Simon Kato

Just your average college student

[Simon Kato @Moncho_K2]

for the modeling of the specialized pathways? What does this difference indicate about the biological process of visualization if anything?

[Simon Kato @Moncho_K2]

This week's journal club had us reading "The functional specialization of visual cortex emerges from training parallel pathways with self-supervised predictive learning."
#SMILEJournalClub
https://t.co/K2hkTPxif7

[Simon Kato @Moncho_K2]

and movement. In particular, they show that a self-supervised model with two pathways can accomplish this task. Of interest, the paper discuss how a naive supervised model with two pathways cannot accomplish this task. What about the self-supervised approach enables...

[Simon Kato @Moncho_K2]

Do the authors believe that it should be up to authors or journals to establish this type of structure? Who is to say which metrics are most insightful? Moreover, opinions about which metrics to report are constantly evolving for good reason, such as effect sizes > p-value.

[Simon Kato @Moncho_K2]

A great paper reminding everyone to be critical when reviewing journal papers. The authors demonstrate how, even among experts, data leakage is a serious concern and question the usefulness of journal paper models due to reproducibility.
https://t.co/JcUGLkPuLU
#SMILEJournalClub

[Simon Kato @Moncho_K2]

The paper advocates for a more well defined basis, structure, metric reporting scheme, and image preprocessing for all machine learning models, especially deep learning models. It seems reasonable for the future of machine learning research to have such guidelines in place.

[Simon Kato @Moncho_K2]

The authors claim that this means that the common practice of using a single, "best" model should be challenged. If we should use multiple models when making prediction, should prediction be done with an ensemble method or simply report each model's output?

[Simon Kato @Moncho_K2]

A paper discussing the "stability" of deep neural networks (DNN) to changes in initial weights. They analyze how the internal layers of DNNs with different initial weights represent the same image differently.
https://t.co/tILNBP0wbH
#SMILEJournalClub

[Simon Kato @Moncho_K2]

The authors claim that this means that the common practice of using a single, "best" model should be challenged. If we should use multiple models when making prediction, should prediction be done with an ensemble method or simply report each model's output?

[Ruogu Fang @RuoguFang]

SMILE Lab at the end of 2020! Keep smiling!

[SMILE Fang Lab @LabFang]

A wonderful time for our SMILE Lab's BBQ party and rock climbing at Lake Wauburg in the beautiful Gainesville's autumn sunshine! Happy Birthday to @kylebsee and @gcfullerton!

[Simon Kato @Moncho_K2]

I would be interested in seeing how a dimension reduction technique, a LASSO method to identify SNPs associated with fundus as a subroutine, or a feature selection technique which allows for interaction such as subset selection would perform with respect to the p-value method.

[Simon Kato @Moncho_K2]

Fun (?) read about how utilizing genetic information along with colour fundus images can provide an improvement over just colour fundus in the diagnosis of late age-related macular degeneration (AMD).
https://t.co/xGFkbmFhI7
#SMILEJournalClub

[Simon Kato @Moncho_K2]

The authors briefly discussed using p-value <1x10^-5 feature selection technique to identify associated SNPs (about 1000) but that this model performed worse. I believe that this feature selection technique suffers from not accounting for interaction between SNPs.