Posts in category: Deepmind

At DeepMind, the Research Platform Team builds infrastructure to empower and accelerate our AI research. Today, we are excited to share how we developed TF-Replicator, a software library that helps researchers deploy their TensorFlow models on GPUs and Cloud TPUs with minimal effort and no previous experience with distributed systems. TF-Replicators programming model has now been open sourced as part of TensorFlows tf.distribute.Strategy. This blog post gives an overview of the ideas and technical challenges underlying TF-Replicator. For a more comprehensive description, please read our arXiv paper.A recurring theme in recent AI breakthroughs – from AlphaFold to BigGAN to AlphaStar – is the need for effortless and reliable scalability. Increasing amounts of computational capacity allow researchers to train ever-larger neural networks with new capabilities. To address this, the Research Platform Team developed TF-Replicator, which allows researchers to target different hardware accelerators for Machine Learning, scale up workloads to many devices, and seamlessly switch between different types of accelerators.Read More

Carbon-free technologies like renewable energy help combat climate change, but many of them have not reached their full potential. Consider wind power: over the past decade, wind farms have become an important source of carbon-free electricity as the cost of turbines has plummeted and adoption has surged. However, the variable nature of wind itself makes it an unpredictable energy sourceless useful than one that can reliably deliver power at a set time.In search of a solution to this problem, last year, DeepMind and Google started applying machine learning algorithms to 700 megawatts of wind power capacity in the central United States. These wind farmspart of Googles global fleet of renewable energy projectscollectively generate as much electricity as is needed by a medium-sized city.Using a neural network trained on widely available weather forecasts and historical turbine data, we configured the DeepMind system to predict wind power output 36 hours ahead of actual generation. Based on these predictions, our model recommends how to make optimal hourly delivery commitments to the power grid a full day in advance. This is important, because energy sources that can be scheduled (i.e.Read More

Games have been used for decades as an important way to test and evaluate the performance of artificial intelligence systems. As capabilities have increased, the research community has sought games with increasing complexity that capture different elements of intelligence required to solve scientific and real-world problems. In recent years, StarCraft, considered to be one of the most challenging Real-Time Strategy (RTS) games and one of the longest-played esports of all time, has emerged by consensus as a grand challenge for AI research.Read More

In late 2017 we introduced AlphaZero, a single system that taught itself from scratch how to master the games of chess, shogi (Japanese chess), and Go, beating a world-champion program in each case. We were excited by the preliminary results and thrilled to see the response from members of the chess community, who saw in AlphaZeros games a ground-breaking, highly dynamic and unconventional style of play that differed from any chess playing engine that came before it.Today, we are delighted to introduce the full evaluation of AlphaZero, published in the journal Science(Open Access version here), that confirms and updates those preliminary results. It describes how AlphaZero quickly learns each game to become the strongest player in history for each, despite starting its training from random play, with no in-built domain knowledge but the basic rules of the game.Read More

Were excited to announce that the team behind Streams our mobile app that supports doctors and nurses to deliver faster, better care to patientswill be joining Google.Its been a phenomenal journey to see Streams go from initial idea to live deployment, and to hear how its helped change the lives of patients and the nurses and doctors who treat them. The arrival of world-leading health expert Dr. David Feinberg at Google will accelerate these efforts, helping to make a difference to the lives of millions of patients around the world.This is a major milestone for DeepMind! One of the reasons for joining forces with Google in 2014 was the opportunity to use Googles scale and experience in building billion-user products to bring our breakthroughs more rapidly to the wider world. Its been amazing to put this into practice in data centre efficiency, Android battery life, text-to-speech applications, and now the work of our Streams team.Over the past three years weve built a team of experts in what it takes to deploy clinical tools in practice – engineers, clinicians, translational researchers and more.Read More

