̽��ֱ�� of Cambridge - computer

New type of computer memory could greatly reduce energy use and improve performance

sc604 — Fri, 23 Jun 2023 13:58:55 +0000

̽��ֱ��researchers, led by the ̽��ֱ�� of Cambridge, developed a device that processes data in a similar way as the synapses in the human brain. ̽��ֱ��devices are based on hafnium oxide, a material already used in the semiconductor industry, and tiny self-assembled barriers, which can be raised or lowered to allow electrons to pass.

This method of changing the electrical resistance in computer memory devices, and allowing information processing and memory to exist in the same place, could lead to the development of computer memory devices with far greater density, higher performance and lower energy consumption. ̽��ֱ��results are reported in the journal Science Advances.

Our data-hungry world has led to a ballooning of energy demands, making it ever more difficult to reduce carbon emissions. Within the next few years, artificial intelligence, internet usage, algorithms and other data-driven technologies are expected to consume more than 30% of global electricity.

“To a large extent, this explosion in energy demands is due to shortcomings of current computer memory technologies,” said first author Dr Markus Hellenbrand, from Cambridge’s Department of Materials Science and Metallurgy. “In conventional computing, there’s memory on one side and processing on the other, and data is shuffled back between the two, which takes both energy and time.”

One potential solution to the problem of inefficient computer memory is a new type of technology known as resistive switching memory. Conventional memory devices are capable of two states: one or zero. A functioning resistive switching memory device however, would be capable of a continuous range of states – computer memory devices based on this principle would be capable of far greater density and speed.

“A typical USB stick based on continuous range would be able to hold between ten and 100 times more information, for example,” said Hellenbrand.

Hellenbrand and his colleagues developed a prototype device based on hafnium oxide, an insulating material that is already used in the semiconductor industry. ̽��ֱ��issue with using this material for resistive switching memory applications is known as the uniformity problem. At the atomic level, hafnium oxide has no structure, with the hafnium and oxygen atoms randomly mixed, making it challenging to use for memory applications.

However, the researchers found that by adding barium to thin films of hafnium oxide, some unusual structures started to form, perpendicular to the hafnium oxide plane, in the composite material.

These vertical barium-rich ‘bridges’ are highly structured, and allow electrons to pass through, while the surrounding hafnium oxide remains unstructured. At the point where these bridges meet the device contacts, an energy barrier was created, which electrons can cross. ̽��ֱ��researchers were able to control the height of this barrier, which in turn changes the electrical resistance of the composite material.

“This allows multiple states to exist in the material, unlike conventional memory which has only two states,” said Hellenbrand.

Unlike other composite materials, which require expensive high-temperature manufacturing methods, these hafnium oxide composites self-assemble at low temperatures. ̽��ֱ��composite material showed high levels of performance and uniformity, making them highly promising for next-generation memory applications.

A patent on the technology has been filed by Cambridge Enterprise, the ̽��ֱ��’s commercialisation arm.

“What’s really exciting about these materials is they can work like a synapse in the brain: they can store and process information in the same place, like our brains can, making them highly promising for the rapidly growing AI and machine learning fields,” said Hellenbrand.

̽��ֱ��researchers are now working with industry to carry out larger feasibility studies on the materials, in order to understand more clearly how the high-performance structures form. Since hafnium oxide is a material already used in the semiconductor industry, the researchers say it would not be difficult to integrate into existing manufacturing processes.

̽��ֱ��research was supported in part by the U.S. National Science Foundation and the Engineering and Physical Sciences Research Council (EPSRC), part of UK Research and Innovation (UKRI).

Reference:
Markus Hellenbrand et al. ‘Thin-film design of amorphous hafnium oxide nanocomposites enabling strong interfacial resistive switching uniformity.’ Science Advances (2023). DOI: 10.1126/sciadv.adg1946

Researchers have developed a new design for computer memory that could both greatly improve performance and reduce the energy demands of internet and communications technologies, which are predicted to consume nearly a third of global electricity within the next ten years.

These materials can work like a synapse in the brain: they can store and process information in the same place, like our brains can

Markus Hellenbrand

Andriy Onufriyenko via Getty Images

Artificial intelligence brain

̽��ֱ��text in this work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Images, including our videos, are Copyright © ̽��ֱ�� of Cambridge and licensors/contributors as identified. All rights reserved. We make our image and video content available in a number of ways – as here, on our main website under its Terms and conditions, and on a range of channels including social media that permit your use and sharing of our content under their respective Terms.

