Hello Deep Learning: Linear combinations - Bert Hubert's writings
a day ago
- - The 'Hello Deep Learning' series introduces neural networks using the EMNIST dataset of handwritten digits (3 and 7).
- - A simple neural network layer (linear combination) is manually configured using the difference between average images of 3s and 7s as weights.
- - By setting a decision rule based on a bias-adjusted score, the model achieves 97.025% accuracy on validation data.
- - Processing includes averaging tensors, computing Hadamard products, and using matrix multiplication for scoring.
- - The code implementation (threeorseven.cc) preprocesses data with MNISTReader, computes weights, applies bias, and evaluates accuracy.