Library
Learn from shared maps
Public maps and explanations, organised around the arXiv paper they explain. 4 maps · 3 sources.
Published notes
Concept write-ups, ordered by how often they helped others.
- Linear Algebra Basics
The study of vectors, vector spaces, matrix operations, and dot products used to represent and transform data.
- Gradient Descent and Backpropagation
An optimization algorithm that calculates loss gradients via the chain rule to iteratively update neural network weights.
- Softmax Function
An activation function that normalizes a vector of real numbers into a probability distribution summing to one.
- Word Embeddings
Dense vector representations of discrete tokens where geometric distance reflects semantic similarity.
- Residual Connections
A network design where a layer's input is added directly to its output, creating shortcut paths for gradient flow.
- Dot-Product Attention
An attention mechanism computing alignment scores by taking the scalar product of query and key vectors.
- Positional Encoding
A method of adding deterministic or learned order signal vectors to input embeddings.
- Transformer Architecture
An encoder-decoder model architecture relying entirely on self-attention mechanisms, bypassing recurrence and convolutions.
- Intrinsic Dimensionality of Model Updates
The property stating that effective learning in high-dimensional parameter spaces can be captured within a much lower-dimensional subspace.
- Matrix Rank and Low-Rank Decomposition
The measure of non-redundant dimensions in a matrix and its factorization into smaller matrices of reduced dimensionality.
- Speaker Verification
The biometric task of authenticating a user's claimed identity based on acoustic features extracted from their voice sample.
- Audio Feature Extraction
The process of converting raw time-domain audio waveforms into compact time-frequency representations such as Mel-spectrograms or MFCCs.
- Linear Algebra Basics
The study of vectors, vector spaces, matrix operations, and dot products used to represent and transform data.
- TinySV Architecture
A resource-efficient speaker verification framework optimized for microcontroller-class devices featuring low-memory on-device adaptation.
- On-Device Learning
The process of fine-tuning or updating machine learning model parameters directly on edge hardware post-deployment.
- Neural Network Quantization
A technique for reducing model size and memory bandwidth by converting floating-point weights and activations to lower-precision representations like 8-bit integers.
- TinyML and Edge Computing
The paradigm of deploying machine learning models onto ultra-low-power microcontrollers with tight memory and power budgets.
- Metric Learning for Verification
A distance-based training approach that shapes an embedding space so that intra-class samples are close and inter-class samples are far apart.
- Speaker Embeddings
Fixed-length vector representations generated by neural networks that capture the unique acoustic and vocal tract characteristics of a speaker.
- Gradient Descent and Backpropagation
An optimization algorithm that calculates loss gradients via the chain rule to iteratively update neural network weights.