We introduce Codex, a GPT language model finetuned on publicly available code from GitHub, and study its Python code-writing capabilities. A distinct production version of Codex powers GitHub Copilot. On HumanEval, a new evaluation set we release to measure functional correctness for synthesizing programs from docstrings, our model solves 28.8% of the problems, while GPT-3 solves 0% and GPT-J solves 11.4%. Furthermore, we find that repeated sampling from the model is a surprisingly effective strategy for producing working solutions to difficult prompts. Using this method, we solve 70.2% of our problems with 100 samples per problem. Careful investigation of our model reveals its limitations, including difficulty with docstrings describing long chains of operations and with binding operations to variables. Finally, we discuss the potential broader impacts of deploying powerful code generation technologies, covering safety, security, and economics.

We introduce an exploration bonus for deep reinforcement learning methods that is easy to implement and adds minimal overhead to the computation performed. The bonus is the error of a neural network predicting features of the observations given by a fixed randomly initialized neural network. We also introduce a method to flexibly combine intrinsic and extrinsic rewards. We find that the random network distillation (RND) bonus combined with this increased flexibility enables significant progress on several hard exploration Atari games. In particular we establish state of the art performance on Montezuma's Revenge, a game famously difficult for deep reinforcement learning methods. To the best of our knowledge, this is the first method that achieves better than average human performance on this game without using demonstrations or having access to the underlying state of the game, and occasionally completes the first level.

Effective training of neural networks requires much data. In the low-data regime, parameters are underdetermined, and learnt networks generalise poorly. Data Augmentation alleviates this by using existing data more effectively, but standard data augmentation produces only limited plausible alternative data. Given the potential to generate a much broader set of augmentations, we design and train a generative model to do data augmentation. The model, based on image conditional Generative Adversarial Networks, uses data from a source domain and learns to take a data item and augment it by generating other withinclass data items. As this generative process does not depend on the classes themselves, it can be applied to novel unseen classes. We demonstrate that a Data Augmentation Generative Adversarial Network (DA-GAN) augments classifiers well on Omniglot, EMNIST and VGG-Face.

The field of few-shot learning has recently seen substantial advancements. Most of these advancements came from casting few-shot learning as a meta-learning problem. Model Agnostic Meta Learning or MAML is currently one of the best approaches for few-shot learning via meta-learning. MAML is simple, elegant and very powerful, however, it has a variety of issues, such as being very sensitive to neural network architectures, often leading to instability during training, requiring arduous hyperparameter searches to stabilize training and achieve high generalization and being very computationally expensive at both training and inference times. In this paper, we propose various modifications to MAML that not only stabilize the system, but also substantially improve the generalization performance, convergence speed and computational overhead of MAML, which we call MAML++.

We introduce Digital microfluidic Isolation of Single Cells for -Omics (DISCO), a platform that allows users to select particular cells of interest from a limited initial sample size and connects single-cell sequencing data to their immunofluorescence-based phenotypes. Specifically, DISCO combines digital microfluidics, laser cell lysis, and artificial intelligence-driven image processing to collect the contents of single cells from heterogeneous populations, followed by analysis of single-cell genomes and transcriptomes by next-generation sequencing, and proteomes by nanoflow liquid chromatography and tandem mass spectrometry. The results described herein confirm the utility of DISCO for sequencing at levels that are equivalent to or enhanced relative to the state of the art, capable of identifying features at the level of single nucleotide variations. The unique levels of selectivity, context, and accountability of DISCO suggest potential utility for deep analysis of any rare cell population with contextual dependencies.

Emerging sequencing technologies are allowing us to characterize environmental, clinical and laboratory samples with increasing speed and detail, including real-time analysis and interpretation of data. One example of this is being able to rapidly and accurately detect a wide range of pathogenic organisms, both in the clinic and the field. Genomes can have radically different GC content however, such that accurate sequence analysis can be challenging depending upon the technology used. Here, we have characterized the performance of the Oxford MinION nanopore sequencer for detection and evaluation of organisms with a range of genomic nucleotide bias. We have diagnosed the quality of base-calling across individual reads and discovered that the position within the read affects base-calling and quality scores. Finally, we have evaluated the performance of the current state-of-the-art neural network-based MinION basecaller, characterizing its behavior with respect to systemic errors as well as context- and sequence-specific errors. Overall, we present a detailed characterization the capabilities of the MinION in terms of generating high-accuracy sequence data from genomes with a wide range of nucleotide content. This study provides a framework for designing the appropriate experiments that are the likely to lead to accurate and rapid field-forward diagnostics.

The Oxford MinION, the first commercial nanopore sequencer, is also the first to implement molecule-by-molecule real-time selective sequencing or “Read Until”. As DNA transits a MinION nanopore, real-time pore current data can be accessed and analyzed to provide active feedback to that pore. Fragments of interest are sequenced by default, while DNA deemed non-informative is rejected by reversing the pore bias to eject the strand, providing a novel means of background depletion and/or target enrichment. In contrast to the previously published pattern-matching Read Until approach, our RUBRIC method is the first example of real-time selective sequencing where on-line basecalling enables alignment against conventional nucleic acid references to provide the basis for sequence/reject decisions. We evaluate RUBRIC performance across a range of optimizable parameters, apply it to mixed human/bacteria and CRISPR/Cas9-cut samples, and present a generalized model for estimating real-time selection performance as a function of sample composition and computing configuration.

The marriage of robotics and fluidics provides a route to AI-guided high-throughput synthesis and testing in two modalities: integrated centralized facilities that produce data, and distributed systems that synthesize products and conduct disease surveillance.