Yes

̽��ֱ��life of Pi: Ten years of Raspberry Pi

sc604 — Fri, 25 Feb 2022 16:57:58 +0000

̽��ֱ��most successful computer ever to come out of the UK celebrates its tenth anniversary this year.

Can federated learning save the world?

sc604 — Mon, 10 May 2021 11:23:31 +0000

Artificial intelligence models are used increasingly widely in today’s world. Many carry out natural language processing tasks – such as language translation, predictive text and email spam filters. They are also used to empower smart assistants such as Siri and Alexa to ‘talk’ to us, and to operate driverless cars.

But to function well these models have to be trained on large sets of data, a process that includes carrying out many mathematical operations for every piece of data they are fed. And the data sets they are being trained on are getting ever larger: one recent natural language processing model was trained on a data set of 40 billion words.

As a result, the energy consumed by the training process is soaring. Most AI models are trained on specialised hardware in large data centres. According to a recent paper in the journal Science, the total amount of energy consumed by data centres made up about 1% of global energy use over the past decade – equalling roughly 18 million US homes. And in 2019, a group of researchers at the ̽��ֱ�� of Massachusetts estimated that training one large AI model used in natural language processing could generate around the same amount of CO₂ emissions as five cars would generate over their total lifetime.

Concerned by this, researchers in Cambridge's Department of Computer Science and Technology set out to investigate more energy-efficient approaches to training AI models. Working with collaborators at the ̽��ֱ�� of Oxford, ̽��ֱ�� College London, and Avignon Université, they explored the environmental impact of a different form of training – called federated learning – and discovered that it had a significantly greener impact.

Instead of training the models in data centres, federated learning involves training models across a large number of individual machines. ̽��ֱ��researchers found that this can lead to lower carbon emissions than traditional learning.

Senior Lecturer Dr Nic Lane explains how it works when the training is performed not inside large data centres but over thousands of mobile devices – such as smartphones – where the data is usually collected by the phone users themselves.

“An example of an application currently using federated learning is the next-word prediction in mobile phones,” he said. “Each smartphone trains a local model to predict which word the user will type next, based on their previous text messages. Once trained, these local models are then sent to a server. There, they are aggregated into a final model that will then be sent back to all users.”

And this method has important privacy benefits as well as environmental benefits, points out Dr Pedro Porto Buarque De Gusmao, a postdoctoral researcher working with Lane.

"Users might not want to share the content of their texts with a third party,” he said. “In federated learning, we can keep data local and use the collective power of millions of mobile devices together to train AI models without users’ raw data ever leaving the phone.”

“And besides these privacy-related gains,” said Lane, “in our recent research, we have shown that federated learning can also have a positive impact in reducing carbon emissions.

“Although smartphones have much less processing power than the hardware accelerators used in data centres, they don’t require as much cooling power as the accelerators do. That’s the benefit of distributing the training of models across a wide pool of devices.”

̽��ֱ��researchers recently co-authored a paper on this called ‘Can Federated Learning save the planet?’ and will be discussing their findings at an international research conference, the Flower Summit 2021, on 11 May.

In their paper, they offer the first-ever systematic study of the carbon footprint of federated learning. They measured the carbon footprint of a federated learning setup by training two models — one in image classification, the other in speech recognition – using a server and two chipsets popular in the simple devices that targeted by federated methods. They recorded the energy consumption during training, and how it might vary depending on where in the world the chipsets and server were located.

They found that while there was a difference between CO₂ emission factors among countries, federated learning under many common application settings was reliably ‘cleaner’ than centralised training.

Training a model to classify images in a large image dataset, they found any federated learning setup in France emitted less CO₂than any centralised setup in both China and the US. And in training the speech recognition model, federated learning was more efficient than centralised training in any country.

Such results are further supported by an expanded set of experiments in a follow-up study (‘A first look into the carbon footprint of federated learning’) by the same lab that explores an even wider variety of data sets and AI models. And this research also provides the beginnings of necessary formalism and algorithmic foundation of even lower carbon emissions for federated learning in the future.

Based on their research, the researchers have made available a first-of-its-kind ‘Federated Learning Carbon Calculator’ so that the public and other researchers can estimate how much CO₂is produced by any given pool of devices. It allows users to detail the number and type of devices they are using, which country they are in, which datasets and upload/download speeds they are using and the number of times each device will train on its own data before sending its model for aggregation.

