Toxicity prevents us from safely deploying powerful pretrained language models for real-world applications. To reduce toxicity in language models, in this post, we will delve into three aspects of the problem: training dataset collection, toxic content detection and model detoxification.

The modern language model with SOTA results on many NLP tasks is trained on large scale free text on the Internet. It is challenging to steer such a model to generate content with desired attributes. Although still not perfect, there are several approaches for controllable text generation, such as guided or learned decoding strategy, smart prompt design, or fine-tuning the model with various methods.

A model that is capable of answering any question with regard to factual knowledge can enable many useful applications. This post delves into how we can build an Open-Domain Question Answering (ODQA) system, assuming we have access to a powerful pretrained language model. Both closed-book and open-book approachs are discussed.

Neural Architecture Search (NAS) automates network architecture engineering. It aims to learn a network topology that can achieve best performance on a certain task. By dissecting the methods for NAS into three components: search space, search algorithm and child model evolution strategy, this post reviews many interesting ideas for better, faster and more cost-efficient automatic neural architecture search.

Exploitation versus exploration is a critical topic in reinforcement learning. This post introduces several common approaches for better exploration in Deep RL.

Inspired by recent progress on various enhanced versions of Transformer models, this post presents how the vanilla Transformer can be improved for longer-term attention span, less memory and computation consumption, RL task solving, etc.

A curriculum is an efficient tool for humans to progressively learn from simple concepts to hard problems. It breaks down complex knowledge by providing a sequence of learning steps of increasing difficulty. In this post, we will examine how the idea of curriculum can help reinforcement learning models learn to solve complicated tasks.

Self-supervised learning opens up a huge opportunity for better utilizing unlabelled data, while learning in a supervised learning manner. This post covers many interesting ideas of self-supervised learning tasks on images, videos, and control problems.

Gradient descent is not the only option when learning optimal model parameters. Evolution Strategies (ES) works out well in the cases where we don’t know the precise analytic form of an objective function or cannot compute the gradients directly. This post dives into several classic ES methods, as well as how ES can be used in deep reinforcement learning.

Meta-RL is meta-learning on reinforcement learning tasks. After trained over a distribution of tasks, the agent is able to solve a new task by developing a new RL algorithm with its internal activity dynamics. This post starts with the origin of meta-RL and then dives into three key components of meta-RL.

If a model or policy is mainly trained in a simulator but expected to work on a real robot, it would surely face the sim2real gap. Domain Randomization (DR) is a simple but powerful idea of closing this gap by randomizing properties of the training environment.

If you are, like me, confused by why deep neural networks can generalize to out-of-sample data points without drastic overfitting, keep on reading.

As a follow up of word embedding post, we will discuss the models on learning contextualized word vectors, as well as the new trend in large unsupervised pre-trained language models which have achieved amazing SOTA results on a variety of language tasks.

Part 4 of the “Object Detection for Dummies” series focuses on one-stage models for fast detection, including SSD, RetinaNet, and models in the YOLO family. These models skip the explicit region proposal stage but apply the detection directly on dense sampled areas.

Meta-learning, also known as “learning to learn”, intends to design models that can learn new skills or adapt to new environments rapidly with a few training examples. There are three common approaches: 1) learn an efficient distance metric (metric-based); 2) use (recurrent) network with external or internal memory (model-based); 3) optimize the model parameters explicitly for fast learning (optimization-based).

In this post, we are looking into the third type of generative models: flow-based generative models. Different from GAN and VAE, they explicitly learn the probability density function of the input data.

Autocoders are a family of neural network models aiming to learn compressed latent variables of high-dimensional data. Starting from the basic autocoder model, this post reviews several variations, including denoising, sparse, and contractive autoencoders, and then Variational Autoencoder (VAE) and its modification beta-VAE.

Attention has been a fairly popular concept and a useful tool in the deep learning community in recent years. In this post, we are gonna look into how attention was invented, and various attention mechanisms and models, such as transformer and SNAIL.

Let’s see how to implement a number of classic deep reinforcement learning models in code.

Abstract: In this post, we are going to look deep into policy gradient, why it works, and many new policy gradient algorithms proposed in recent years: vanilla policy gradient, actor-critic, off-policy actor-critic, A3C, A2C, DPG, DDPG, D4PG, MADDPG, TRPO, PPO, ACER, ACTKR, SAC, TD3 & SVPG.

In this post, we are gonna briefly go over the field of Reinforcement Learning (RL), from fundamental concepts to classic algorithms. Hopefully, this review is helpful enough so that newbies would not get lost in specialized terms and jargons while starting. [WARNING] This is a long read.

The multi-armed bandit problem is a class example to demonstrate the exploration versus exploitation dilemma. This post introduces the bandit problem and how to solve it using different exploration strategies.

In Part 3, we would examine four object detection models: R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN. These models are highly related and the new versions show great speed improvement compared to the older ones.

Part 2 introduces several classic convolutional neural work architecture designs for image classification (AlexNet, VGG, ResNet), as well as DPM (Deformable Parts Model) and Overfeat models for object recognition.

In this series of posts on “Object Detection for Dummies”, we will go through several basic concepts, algorithms, and popular deep learning models for image processing and objection detection. Hopefully, it would be a good read for people with no experience in this field but want to learn more. The Part 1 introduces the concept of Gradient Vectors, the HOG (Histogram of Oriented Gradients) algorithm, and Selective Search for image segmentation.

Word embedding is a dense representation of words in the form of numeric vectors. It can be learned using a variety of language models. The word embedding representation is able to reveal many hidden relationships between words. For example, vector(“cat”) - vector(“kitten”) is similar to vector(“dog”) - vector(“puppy”). This post introduces several models for learning word embedding and how their loss functions are designed for the purpose.

This post is a summary of Prof Naftali Tishby’s recent talk on “Information Theory in Deep Learning”. It presented how to apply the information theory to study the growth and transformation of deep neural networks during training.

This post explains the maths behind a generative adversarial network (GAN) model and why it is hard to be trained. Wasserstein GAN is intended to improve GANs’ training by adopting a smooth metric for measuring the distance between two probability distributions.

This post reviews some research in model interpretability, covering two aspects: (i) interpretable models with model-specific interpretation methods and (ii) approaches of explaining black-box models. I included an open discussion on explainable artificial intelligence at the end.

This post is a continued tutorial for how to build a recurrent neural network using Tensorflow to predict stock market prices. Part 2 attempts to predict prices of multiple stocks using embeddings. The full working code is available in lilianweng/stock-rnn.

This post is a tutorial for how to build a recurrent neural network using Tensorflow to predict stock market prices. Part 1 focuses on the prediction of S&P 500 index. The full working code is available in lilianweng/stock-rnn.

Starting earlier this year, I grew a strong curiosity of deep learning and spent some time reading about this field. To document what I’ve learned and to provide some interesting pointers to people with similar interests, I wrote this overview of deep learning models and their applications.