Substantial reorganization of notebooks with consistent formatting

chapters/11A.ipynb

@@ -37,7 +37,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "# Inference in AI/ML:"
+    "## **Inference in AI/ML**"
    ],
    "metadata": {
     "id": "Zdkyf97QSghd"
@@ -46,7 +46,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "## Difference between training and inference."
+    "### **Difference between training and inference.**"
    ],
    "metadata": {
     "id": "THKGYDJVTUrI"
@@ -82,7 +82,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "###Inference: Using the deep learning model"
+    "### **Inference: Using the deep learning model**"
    ],
    "metadata": {
     "id": "i8GdGUetkwJC"
@@ -100,7 +100,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "### Comparison of training and inference in terms of time :"
+    "### **Comparison of training and inference in terms of time :**"
    ],
    "metadata": {
     "id": "2oWQzc-uJpha"
@@ -124,7 +124,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "##Data preprocessing and transformation for inference"
+    "### **Data preprocessing and transformation for inference**"
    ],
    "metadata": {
     "id": "doKuhtauJ95r"
@@ -133,7 +133,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "###Preprocessing:\n",
+    "### **Preprocessing:**\n",
     "It is crucial to understand the limitations of inference. First, the input data must be of the same nature as used during training. For example, a trained network used for brain tumor classification on axial brain MRI slices cannot be used for inference on axial brain CT slices, or even sagittal brain MRI slices. It may even require precisely the same slice thickness or even device manufacturer to work well. When an algorithm is overly sensitive to the nature of the data, it is often called 'brittle' and it is not generalized.\n",
     "Second, it is crucial to understand that a preprocessing pipeline applied to training data must also be applied to inference data. As an example, consider diagnosing Alzheimer's Disease (AD) using MRI. In most cases, people use a preprocessing pipeline that includes skull stripping (zeroing out non-brain tissues), reorientation to a consistent image plane, bias field correction, and intensity normalization. Because the model was only exposed to preprocessed data during training, the model will not perform well if inference data is not handled in the same way.\n",
     "\n"
@@ -145,7 +145,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "###Transformation:\n",
+    "### **Transformation:**\n",
     "Data transformation includes a number of different manipulations that alter its properties, but hopefully in a way that preserves the key information. Some transformations alter the size of pixels and thus also the X, Y and Z dimensions. Since MRIs may be acquired with a variety of different pixel sizes and X, Y, and Z dimensions, it is usually a good thing to normalize the size to be consistent prior to using the data for training or inference. \n",
     "Another family of transformations alters intensity. Sometimes this is to set the grayscale range to be similar for all images. In some cases, we may wish to filter images to reduce noise. \n",
     "Another class of spatial transformations alters the primary plane of the images, for instance from Axial to Coronal or sagittal. \n",
@@ -200,7 +200,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "##Inference with torch.no_grad and eval() mode:\n",
+    "### **Inference with torch.no_grad and eval() mode:**\n",
     "\n",
     "PyTorch's model instance has two modes: train and eval: model.train() is for training, and model.eval() is for evaluation which requires inference. The difference is that model.eval() allows the batchnorm or dropout layers to run in evaluate mode rather than training mode. In training mode, dropout is active and, as a result, in every pass, some nodes or weights will be set off randomly. In eval mode, the dropout is off and, based on the trained network, some nodes are on and others are off.\n",
     "Occasionally, for uncertainty quantification, one will want to leave the dropout on in inference, but it is beyond the scope of this chapter, and we will discuss that in the future.\n",
@@ -238,7 +238,7 @@
   {
    "cell_type": "markdown",
    "source": [
-    "##Using visualization techniques to check the model:\n",
+    "### **Using visualization techniques to check the model:**\n",
     "When we train a model and later use it for inference, we can check it to see if it is looking at the correct features or parts of an image. This will enable us to detect potential biases or other errors. This is covered in chapter 10C ([Model Interpretation and Visualization Techniques\n",
     "Validation of MIDeL algorithms](https://)).\n",
     "\n"