They also offer a similar calculator for estimating the carbon emissions of centralised machine learning.

“ ̽��ֱ��development and usage of AI is playing an increasing role in the tragedy that is climate change,” said Lane, “and this problem will only worsen as this technology continues to proliferate through society. We urgently need to address this which is why we are keen to share our findings showing that federated learning methods can produce less CO₂ than data centres under important application scenarios.

“But even more importantly, our research also shines a light as to how federated learning should evolve towards being even more broadly environmentally friendly. Decentralized methods like this will be key in the invention of future sustainable forms of AI in the years ahead.”

̽��ֱ��researchers will be presenting their work at the Flower Summit 2021, which takes place on Tuesday 11 May 2021.
Can Federated Learning Save the Planet, Xinchi Qiu, Titouan Parcollet, Daniel J. Beutel, Taner Topal, Akhil Mathur, Nicholas D. Lane.
Federated Learning Carbon Calculator
A first look into the carbon footprint of federated learning, Xinchi Qiu, Titouan Parcollet, Javier Fernandez-Marques, Pedro Porto Buarque de Gusmao, Daniel J. Beutel, Taner Topal, Akhil Mathur, Nicholas D. Lane

Training the artificial intelligence models that underpin web search engines, power smart assistants and enable driverless cars, consumes megawatts of energy and generates worrying carbon dioxide emissions. But new ways of training these models are proven to be greener.

̽��ֱ��development and usage of AI is playing an increasing role in the tragedy that is climate change, and this problem will only worsen as this technology continues to proliferate through society

Nic Lane

Image via www.vpnsrus.com

Machine Learning & Artificial Intelligence

̽��ֱ��text in this work is licensed under a Creative Commons Attribution 4.0 International License. Images, including our videos, are Copyright © ̽��ֱ�� of Cambridge and licensors/contributors as identified. All rights reserved. We make our image and video content available in a number of ways – as here, on our main website under its Terms and conditions, and on a range of channels including social media that permit your use and sharing of our content under their respective Terms.

Yes

Licence type:

Attribution

Codecheck confirms reproducibility of COVID-19 model results

sc604 — Mon, 08 Jun 2020 23:00:01 +0000

̽��ֱ��code, script and documentation of the 16 March report, which is available on Github, was subject to an independent review led by Dr Stephen Eglen, from Cambridge’s Department of Applied Mathematics and Theoretical Physics.

Eglen co-founded Codecheck last year to help evaluate the computer programs behind scientific studies. Researchers provide their code and data to Codecheck, who run the code independently to ensure the work can be reproduced.

Last week, Codecheck certified the reproducibility of arguably the most talked-about computational model of the COVID-19 pandemic, that of the Imperial College group led by Professor Neil Ferguson. ̽��ֱ��model suggested that there could be up to half a million deaths in the UK if no measures were taken to slow the spread of the virus, and has been cited as one of the main reasons that lockdown went into effect soon after. However, the Imperial group did not immediately make their code publicly available.

Codecheck.org.uk provided an independent review of the replication of key findings from Report 9 using CovidSim reimplementation. ̽��ֱ��process matches domain expertise and technical skills, taking place as an open peer review. ̽��ֱ��reviewer conducts the codecheck and submits the resulting certificate as part of their review.

̽��ֱ��results confirm that the key finding of Report 9 - on the impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand - are reproducible. Eglen did not review the epidemiology that went into the Imperial model, however.

In his analysis, Dr Eglen said: “Each run generated a tab-delimited file in the output folder. Two R scripts provided by Prof Ferguson were used to summarise these runs into two summary files... These files were compared against the values generated by Prof Ferguson... ̽��ֱ��results were found to be identical. Inserting my results into his Excel spreadsheet generated the same pivot tables.”

̽��ֱ��codecheck found that: “Small variations (mostly under 5%) in the numbers were observed between Report 9 and our runs.” ̽��ֱ��codecheck confirmed the trends and findings of the original report.

Building in part on code originally developed, published and peer-reviewed in 2005 and 2006, the code used for Report 9 continues to be actively developed to allow examination of the wider range of control policies now being deployed as countries relax lockdown. ̽��ֱ��Imperial team is sharing the code to enhance transparency and to allow others to contribute and make use of the simulation.

Refactoring the code has allowed changes to be made more quickly and reliably, including incorporating new data that has become available as the pandemic has progressed.