In August, we announced the first stage of our joint research partnership with Moorfields Eye Hospital, which showed how AI could match world-leading doctors at recommending the correct course of treatment for over 50 eye diseases, and also explain how it arrives at its recommendations.Now were excited to start working on the next research challenge whether we can help clinicians predict eye diseases before symptoms set in.There are two types of age-related macular degeneration (AMD), one of the most common blinding eye diseases, with 170 million sufferers worldwide. The dry form is relatively common among those over 65, and often only causes mild sight loss. However, about 15% of patients with dry AMD go on to develop the more serious form of the disease wet AMD which can cause permanent, blinding sight loss.Currently, ophthalmologists diagnose wet AMD by analysing highly detailed 3D scans of the back of the eye, called OCT scans.Read More

Today we are open sourcing a new library of useful building blocks for writing reinforcement learning (RL) agents in TensorFlow. Named TRFL (pronounced truffle), it represents a collection of key algorithmic components that we have used internally for a large number of our most successful agents such as DQN, DDPG and the Importance Weighted Actor Learner Architecture.A typical deep reinforcement learning agent consists of a large number of interacting components: at the very least, these include the environment and some deep network representing values or policies, but they often also include components such as a learned model of the environment, pseudo-reward functions or a replay system.These parts tend to interact in subtle ways (often not well-documented in papers, as highlighted by Henderson and colleagues), thus making it difficult to identify bugs in such large computational graphs. A recent blog post by OpenAI highlighted this issue by analysing some of the most popular open-source implementations of reinforcement learning agents and finding that six out of 10 had subtle bugs found by a community member and confirmed by the author.One approach to addressing this issue, and helping those in the research community attempting to reproduce results from papers, is through open-sourcing complete agent implementations.Read More

Japanese version followsSix months ago, we joined a groundbreaking new research partnership led by the Cancer Research UK Imperial Centre at Imperial College London to explore whether AI technology could help clinicians diagnose breast cancers on mammograms quicker and more effectively.Breast cancer is a huge global health problem. Around the world, over 1.6 million people are diagnosed with the disease every single year, and 500,000 lose their life to it partly because accurately detecting and diagnosing breast cancer still remains a huge challenge.Working alongside leading breast cancer experts, clinicians and academics in the UK, weve been exploring whether machine learning (a form of AI) could help address this issue.Today, were delighted to announce that this project is expanding internationally, with The Jikei University Hospital, one of Japans foremost medical institutions, joining the collaborationas part of a wider five year partnership they have signed with DeepMind Health.For the purposes of this research, they will be working with us to analyse historic, de-identified mammograms from around 30,000 women taken at the hospital between 2007 and 2018.Read More

Early results from our partnership with the Radiotherapy Department at University College London Hospitals NHS Foundation Trust suggest that we are well on our way to developing an artificial intelligence (AI) system that can analyse and segment medical scans of head and neck cancer to a similar standard as expert clinicians. This segmentation process is an essential but time-consuming step when planning radiotherapy treatment. The findingsalso show that our system can complete this process in a fraction of the time.Speeding up the segmentation processMore than half a million people are diagnosed each year with cancers of the head and neck worldwide. Radiotherapy is a key part of treatment, but clinical staff have to plan meticulously so that healthy tissue doesnt get damaged by radiation: a process which involves radiographers, oncologists and/or dosimetrists manually outlining the areas of anatomy that need radiotherapy, and those areas that should be avoided.Although our work is still at an early stage, we hope it could one day reduce the waiting time between diagnosis and treatment, which could potentially improve outcomes for cancer patients.Read More

Multi-task learning – allowing a single agent to learn how to solve many different tasks – is a longstanding objective for artificial intelligence research. Recently, there has been a lot of excellent progress, with agents likeDQN able to use the same algorithm to learn to play multiple games including Breakout and Pong. These algorithms were used to train individual expert agents for each task. As artificial intelligence research advances to more complex real world domains, building a single general agent – as opposed to multiple expert agents – to learn to perform multiple tasks will be crucial. However, so far, this has proven to be a significant challenge.One reason is that there are often differences in the reward scales our reinforcement learning agents use to judge success, leading them to focus on tasks where the reward is arbitrarilyhigher. For example, in the Atari game Pong, the agent receives a reward of either -1, 0, or +1 per step. In contrast, an agent playing Ms. Pac-Man can obtain hundreds or thousands of points in a single step. Even if the size of individual rewards is comparable, the frequency of rewards can change over time as the agent gets better.Read More