AI Technology & Industry Review
In the new paper Evolution Through Large Models, an OpenAI research team shows that large-scale language models (LLMs) trained to generate modern programming language can suggest intelligent mutations that can be leveraged to realize dramatically improved mutation operators for genetic programming.
In the new paper LegoNN: Building Modular Encoder-Decoder Models, Meta AI researchers propose LegoNN, a procedure for building encoder-decoder architectures with decoder modules that can be shared across different tasks without finetuning or significant performance reductions.
In the new paper Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models, 444 authors from 132 institutions introduce Beyond the Imitation Game (BIG-bench), a large-scale, extremely difficult and diverse benchmark that includes 204 tasks for predicting the potentially transformative effects of large language models.
In the new paper UViM: A Unified Modeling Approach for Vision with Learned Guiding Codes, a Google Brain research team proposes UViM, a unified approach that leverages language modelling and discrete representation learning to enable the modelling of a wide range of computer vision tasks without task-specific modifications.
In the new paper Tracing Knowledge in Language Models Back to the Training Data, a team from MIT CSAIL and Google Research proposes a benchmark for tracing language models’ assertions to the associated training data, aiming to establish a principled ground truth and mitigate high compute demands for large neural language model training.
In the new paper Large Language Models are Zero-Shot Reasoners, a research team from the University of Tokyo and Google Brain demonstrates that large language models (LLMs) can become good zero-shot reasoners through the addition of a simple prompt — “Let’s think step by step” — that elicits a step-by-step thinking process before each question is answered. Their Zero-shot-CoT model achieves huge performance gains compared to the zero-shot baseline.
In the new paper Standing on the Shoulders of Giant Frozen Language Models, AI21 Labs researchers propose three novel methods for learning small neural modules that specialize a frozen language model to different tasks. Their compute-saving approach outperforms conventional frozen model methods and challenges fine-tuning performance without sacrificing model versatility.
In the new paper Unifying Language Learning Paradigms, a Google Research/Brain team proposes a framework for pretraining universal language models that are effective across many different tasks. Their 20B parameter model surpasses 175B GPT-3 on the zero-shot SuperGLUE benchmark and triples the performance of T5-XXL on one-shot summarization tasks.
In the new paper Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language, Google researchers argue that the diversity of different foundation models is symbiotic and that it is possible to build a framework that uses structured Socratic dialogue between pre-existing foundation models to formulate new multimodal tasks as a guided exchange between the models without additional finetuning.
A Stanford research team proposes Time Control (TC), a language model that implicitly plans via a latent stochastic process and generates texts consistent with this latent plan to improve performance on long text generation.
DeepMind trains agents to use keyboard and mouse commands with pixel and Document Object Model (DOM) observations to control computers, achieving state-of-the-art and human-level mean performance across all tasks on the MiniWob++ benchmark.
A research team from Alibaba Group’s DAMO Academy proposes OFA (One For All), a pretrained model that unifies modalities and tasks to a simple Seq2Seq learning framework and achieves SOTA results on a series of multimodal tasks.
An OpenAI research team presents an expert iteration-based neural theorem prover capable of solving a curriculum of increasingly difficult mathematical problems (such as high-school olympiad-level problems) from a set of formal statements of sufficiently varied difficulty and without the need for associated ground-truth proofs.
A research team from Microsoft and NVIDIA leverages the NVIDIA Megatron-LM and Microsoft’s DeepSpeed to create an efficient and scalable 3D parallel system that combines data, pipeline, and tensor-slicing based parallelism, achieving superior zero-, one-, and few-shot learning accuracies and new state-of-the-art results on NLP benchmarks.
An OpenAI research team leverages reinforcement learning from human feedback (RLHF) to make significant progress on aligning language models with the users’ intentions. The proposed InstructGPT models are better at following instructions than GPT-3 while also more truthful and less toxic.
A team from Google Research, University of Pennsylvania and Cornell University proposes a principled perspective to filter out common memorization for LMs, introducing “counterfactual memorization” to measure the expected change in a model’s prediction and distinguish “rare” (episodic) memorization from “common” (semantic) memorization in neural LMs.
