▮ Data The growth of data volume has been exponentially fast, especially these past few years. The plot below by Statista shows that the data volume this year(2023) has nearly doubled compared to 2020. However, despite the abundance of data…
▮ Error Analysis Error analysis is a process of examining the dev set that your ML model misclassified to understand the underlying causes of error. This can help you decide what to prioritize and the direction where the project should…
What is it? Confound variables are like extra independent variables that affect the results. Issues This can cause the following issues. Increase Variance Introduce Bias Avoidance Here are some methods to avoid the above. Control considering variables Random assignment Counterbalancing
Pearson Correlation Evaluates the linear relationship between 2 variables. Ranges from -1(When the value of one variable increases while the other decreases) to 1(When the value of one variable increases while the other increases as well). Spearman Rank-Order Correlation Evaluates…
Adding Linear Complexity When we want to train a model, we can easily imagine that we are unable to capture patterns of the training data if only using straight lines. Polynomial features is useful when you want to add more…
Bayes Theorem Bayes Theorem is about considering the posterior distribution considering prior distribution and currently available data. Let’s say we want to predict a man’s occupation. Is this man a librarian or a farmer given the following description? – He…
KL Divergence measures the distance between 2 distributions. This can be used to understand Cross-Entropy and deep learning model architectures such as VAE. For Example, lets say there is a coin which has 50% chance of being HEADS and 50%…
There are mainly two approaches for optimizing gradient descent. Adjusting the Learning Rate or Adjusting the Gradients.
Why do we use squared loss instead of absolute loss? One reason is because by squaring the loss you can magnify it which can help train the model. Another reason is because absolute loss is not differentiable when equals 0.…
Multi-Class = 1 class per image Multi-Label = Includes multiple label in a single image Softmax = Scale output to 0~1 and make the sum equal to 1 so that it becomes probabilities. Useful for multi-class classification. Sigmoid = Scale…