Programming competition problems are pervasive in the AI community. They can be used to evaluate programmers’ abilities to solve artificial tasks as well as to test the limits of state-of-the-art algorithms.

A research team from MIT, Allen Institute for AI and Microsoft Research recently introduced Python Programming Puzzles (P3), a novel and open-source collection of programming challenges that capture the essence of puzzles and can be used to teach and evaluate an AI’s programming proficiency.


The continuing success of deep learning models has been largely due to their scalability, which allows them to deal with large-scale data and billions of model parameters. The deployment of such huge models on devices with limited resources however remains a challenge in the research community.

A variety of model compression and acceleration techniques have been developed to address this issue, and one of the most popular is knowledge distillation, which effectively learns a small student model from a large teacher model. Knowledge distillation seems a practical and effective solution, but just how well does it really work?

In the…


For standard reinforcement learning (RL) algorithms, the maximization of expected return is achieved by selecting the single highest-reward sequence of actions. But for tasks in a combinatorial domain — such as drug molecule synthesis, where exploration is important — the desired goal is no longer to simply generate the single highest-reward sequence of actions, but rather to carefully sample a diverse set of high-return solutions.

To address this specific machine learning problem, a research team from Mila, McGill University, Université de Montréal, DeepMind and Microsoft has proposed GFlowNet, a novel flow-network-based generative method that can turn a given positive reward…


The success of contemporary machine learning algorithms has been largely driven by increasing model and dataset sizes. The distribution of computation across multiple devices is becoming the pervasive approach for scaling the training of these complex models with large-scale data.

Distributing the learning process however also complicates the implementation process, which can be problematic for the many machine learning practitioners unfamiliar with distributed system mechanisms, especially those with complicated communication topologies.

In a new paper, a research team from DeepMind and Google Brain addresses this issue with Launchpad, a programming model that simplifies the process of defining and launching instances…


Contemporary pretrained multilingual language models (LMs) aim to represent more than 100 languages in a single model. However, despite their state-of-the-art results in cross-lingual transfer, such multilingual models are often incapable of equitably representing their diverse set of languages due to limited capacity, skewed pretraining data and suboptimal vocabularies.

Although language-specific models trained on large custom vocabularies can avoid these issues, they lack the strong cross-lingual transfer abilities of multilingual LMs.

In a bid to encompass the “best of both worlds,” a team from Google Research has proposed MergeDistill, a framework for merging pretrained teacher LMs from multiple monolingual and…


Imagine you’re in an airport, searching for your departure gate. Humans have an excellent ability to extract relevant information from unfamiliar environments to guide us toward a specific goal. This practical conscious processing of information, aka consciousness in the first sense (C1), is achieved by focusing on a small subset of relevant variables from an environment — in the airport scenario we would ignore souvenir shops and so on and focus only on gate-number signage — and it enables us to generalize and adapt well to new situations and to learn new skills or concepts from only limited examples.

In…


It’s no secret that today’s increasingly powerful artificial neural networks (ANNs) bring with them increasing powerful computational appetites. The Open AI paper AI and Compute estimates the compute used by 2018’s AlphaGo Zero was some 300,000 times higher than 2012’s AlexNet. Human brains meanwhile are much more efficient: Stanford Professor of Neurology and Neurosurgery Robert Sapolsky told ESPN that chess grandmasters can burn some 6,000 calories on a high-pressure competition day, which is only about three times the typical human requirement.

Power-efficient neuromorphic intelligence systems have been attracting attention in recent years as a possible way to reduce the versatility…


Recent studies have shown that transformers can model high-dimensional distributions of semantic concepts at scale, opening up the intriguing possibility of formalizing sequential decision-making problems as reinforcement learning (RL). New research from a UC Berkeley, Facebook AI Research and Google Brain team that includes esteemed Belgian professor Pieter Abbeel explores whether generative trajectory modelling — i.e. modelling the joint distribution of a sequence of states, actions, and rewards — could serve as a replacement for conventional RL algorithms.

In the paper Decision Transformer: Reinforcement Learning via Sequence Modeling, the researchers abstract RL as a sequence modelling problem. Their proposed Decision…


AI’s mastery of complex games like Go and StarCraft has boosted research interest in reinforcement learning (RL), where agents provided only with the game rules engage in self-play to elevate their performance to human level and beyond. But how to build reward functions for real-world tasks that lack a clearly defined win condition? Enter Adversarial Imitation Learning (AIL), a framework for continuous control that has been gaining popularity in recent years for solving such complex tasks.

A number of AIL algorithm improvements have been proposed and implemented, such as changing the discriminator’s loss function or switching from on-policy to off-policy…


A new Google Research study has proposed a unified, efficient and modular approach for implicit differentiation of optimization problems that combines the benefits of implicit differentiation and automatic differentiation (autodiff). The researchers say solvers equipped with implicit differentiation set up by the proposed framework can make the autodiff process more efficient for end-users.

Autodiff is a revolutionary technique used in machine learning (ML) solvers for optimization problems. A key advantage of autodiff is that it frees human experts from the tedious burden of manually computing the derivatives of a system’s complex optimization functions.

In most cases, however, the inputs of…

Synced

AI Technology & Industry Review — syncedreview.com | Newsletter: http://bit.ly/2IYL6Y2 | Share My Research http://bit.ly/2TrUPMI | Twitter: @Synced_Global

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store