Skip to content

Machine Learning

Mastering Neural Network - Linear Layer and SGD

The human brain remains one of the greatest mysteries, far more complex than anything else we know. It is the most complicated object in the universe that we know of. The underlying processes and the source of consciousness, as well as consciousness itself, remain unknown. Neural Nets are good for popularizing Deep Learning algorithms, but we can't say for sure what mechanism behind biological Neural Networks enables intelligence to arise.

Training result

Visualized Boundaries

Weight Initialization Methods in Neural Networks

Weight initialization is crucial in training neural networks, as it sets the starting point for optimization algorithms. The activation function applies a non-linear transformation in our network. Different activation functions serve different purposes. Choosing the right weight initialization and activation function is key to better neural network performance. Xavier initialization is ideal for Sigmoid or Tanh in feedforward networks. He initialization pairs well with ReLU for faster convergence, especially in CNNs. Matching these improves training efficiency and model performance.

Initialization methods comparison

Comparison of different initialization methods

The journey from Logits to Probabilities

Logits are the raw outputs generated by a model before applying any activation function, such as the sigmoid or softmax functions. These values are unbounded, meaning they can be positive, negative, or zero. Logits represent the model's unnormalized confidence scores for assigning an input to different classes.

Logits vs Probs

Logits vs Probabilities

In a classification problem, especially in multi-class classification, the model produces a set of logits—one for each class. These logits indicate the model's relative confidence in each class without being normalized into probabilities.