Satish Bajagain — Noteshttps://satishbajagain.com.np/notes/Learning notes on neural networks and computational neuroscience.enWhy a shared value needs gradient accumulationhttps://satishbajagain.com.np/notes/micrograd/https://satishbajagain.com.np/notes/micrograd/Follow the two paths through a tiny computation graph to see why backward rules must add their contributions.Wed, 09 Sep 2026 00:00:00 GMTA bigram model is a table of conditional probabilitieshttps://satishbajagain.com.np/notes/bigrams/https://satishbajagain.com.np/notes/bigrams/Count adjacent characters, normalize each row, and understand what the loss actually measures.Wed, 09 Sep 2026 00:00:00 GMTFrom character IDs to a learned contexthttps://satishbajagain.com.np/notes/mlp/https://satishbajagain.com.np/notes/mlp/Track the shapes from an embedding lookup through a multilayer perceptron language model.Wed, 09 Sep 2026 00:00:00 GMTBatch normalization does not make a distribution Gaussianhttps://satishbajagain.com.np/notes/batchnorm/https://satishbajagain.com.np/notes/batchnorm/Separate centering and scaling from distribution shape, and distinguish training from inference.Wed, 09 Sep 2026 00:00:00 GMTWhy the softmax gradient is prediction minus targethttps://satishbajagain.com.np/notes/backprop-ninja/https://satishbajagain.com.np/notes/backprop-ninja/Derive the logit gradient of mean cross-entropy and keep the batch scaling explicit.Wed, 09 Sep 2026 00:00:00 GMTGrowing context by combining neighboring representationshttps://satishbajagain.com.np/notes/wavenet/https://satishbajagain.com.np/notes/wavenet/Follow a hierarchy of character groups and distinguish the learning exercise from the original WaveNet.Wed, 09 Sep 2026 00:00:00 GMTWhat an EEG band can—and cannot—tell ushttps://satishbajagain.com.np/notes/eeg-bands/https://satishbajagain.com.np/notes/eeg-bands/Frequency bands describe parts of a signal; they do not provide a direct label for a person's thoughts.Wed, 09 Sep 2026 00:00:00 GMT