ImagineAILab/ai-by-hand-excel
Ai By Hand Excel
Este proyecto consiste en modelos interactivos basados en hojas de cálculo diseñados para demostrar la mecánica matemática de la retropropagación (backpropagation), perceptrones multicapa y atención de transformadores. Sirve como un simulador de red neuronal basado en Excel para calcular manualmente operaciones de tensores y multiplicaciones de matrices para visualizar el flujo de datos.
Los modelos proporcionan una forma de visualizar la lógica interna de las redes neuronales implementando auto-atención y retropropagación a través de fórmulas de celda explícitas. Incluye ejercicios matemáticos específicos para modelar la arquitectura de transformadores y las capas de perceptrones multicapa.
La simulación cubre los fundamentos del aprendizaje profundo, incluyendo el cálculo de pases hacia adelante (forward passes) y el uso de descenso de gradiente para actualizar pesos. Utiliza simulaciones de tensores basadas en celdas para rastrear actualizaciones de parámetros y mapear relaciones ponderadas dentro de los mecanismos de atención.
Features
- Spreadsheet AI Simulations - Builds simplified versions of multi-layer perceptrons in Excel to visualize internal neural network behavior.
- Attention Visualizations - Visualizes self-attention weights and relationships through matrix multiplication cells in a spreadsheet.
- Backpropagation Implementations - Implements gradient-based optimization logic through explicit formulas to demonstrate how weights are updated.
- Educational Neural Network Implementations - Offers pedagogical implementations of neural network components built from first principles using spreadsheet calculations.
- Simulated Transformer Models - Simulates self-attention and the mathematical operations of transformer networks using manual Excel calculations.
- Transformer Architecture Visualizations - Maps the internal mathematical operations and data flow of transformer architectures using cell-based calculations.
- Deep Learning Fundamentals - Provides practical exercises on backpropagation and matrix multiplication to teach core deep learning logic.
- Neural Network Forward Pass Simulators - Provides an interactive simulation of the step-by-step arithmetic involved in a neural network forward pass.
- Neural Network Simulators - Provides a spreadsheet-based environment for manually calculating tensor operations to visualize neural network data flow.
- Cell-Based Tensor Simulations - Implements tensor elements and linear algebra operations using individual spreadsheet cells for visualization.
- MLP Mathematical Exercises - Offers a set of manual calculations in a spreadsheet for modeling the layers and weights of a multi-layer perceptron.
- Educational Tensor Operations - Provides tools for calculating core tensor operations and matrix multiplications to track neural network data flow.
- Reference-Linked Updates - Updates network parameters by linking output error cells directly back to input weight cells.
- Mathematical Architecture Modeling - Models the mathematical operations of transformers and multi-layer perceptrons using manual calculations.
- Spreadsheet Parameter Tracking - Stores model weights and biases in static cell ranges for manual adjustment and observation.
