting/tuning a model are also iteratively used to test it, with allegedly "clever" methods to avoid fooling oneself) and other #MachineLearning mathematical tricks where many dimensions/parameters are tuned by using much less data. So easy to fall into #overfitting, or modelling not only the studied "signal" in these few data but also (or mostly) their useless noise.

That reads "dreamstime" to you, right?

That reads dreamstime to me.

#overfitting #stabilityai #stablediffusion #SD15

A Simple Unsupervised Data Depth-based Method to Detect Adversarial Images

https://openreview.net/forum?id=bwRXFDNYdk

#adversarial #imagenet #overfitting

On Intriguing Layer-Wise Properties of Robust Overfitting in Adversarial Training

https://openreview.net/forum?id=BaoCnmosJz

#adversarial #overfitting #robustness

This study finds that badly trained ML models overfit. IMHO nothing new. https://twitter.com/Eric_Wallace_/status/1620449934863642624
#ai #machinelearning #ml #overfitting