Policy Gradients: REINFORCE from Scratch
Deep dive into Policy Gradient methods, specifically REINFORCE deriving the mathematical foundations as well as the implementation from scratch.
Topic
Machine Learning
A collection of 18 issues
Advanced Feature Attribution Techniques for Deep Learning Models
Explainable AI: Exploring various techniques built to shed light on predictions made by deep learning neural networks.
Introduction to Graph Neural Networks: The Message Passing Framework
Exploring the fundamentals of Graph Neural Networks (GNNs), their applications and the underlying mathematical principles.
Artificial Neural Network Backpropagation
Diving deep into how backpropagation of (deep) neural networks work, from derivatives and the chain rule to gradient descent and delta updatess
From Bayesian Priors to Weight Decay
Deriving via MAP estimation how Gaussian priors in the Bayesian Framework lead to the well known L₂ weight decay regularization.
Relating Bayesian Inference, Expected Risk Minimization and Maximum Likelihood Estimation
A birds-eye view on Bayesian Inference, Empirical Risk Minimization and Maximum Likelihood Estimation, relating the different concepts to find core similarities and differences.
An Overview on Matrix Factorization and Dimensionality Reduction
Exploring matrix factorization techniques such as SVD, PCA, NMF, (...) and their applications to simplify complexity in data, improve computational efficiency and reveal hidden patterns.
The Transformer Architecture
Diving deep into the Transformer architecture and its mathematical underpinnings, covering scaled dot-product attention, multi-head attention, and positional encodings, ... .
Exploring how encoder-decoder, encoder-only, and decoder-only models work for NLP, translation and generative AI.
Enhancing HumSet: Improving Humanitarian Crisis Response with AdapterFusion
Exploring ways to leverage NLP and the hierarchical classification framework of HumSet to enhance responses to humanitarian crises beyond baseline results.
Language Models are Few Shot Learners - Meta Learning with GPT-3
Analyzing the paper "Language Models are Few-Shot Learners", exploring GPT-3's meta-learning abilities in Zero, One, and Few-Shot scenarios.