In vjcitn/CSHstats: discussions and exercises on classical statistics for CSHL 2022

suppressPackageStartupMessages({
suppressMessages({
library(littleDeep)  # vjcitn/littleDeep, includes python interface
library(DT)
})
})

Introductory remarks

• learning a nonlinear regression model and putting it to use
• "layers" of a multilayer neural net and their correspondence to patient characteristics (molecular, clinical) and treatment outcomes
  • input, output, loss function
• excellent high level overview comparing convolutional neural nets and autoencoders

Road map

• ImageArray

• ISLR (also a python version Convolutional Neural Net example (also see Jeremy Jordan

• Sfaira model zoo of embeddings and classifiers

Images and categories

library(littleDeep)
data(ciftrain1k)
ciftrain1k
n <- preview(ciftrain1k)

We can filter an ImageArray

preview(filterByType(ciftrain1k, "orange"))

Question on bias: Does the machine recognize the orange or just the "more orange" color distribution? How do you "normalize" images so that the "object itself" presents only the "essential features"?

A trained classifier that ingests JPGs as arrays and returns scores for resemblance to

pre-trained categories

Check here

I'll demonstrate 'jpeg_shrinker' which could be installed on intel macs but not on M1 macs at this time. This program will convert a jpeg image to 32x32 resolution and use a trained CNN to classify the content.

islr_cnn

Sfaira: a model zoo of pretrained embedders and cell-type classifiers from Theis Lab

We'll look at the pkgdown site, specifically "get started" tab.

vjcitn/CSHstats documentation built on Jan. 9, 2025, 8:37 p.m.