Update 7a-OOD-detection-output-based.md

episodes/7a-OOD-detection-output-based.md

@@ -62,7 +62,7 @@ Another critical OOD issue is the generation of fake or non-existent medical ref
 In [evaluations of GPT-3's ability to generate medical literature references](https://hai.stanford.edu/news/generating-medical-errors-genai-and-erroneous-medical-references) , it was found that a significant portion of the references were either entirely fabricated or did not support the claims being made. This was especially true for complex medical inquiries that the model had not seen in its training data.
 
 
-# Detecting and handling OOD data
+## Detecting and handling OOD data
 Given the problems posed by OOD data, a reliable model should identify such instances, and then:
 
 1. Reject them during inference
@@ -72,16 +72,16 @@ The second step is much more complicated/involved since it requires matching OOD
 
 How can we determine whether a given instance is OOD or ID? Over the past several years, there have been a wide assortment of new methods developed to tackle this task. In this episode, we will cover a few of the most common approaches and discuss advantages/disadvantages of each.
 
-## Threshold-based methods
+### Threshold-based methods
 Threshold-based methods are one of the simplest and most intuitive approaches for detecting out-of-distribution (OOD) data. The central idea is to define a threshold on a certain score or confidence measure, beyond which the data point is considered out-of-distribution. Typically, these scores are derived from the model's output probabilities or other statistical measures of uncertainty. There are two general classes of threshold-based methods: output-based and distance-based.
 
-### Output-based thresholds
+#### Output-based thresholds
 Output-based Out-of-Distribution (OOD) detection refers to methods that determine whether a given input is out-of-distribution based on the output of a trained model. These methods typically analyze the model’s confidence scores, energy scores, or other output metrics to identify data points that are unlikely to belong to the distribution the model was trained on. The main approaches within output-based OOD detection include:
 
 - **Softmax scores**: The softmax output of a neural network represents the predicted probabilities for each class. A common threshold-based method involves setting a confidence threshold, and if the maximum softmax score of an instance falls below this threshold, it is flagged as OOD.
 - **Energy**: The energy-based method also uses the network's output but measures the uncertainty in a more nuanced way by calculating an energy score. The energy score typically captures the confidence more robustly, especially in high-dimensional spaces, and can be considered a more general and reliable approach than just using softmax probabilities.
 
-### Distance-based thresholds
+#### Distance-based thresholds
 Distance-based methods calculate the distance of an instance from the distribution of training data features learned by the model. If the distance is beyond a certain threshold, the instance is considered OOD. Common distance-based approaches include:
 
 - **Mahalanobis distance:** This method calculates the Mahalanobis distance of a data point from the mean of the training data distribution. A high Mahalanobis distance indicates that the instance is likely OOD.