In addition to the features presented in Imperial Report 9, further strategies can now be examined such as testing and contact tracing, which was not a UK policy option in March.

Users also now have the ability to vary intensity of interventions over time and to calibrate the model to country-specific epidemic data.

Adapted from a piece originally published on the Imperial College London website

How you can support Cambridge's COVID-19 research effort

Donate to support COVID-19 research at Cambridge

Cambridge researcher confirms reproducibility of high-profile Imperial College coronavirus computational model.

Photo by Christian Wiediger on Unsplash

Closeup of computer keyboard

Yes

What makes a faster typist?

sc604 — Thu, 05 Apr 2018 08:31:58 +0000

̽��ֱ��data was collected by researchers from Aalto ̽��ֱ�� in Finland and the ̽��ֱ�� of Cambridge. Volunteers from over 200 countries took the typing test, which is freely available online. Participants were asked to transcribe randomised sentences, and their accuracy and speed were assessed by the researchers.

Unsurprisingly, the researchers found that faster typists make fewer mistakes. However, they also found that the fastest typists also performed between 40 and 70 percent of keystrokes using rollover typing, in which the next key is pressed down before the previous key is lifted. ̽��ֱ��strategy is well-known in the gaming community but has not been observed in a typing study. ̽��ֱ��results will be presented later this month at the ACM CHI Conference on Human Factors in Computing Systems in Montréal.

“Crowdsourcing experiments that allow us to analyse how people interact with computers on a large scale are instrumental for identifying solution principles for the design of next-generation user interfaces,” said study co-author Dr Per Ola Kristensson from Cambridge’s Department of Engineering.

Most of our knowledge of how people type is based on studies from the typewriter era. Now, decades after the typewriter was replaced by computers, people make different types of mistakes. For example, errors where one letter is replaced by another are now more common, whereas in the typewriter era typists often added or omitted characters.

Another difference is that modern users use their hands differently. “Modern keyboards allow us to type keys with different fingers of the same hand with much less force than what was possible with typewriters,” said co-author Anna Feit from Aalto ̽��ֱ��. “This partially explains why self-taught typists using fewer than ten fingers can be as fast as touch typists, which was probably not the case in the typewriter era.”

̽��ֱ��average user in the study typed 52 words per minute, much slower than the professionally trained typists in the 70s and 80s, who typically reached 60-90 words per minute. However, performance varied largely. “ ̽��ֱ��fastest users in our study typed 120 words per minute, which is amazing given that this is a controlled study with randomised phrases,” said co-author Dr Antti Oulasvirta, also from Aalto. “Many informal tests allow users to practice the sentences, resulting in unrealistically high performance.”

̽��ֱ��researchers found that users who had previously taken a typing course actually had a similar typing behaviour as those who had never taken such a course, in terms of how fast they type, how they use their hands and the errors they make - even though they use fewer fingers.

̽��ֱ��researchers found that users display different typing styles, characterised by how they use their hands and fingers, the use of rollover, tapping speeds, and typing accuracy.

For example, some users could be classified as ‘’careless typists’’ who move their fingers quickly but have to correct many mistakes; and others as attentive error-free typists, who gain speed by moving hands and fingers in parallel, pressing the next key before the first one is released.

It is now possible to classify users’ typing behaviour based on the observed keystroke timings which does not require the storage of the text that users have typed. Such information can be useful for example for spell checkers, or to create new personalised training programmes for typing.

“You do not need to change to the touch typing system if you want to type faster,” said Feit. “A few simple exercises can help you to improve your own typing technique.”

̽��ֱ��anonymised dataset is available at the project homepage: http://userinterfaces.aalto.fi/136Mkeystrokes/

Reference:
Dhakal, V., Feit, A., Kristensson, P.O. and Oulasvirta, A. 2018. 'Observations on typing from 136 million keystrokes.' In Proceedings of the 36th ACM Conference on Human Factors in Computing Systems (CHI 2018). ACM Press.

Adapted from an Aalto ̽��ֱ�� press release.

̽��ֱ��largest-ever dataset on typing speeds and styles, based on 136 million keystrokes from 168,000 volunteers, finds that the fastest typists not only make fewer errors, but they often type the next key before the previous one has been released.

Crowdsourcing experiments that allow us to analyse how people interact with computers on a large scale are instrumental for identifying solution principles for the design of next-generation user interfaces.

Per Ola Kristensson

Photo by Cytonn Photography on Unsplash

Want to type faster?

