Jong Wook Kim

State-of-the-art computer vision systems are trained to predict a fixed set of predetermined object categories. This restricted form of supervision limits their generality and usability since additional labeled data is needed to specify any other visual concept. Learning directly from raw text about images is a promising alternative which leverages a much broader source of supervision. We demonstrate that the simple pre-training task of predicting which caption goes with which image is an efficient and scalable way to learn SOTA image representations from scratch on a dataset of 400 million (image, text) pairs collected from the internet. After pre-training, natural language is used to reference learned visual concepts (or describe new ones) enabling zero-shot transfer of the model to downstream tasks. We study the performance of this approach by benchmarking on over 30 different existing computer vision datasets, spanning tasks such as OCR, action recognition in videos, geo-localization, and many types of fine-grained object classification. The model transfers non-trivially to most tasks and is often competitive with a fully supervised baseline without the need for any dataset specific training. For instance, we match the accuracy of the original ResNet-50 on ImageNet zero-shot without needing to use any of the 1.28 million training examples it was trained on. We release our code and pre-trained model weights at https://github.com/OpenAI/CLIP.

We study the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results but in a zero-shot transfer setting without the need for any fine-tuning. When compared to humans, the models approach their accuracy and robustness. We are releasing models and inference code to serve as a foundation for further work on robust speech processing.

Q-learning is arguably one of the most applied representative reinforcement learning approaches and one of the off-policy strategies. Since the emergence of Q-learning, many studies have described its uses in reinforcement learning and artificial intelligence problems. However, there is an information gap as to how these powerful algorithms can be leveraged and incorporated into general artificial intelligence workflow. Early Q-learning algorithms were unsatisfactory in several aspects and covered a narrow range of applications. It has also been observed that sometimes, this rather powerful algorithm learns unrealistically and overestimates the action values hence abating the overall performance. Recently with the general advances of machine learning, more variants of Q-learning like Deep Q-learning which combines basic Q learning with deep neural networks have been discovered and applied extensively. In this paper, we thoroughly explain how Q-learning evolved by unraveling the mathematical complexities behind it as well its flow from reinforcement learning family of algorithms. Improved variants are fully described, and we categorize Q-learning algorithms into single-agent and multi-agent approaches. Finally, we thoroughly investigate up-to-date research trends and key applications that leverage Q-learning algorithms.

There is a need to extract meaningful information from big data, classify it into different categories, and predict end-user behavior or emotions. Large amounts of data are generated from various sources such as social media and websites. Text classification is a representative research topic in the field of natural-language processing that categorizes unstructured text data into meaningful categorical classes. The long short-term memory (LSTM) model and the convolutional neural network for sentence classification produce accurate results and have been recently used in various natural-language processing (NLP) tasks. Convolutional neural network (CNN) models use convolutional layers and maximum pooling or max-overtime pooling layers to extract higher-level features, while LSTM models can capture long-term dependencies between word sequences hence are better used for text classification. However, even with the hybrid approach that leverages the powers of these two deep-learning models, the number of features to remember for classification remains huge, hence hindering the training process. In this study, we propose an attention-based Bi-LSTM+CNN hybrid model that capitalize on the advantages of LSTM and CNN with an additional attention mechanism. We trained the model using the Internet Movie Database (IMDB) movie review data to evaluate the performance of the proposed model, and the test results showed that the proposed hybrid attention Bi-LSTM+CNN model produces more accurate classification results, as well as higher recall and F1 scores, than individual multi-layer perceptron (MLP), CNN or LSTM models as well as the hybrid models.

Large pre-trained models such as CLIP or ALIGN offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning methods substantially improve accuracy on a given target distribution, they often reduce robustness to distribution shifts. We address this tension by introducing a simple and effective method for improving robustness while fine-tuning: ensembling the weights of the zero-shot and fine-tuned models (WiSE-FT). Compared to standard fine-tuning, WiSE-FT provides large accuracy improvements under distribution shift, while preserving high accuracy on the target distribution. On ImageNet and five derived distribution shifts, WiSE-FT improves accuracy under distribution shift by 4 to 6 percentage points (pp) over prior work while increasing ImageNet accuracy by 1.6 pp. WiSE-FT achieves similarly large robustness gains (2 to 23 pp) on a diverse set of six further distribution shifts, and accuracy gains of 0.8 to 3.3 pp compared to standard fine-tuning on commonly used transfer learning datasets. These improvements come at no additional computational cost during fine-tuning or inference.