helylin

[helylin @helylin1]

From my understanding, with so many parameters NLP prevents overfitting by using really large training datasets. Is this correct?

[helylin @helylin1]

#SMILEJournalClub
It seems like natural language processing models have a lot of parameters. The GatorTron models have, starting from 345 million parameters, to 8.9 billion parameters.

[helylin @helylin1]

#SMILEJournalClub

[helylin @helylin1]

From this article, we can see that vision transformers can be a very powerful tool for classification.

[helylin @helylin1]

Based on my understanding, the difference between vision transformers and CNNs is that CNNs uses patches of an image to draw information while visual transformers use the whole image to draw information. Is this correct?

[helylin @helylin1]

so the treatments effects depend on variables with interaction terms that are statistically significant. I wonder how using causal forests on this simulated dataset would compare to using multiple linear regression on this dataset?

[helylin @helylin1]

#SMILEJournalClub
This week’s article uses causal forests to evaluate the difference in treatment effects. From my understanding, causal forests can determine which subgroup does a treatment have good or negative effects on or for which subgroup are there big or small effects.

[helylin @helylin1]

In statistics, multiple linear regression can also estimate the effects of treatments in different subgroups. In multiple linear regression, there can be interaction terms between our variables and treatment indicator

[helylin @helylin1]

To get the final prediction, the model splits the video into 5 second windows and makes a prediction on each of these windows and takes the median of the predictions. Is this how the model accounts for temporal dependence?

[helylin @helylin1]

#SMILEJournalClub

This week’s article covered using deep learning to diagnose infants for cerebral palsy. Here, the data used are 5-9 minute videos of infants in their waking hours.

[helylin @helylin1]

The goal of transfer learning is to address problems that do not have a large enough sample size by using an already developed AI model and applying it to another problem that does not have a large sample size.

[helylin @helylin1]

#SMILEJournalClub
This week’s article discusses using pretrained convolutional neural networks as a basis for transfer learning to address the limitations of small sample size in developing high accuracy models.

[helylin @helylin1]

However, the dataset for this paper was of healthy white people. Why was this cohort alone used as the dataset? Also, how would the results of this model compare if we used a more diverse dataset?

[helylin @helylin1]

#SMILEJournalClub
This paper used polygenic risk scores (PRS) as an association to multimodal brain features of schizophrenia and developed a PRS -associated biomarker for schizophrenia. Classification using new biomarker was then performed and high accuracy scores were achieved.

[helylin @helylin1]

Are there any instances where you would use more than 1 full-connected layer and if so, why?

[helylin @helylin1]

#SMILEJournalClub
This week’s article provides an overview of Convolutional Neural Networks (CNNs). It talked about the purposes of using CNNs, their architecture, and how to analyze CNN models.

[helylin @helylin1]

Usually, CNNs have a fully connected layer after the convolutional and pooling layers. From my understanding, the fully-connected layer is a list of features that is extracted from the convolutional layers.

[helylin @helylin1]

I wonder what the accuracy and statistical measurements for this model would be on another dataset source such as the UK biobank?

[helylin @helylin1]

#SMILEJournalClub
In my research, I have used retinal biomarkers for Alzheimer’s disease detection. It was interesting to see how the retina can be indicators for other diseases such as diabetes.

[helylin @helylin1]

This article is about using deep learning to detect stages of diabetic retinopathy from retinal images. Currently, the dataset used consists of images from patients who participated in the Shanghai Integrated Diabetes Prevention and Care System.