Pay attention to errors, as they are costly to correct. Slow down to avoid them and you will be faster in the long run.
Learn to type without looking at fingers; your motor system will automatically pick up very fast ‘’trills’’ for frequently occurring letter combinations (“the”), which will speed up your typing. Being able to look at the screen while typing also allows you to quickly detect mistakes.
Practice rollover: use different fingers for successive letter keys instead of moving a single finger from one key to another. Then, when typing a letter with one finger, press the next one with the other finger.
Take an online typing test to track performance and identify weaknesses such as high error rates. Make sure that the test requires you to type new sentences so you do not over-practice the same text.
Dedicate time to practice deliberately. People may forget the good habits and relapse to less efficient ways of typing.

̽��ֱ��text in this work is licensed under a Creative Commons Attribution 4.0 International License. For image use please see separate credits above.

Yes

New type of supercomputer could be based on ‘magic dust’ combination of light and matter

sc604 — Mon, 25 Sep 2017 15:00:00 +0000

̽��ֱ��researchers, from Cambridge, Southampton and Cardiff Universities in the UK and the Skolkovo Institute of Science and Technology in Russia, have used quantum particles known as polaritons – which are half light and half matter – to act as a type of ‘beacon’ showing the way to the simplest solution to complex problems. This entirely new design could form the basis of a new type of computer that can solve problems that are currently unsolvable, in diverse fields such as biology, finance or space travel. ̽��ֱ��results are reported in the journal Nature Materials.

Our technological progress -- from modelling protein folding and behaviour of financial markets to devising new materials and sending fully automated missions into deep space -- depends on our ability to find the optimal solution of a mathematical formulation of a problem: the absolute minimum number of steps that it takes to solve that problem.

̽��ֱ��search for an optimal solution is analogous to looking for the lowest point in a mountainous terrain with many valleys, trenches, and drops. A hiker may go downhill and think that they have reached the lowest point of the entire landscape, but there may be a deeper drop just behind the next mountain. Such a search may seem daunting in natural terrain, but imagine its complexity in high-dimensional space. “This is exactly the problem to tackle when the objective function to minimise represents a real-life problem with many unknowns, parameters, and constraints,” said Professor Natalia Berloff of Cambridge’s Department of Applied Mathematics and Theoretical Physics and the Skolkovo Institute of Science and Technology, and the paper’s first author.

Modern supercomputers can only deal with a small subset of such problems when the dimension of the function to be minimised is small or when the underlying structure of the problem allows it to find the optimal solution quickly even for a function of large dimensionality. Even a hypothetical quantum computer, if realised, offers at best the quadratic speed-up for the “brute-force” search for the global minimum.

Berloff and her colleagues approached the problem from an unexpected angle: What if instead of moving along the mountainous terrain in search of the lowest point, one fills the landscape with a magical dust that only shines at the deepest level, becoming an easily detectible marker of the solution?

“A few years ago our purely theoretical proposal on how to do this was rejected by three scientific journals,” said Berloff. “One referee said, ‘Who would be crazy enough to try to implement this?!’ So we had to do it ourselves, and now we’ve proved our proposal with experimental data.”

Their ‘magic dust’ polaritons are created by shining a laser at stacked layers of selected atoms such as gallium, arsenic, indium, and aluminium. ̽��ֱ��electrons in these layers absorb and emit light of a specific colour. Polaritons are ten thousand times lighter than electrons and may achieve sufficient densities to form a new state of matter known as a Bose-Einstein condensate, where the quantum phases of polaritons synchronise and create a single macroscopic quantum object that can be detected through photoluminescence measurements.

̽��ֱ��next question the researchers had to address was how to create a potential landscape that corresponds to the function to be minimised and to force polaritons to condense at its lowest point. To do this, the group focused on a particular type of optimisation problem, but a type that is general enough so that any other hard problem can be related to it, namely minimisation of the XY model which is one of the most fundamental models of statistical mechanics. ̽��ֱ��authors have shown that they can create polaritons at vertices of an arbitrary graph: as polaritons condense, the quantum phases of polaritons arrange themselves in a configuration that correspond to the absolute minimum of the objective function.

“We are just at the beginning of exploring the potential of polariton graphs for solving complex problems,” said co-author Professor Pavlos Lagoudakis, Head of the Hybrid Photonics Lab at the ̽��ֱ�� of Southampton and the Skolkovo Institute of Science and Technology, where the experiments were performed. “We are currently scaling up our device to hundreds of nodes, while testing its fundamental computational power. ̽��ֱ��ultimate goal is a microchip quantum simulator operating at ambient conditions.”

