Reporting and perspectives from Lilian Weng. Headlines and excerpts link to the original articles.
Lilian Weng published a post examining methods for reducing toxic content in pretrained language models. It covers toxicity definitions and categorizations, data collection and annotation, detection approaches, and detoxification techniques.
Lilian Weng's blog post surveys methods for controlling neural text generation with pretrained language models, covering decoding strategies, guided and trainable decoding, prompt design, and fine-tuning approaches. It was updated through September 2021 to add P-tuning, Prompt Tuning, and unlikelihood training.
Lilian Weng published a technical overview of approaches for building open-domain question answering systems, covering open-book retriever-reader and retriever-generator frameworks, closed-book generative QA, and related techniques.
A survey-style report describes Neural Architecture Search (NAS), the automatic design of neural network architectures. It outlines the three components defined by Elsken et al. 2019—search space, search algorithm, and evaluation strategy—and reviews search spaces, search algorithms, and evaluation strategies from prior work.
An updated overview of exploration strategies in deep reinforcement learning, covering classic methods, key problems like hard exploration and noisy-TV, and intrinsic reward techniques including count-based and prediction-based approaches.
Lilian Weng updated her post on the Transformer family on 2023-01-27, refactoring it to incorporate new Transformer models since 2020. The revised version, titled The Transformer Family Version 2.0, is the recommended reference on the topic.
A technical overview of curriculum learning for reinforcement learning, covering task-specific curricula, teacher-guided methods, self-play, automatic goal generation, skill-based curricula, and curriculum through distillation. The article notes updates made on 2020-02-03 and 2020-02-04 and cites work including Elman (1993) and Bengio et al. (2009).
Lilian Weng's self-supervised representation learning post covers pretext tasks for unlabeled data, including image distortion, patch prediction, colorization, generative modeling, and contrastive learning, with update notes from 2020 through 2021.
Lilian Weng published a technical post on Evolution Strategies, covering Gaussian ES, CMA-ES parameter updates, Natural Evolution Strategies, and applications in deep reinforcement learning including OpenAI ES and exploration methods.
Lilian Weng's post surveys Meta Reinforcement Learning, tracing its origin to a 2001 Hochreiter et al. paper and its 2016 proposals by Wang et al. and Duan et al. It covers formulations, key components, and algorithms including meta-gradient RL, Evolved Policy Gradient, MAESN, and episodic control.
Lilian Weng's article explains domain randomization, a sim2real transfer technique that trains robot policies in simulators with randomized visual and physical properties so they can adapt to the real world. It covers uniform and guided approaches, SimOpt, RCAN, DeceptionNet, and ADR.
Lilian Weng published a post examining why deep neural networks generalize despite having many parameters and achieving near-perfect training error, covering classic compression theorems, expressivity results, and the Lottery Ticket Hypothesis (added in a May 27, 2019 update).