Michael McGaha

UF '22. SMILE Lab Member.

[Michael McGaha @mqmcgaha1]

I think this might separate it from other papers that use predictive coding for object recognition that only suggest predictive coding makes the model more robust to noise, but do not significantly improve overall accuracy.

[Michael McGaha @mqmcgaha1]

For this week’s #SMILEJournalClub, we read “The functional specialization of visual cortex emerges from training parallel pathways with self-supervised predictive learning”.
https://t.co/cXsowovtZL

[Michael McGaha @mqmcgaha1]

Moreover, I find it intuitive that contrastive predictive coding (CPC) may improve model accuracy in this case since the data is a sequence of observations which are likely autocorrelated.

[Michael McGaha @mqmcgaha1]

I do wonder what are the merits of modeling both pathways in a single ANN since they are distinct biologically, and it is possible to only impair one pathway independently of the other which I do not think is possible with this single ANN model.

[Michael McGaha @mqmcgaha1]

The paper suggests that we can more effectively model the dorsal and ventral pathways with parallel pathways with self-supervised learning. These pathways are responsible for visual recognition and motion perception respectively.

[Michael McGaha @mqmcgaha1]

How is the paper able to conclude that many models do not follow an end-to-end framework? When would a non end-to-end framework be merited?

[Michael McGaha @mqmcgaha1]

This week I read "Convolutional neural networks for classification of Alzheimer’s
disease: Overview and reproducible evaluation".

https://t.co/SlzcyCXGuh

Overall, the paper presents many issues that I have also observed when reading other ML papers.
#SMILEJournalClub

[Michael McGaha @mqmcgaha1]

This paper also showed the importance of using balanced accuracy metrics when there is significant data imbalance.

[Michael McGaha @mqmcgaha1]

and that bias can be introduced via transfer learning.
Lastly, I thought it was interesting how a linear SVM with voxels could outperform many models on 2D data in Alzheimer's Disease classification (AD).

[Michael McGaha @mqmcgaha1]

This essentially means that the training, validation, and testing sets were not fully disjoint or couldnt be believed to be independent. This is not a matter of a simple indexing issue, but rather the paper identifies four major causes.

[Michael McGaha @mqmcgaha1]

Common problems include: an architecture's code not being visible, lack of detail on how images were preprocessed, and the way in which cross validation and training were performed. A major emphasis of this paper was the assertion that many papers may have data leakage problems.

[Michael McGaha @mqmcgaha1]

Bernoulli dropout was a suggestion to ameliorate low shared variance. Shared variance was maximized with p =0.64. Does this suggest that there is overfitting without this regularization? Higher values of p reduced shared variance since it breaks centroid consistency.

[Michael McGaha @mqmcgaha1]

Today I read the paper "Individual differences among deep neural network models" which investigates using multiple networks as opposed to a single network and comparing internal representations among networks.

https://t.co/0FJ2eIIL5a
#SMILEJournalClub

[Michael McGaha @mqmcgaha1]

The researchers suggested that individual differences were the result of training and the arrangement of category exemplars within clusters since the models were robust in identifying category centroids.

[Michael McGaha @mqmcgaha1]

I think that this is a novel idea and is quite analogous to bootstrap since they posit that scientists can use representational consistency to estimate expected network variability, a similar goal of using resampling to estimate model variability.

[Michael McGaha @mqmcgaha1]

the model also produced a secondary output (AMD Severity). Interestingly, the models with AMD severity as a secondary output yielded slightly higher AUC scores.

https://t.co/xO1vyTHdFy

[Michael McGaha @mqmcgaha1]

Interesting to read about how single nucleotide polymorphisms (SNPs) when combined with fundus images can improve accuracy and AUC scores. Rather than yielding a single output (probability probability that
the time to late AMD exceeds the inquired duration),

#SMILEJournalClub