Reference:
Natalia G. Berloff et al. ‘Realizing the classical XY Hamiltonian in polariton simulators.’ Nature Materials (2017). DOI: 10.1038/nmat4971

A team of researchers from the UK and Russia have successfully demonstrated that a type of ‘magic dust’ which combines light and matter can be used to solve complex problems and could eventually surpass the capabilities of even the most powerful supercomputers.

One referee said, ‘Who would be crazy enough to try to implement this?!’

Natalia Berloff

Kirill Kalinin

Creating polariton condensates in the vertices of an arbitrary graph and reading out the quantum phases that represent the absolute minimum of an XY Model

̽��ֱ��text in this work is licensed under a Creative Commons Attribution 4.0 International License. For image use please see separate credits above.

Yes

Cambridge to host transatlantic cyber security competition

Anonymous — Thu, 20 Jul 2017 08:55:49 +0000

̽��ֱ��“Cambridge2Cambridge” cyber security competition, backed by government and industry, is the brainchild of the ̽��ֱ�� of Cambridge and the Massachusetts Institute of Technology (MIT) in the US, and will see talented students pitted against each other in a three-day showdown.

In total, 110 students from 25 universities from the UK and USA will form mixed transatlantic teams and battle against a fictional rogue state in the life-like cyber security competition backed by the National Cyber Security Centre (NCSC) and Cabinet Office.

̽��ֱ��annual event is now in its second year with prize money up for grabs for the winners. It will be held from 24-26 July at Trinity College, Cambridge.

With major cyber-attacks on the increase, according to the NCSC, the need for cyber security experts is more important than ever before.

Professor Frank Stajano, Head of the Academic Centre of Excellence in Cyber Security Research at Cambridge’s Computer Laboratory and the co-founder of Cambridge2Cambridge, said that the competition has been designed to promote greater cyber security collaboration between the UK and USA, and to give students the platform to explore creative ways to combat global cyber-attacks.

“ ̽��ֱ��aim of the competition is also to bring together different individuals in a fun and inclusive environment, where they can apply their cyber security abilities in a collaborative and competitive setting, allowing students to implement the skills they have been taught, while learning new ones in the process,” he said.

It also gives budding cyber enthusiasts the opportunity to meet like-minded individuals, and learn more about careers in the sector by introducing them to key players in the industry and government.

https://cambridge2cambridge.csail.mit.edu/

A major cyber security challenge, aimed at educating and inspiring the next generation of cyber defenders from across the UK and US, will be held at the ̽��ֱ�� of Cambridge next week.

̽��ֱ��aim of the competition is to bring together different individuals in a fun and inclusive environment, where they can apply their cyber security abilities in a collaborative and competitive setting.

Frank Stajano

Inter-ACE Cyber Challenge 2017

̽��ֱ��text in this work is licensed under a Creative Commons Attribution 4.0 International License. For image use please see separate credits above.

Yes

Artificial intelligence: computer says YES (but is it right?)

lw355 — Thu, 20 Oct 2016 14:17:17 +0000

There would always be a first death in a driverless car and it happened in May 2016. Joshua Brown had engaged the autopilot system in his Tesla when a tractor-trailor drove across the road in front of him. It seems that neither he nor the sensors in the autopilot noticed the white-sided truck against a brightly lit sky, with tragic results.

Of course many people die in car crashes every day – in the USA there is one fatality every 94 million miles, and according to Tesla this was the first known fatality in over 130 million miles of driving with activated autopilot. In fact, given that most road fatalities are the result of human error, it has been said that autonomous cars should make travelling safer.

Even so, the tragedy raised a pertinent question: how much do we understand – and trust – the computers in an autonomous vehicle? Or, in fact, in any machine that has been taught to carry out an activity that a human would do?

We are now in the era of machine learning. Machines can be trained to recognise certain patterns in their environment and to respond appropriately. It happens every time your digital camera detects a face and throws a box around it to focus, or the personal assistant on your smartphone answers a question, or the adverts match your interests when you search online.

Machine learning is a way to program computers to learn from experience and improve their performance in a way that resembles how humans and animals learn tasks. As machine learning techniques become more common in everything from finance to healthcare, the issue of trust is becoming increasingly important, says Zoubin Ghahramani, Professor of Information Engineering in Cambridge's Department of Engineering.