In the new paper Towards Interactive Language Modeling, a research team from the University of Amsterdam and Meta AI Labs presents a road map detailing the steps to be taken towards interactive language modelling.
A DeepMind research team proposes RETRO (Retrieval-Enhanced Transformer), an enhanced auto-regressive language model that conditions on document chunks retrieved from a large corpus and achieves performance comparable to GPT-3 and Jurassic-1 on the Pile dataset while using 25× fewer parameters.
In the new paper Shaking the Foundations: Delusions in Sequence Models for Interaction and Control, a DeepMind research team explores the origins of mismatch problems in sequence models that lack understanding of the cause and effect of their actions, and addresses the problem by treating actions as causal interventions.
A Google AI research team explores zero-label learning (training with synthetic data only) in natural language processing, and introduces Unsupervised Data Generation (UDG), a training data creation procedure designed to synthesize high-quality training data without human annotations.
A new MIT CSAIL paper presents an alternative, hierarchical approach to sequence-to-sequence learning with quasi-synchronous grammars to improve sequence-to-sequence models’ compositional generalization on diagnostic tasks.
Researchers from Instituto de Telecomunicações, DeepMind, Institute of Systems and Robotics, Instituto Superior Técnico and Unbabel propose “∞-former” — a transformer model with unbounded long-term memory (LTM) that can attend to arbitrarily long contexts.
A Huawei research team proposes a topic-based personalized ranking model (TPRM) that integrates pretrained contextualized term representations and user profiles constructed by a topic model to tailor the output ranking list.
A Google Research team proposes Wordcraft, a text editor with a built-in AI-powered creative writing assistant. Wordcraft uses few-shot learning and the natural affordances of conversation to support a variety of user interactions; and can help with story planning, writing and editing.
A research team from Baidu proposes ERNIE 3.0, a unified framework for pretraining large-scale, knowledge-enhanced models that can easily be tailored for both natural language understanding and generation tasks with zero-shot learning, few-shot learning or fine-tuning, and achieves state-of-the-art results on NLP tasks.
A research team from OpenAI proposes Codex, a specialized GPT model fine-tuned on publicly available code from GitHub that can produce functionally correct Python code bodies from natural language docstrings and could excel at a variety of coding tasks.
A research team from Georgia Tech, Microsoft Research and Microsoft Azure AI studies the collections of “lottery tickets” in extremely over-parametrized models, revealing the generalization performance pattern of winning tickets and proving the existence of “super tickets.”
A research team from New York University, Facebook AI, and a CIFAR Fellow in Learning in Machines & Brains raise doubts regarding large-scale pretrained language models’ few-shot learning abilities. The researchers re-evaluate such abilities with held-out examples unavailable, which they propose constitutes “true few-shot learning.”
A research team from NVIDIA, Stanford University and Microsoft Research propose a novel pipeline parallelism approach that improves throughput by more than 10 percent with a comparable memory footprint, showing such strategies can achieve high aggregate throughput while training models with up to a trillion parameters.
A research team from Hugging Face shows that prompting is indeed beneficial for fine-tuning pretrained language models, and that this benefit can be quantified as worth hundreds of data points on average across classification tasks.
A research team from UC Berkeley, University of Maryland and UC Irvine identifies pitfalls that cause instability in the GPT-3 language model and proposes a contextual calibration procedure that improves accuracy by up to 30 percent.
A team from the Allen Institute for Artificial Intelligence and the University of Washington this week introduced TLDR generation, a new automatic summarization task for scientific papers.
Now, a team from Facebook AI Research, Inria, and Sorbonne Université have released CamemBERT, essentially a French version of Google AI’s game-changing pretrained language model BERT.
In a blog post today OpenAI today announced the final staged release of its 1.5 billion parameter language model GPT-2, along with all associated code and model weights.
Google’s new “ALBERT” language model has achieved state-of-the-art results on three popular benchmark tests for natural language understanding (NLU): GLUE, RACE, and SQuAD 2.0.
Synced Global AI Weekly February 24th
Github developer Hugging Face has updated its repository with a PyTorch reimplementation of the GPT-2 language model small version that OpenAI open-sourced last week, along with pretrained models and fine-tuning examples.