Faced with a life or death decision, would a driverless car decide to hit pedestrians, or avoid them and risk the lives of its occupants? Providing a medical diagnosis, could a machine be wildly inaccurate because it has based its opinion on a too-small sample size? In making financial transactions, should a computer explain how robust is its assessment of the volatility of the stock markets?

“Machines can now achieve near-human abilities at many cognitive tasks even if confronted with a situation they have never seen before, or an incomplete set of data,” says Ghahramani. “But what is going on inside the ‘black box’? If the processes by which decisions were being made were more transparent, then trust would be less of an issue.”

His team builds the algorithms that lie at the heart of these technologies (the “invisible bit” as he refers to it). Trust and transparency are important themes in their work: “We really view the whole mathematics of machine learning as sitting inside a framework of understanding uncertainty. Before you see data – whether you are a baby learning a language or a scientist analysing some data – you start with a lot of uncertainty and then as you have more and more data you have more and more certainty.

“When machines make decisions, we want them to be clear on what stage they have reached in this process. And when they are unsure, we want them to tell us.”

One method is to build in an internal self-evaluation or calibration stage so that the machine can test its own certainty, and report back.

Two years ago, Ghahramani’s group launched the Automatic Statistician with funding from Google. ̽��ֱ��tool helps scientists analyse datasets for statistically significant patterns and, crucially, it also provides a report to explain how sure it is about its predictions.

“ ̽��ֱ��difficulty with machine learning systems is you don’t really know what’s going on inside – and the answers they provide are not contextualised, like a human would do. ̽��ֱ��Automatic Statistician explains what it’s doing, in a human-understandable form.”

Where transparency becomes especially relevant is in applications like medical diagnoses, where understanding the provenance of how a decision is made is necessary to trust it.

Dr Adrian Weller, who works with Ghahramani, highlights the difficulty: “A particular issue with new artificial intelligence (AI) systems that learn or evolve is that their processes do not clearly map to rational decision-making pathways that are easy for humans to understand.” His research aims both at making these pathways more transparent, sometimes through visualisation, and at looking at what happens when systems are used in real-world scenarios that extend beyond their training environments – an increasingly common occurrence.

“We would like AI systems to monitor their situation dynamically, detect whether there has been a change in their environment and – if they can no longer work reliably – then provide an alert and perhaps shift to a safety mode.” A driverless car, for instance, might decide that a foggy night in heavy traffic requires a human driver to take control.

Weller’s theme of trust and transparency forms just one of the projects at the newly launched £10 million Leverhulme Centre for the Future of Intelligence (CFI). Ghahramani, who is Deputy Director of the Centre, explains: “It’s important to understand how developing technologies can help rather than replace humans. Over the coming years, philosophers, social scientists, cognitive scientists and computer scientists will help guide the future of the technology and study its implications – both the concerns and the benefits to society.”

CFI brings together four of the world’s leading universities (Cambridge, Oxford, Berkeley and Imperial College, London) to explore the implications of AI for human civilisation. Together, an interdisciplinary community of researchers will work closely with policy-makers and industry investigating topics such as the regulation of autonomous weaponry, and the implications of AI for democracy.

Ghahramani describes the excitement felt across the machine learning field: “It’s exploding in importance. It used to be an area of research that was very academic – but in the past five years people have realised these methods are incredibly useful across a wide range of societally important areas.

“We are awash with data, we have increasing computing power and we will see more and more applications that make predictions in real time. And as we see an escalation in what machines can do, they will challenge our notions of intelligence and make it all the more important that we have the means to trust what they tell us.”

Artificial intelligence has the power to eradicate poverty and disease or hasten the end of human civilisation as we know it – according to a speech delivered by Professor Stephen Hawking 19 October 2016 at the launch of the Centre for the Future of Intelligence.

Computers that learn for themselves are with us now. As they become more common in ‘high-stakes’ applications like robotic surgery, terrorism detection and driverless cars, researchers ask what can be done to make sure we can trust them.

As we see an escalation in what machines can do, they will challenge our notions of intelligence and make it all the more important that we have the means to trust what they tell us

Zoubin Ghahramani

Thierry Ehrmann

2019 by ExperiensS

̽��ֱ��text in this work is licensed under a Creative Commons Attribution 4.0 International License. For image use please see separate credits above.

Yes

Licence type:

Attribution-ShareAlike