Krzysztof Strug
902 stories

New Theory Cracks Open the Black Box of Deep Learning

1 Share

Even as machines known as “deep neural networks” have learned to converse, drive cars, beat video games and Go champions, dream, paint pictures and help make scientific discoveries, they have also confounded their human creators, who never expected so-called “deep-learning” algorithms to work so well. No underlying principle has guided the design of these learning systems, other than vague inspiration drawn from the architecture of the brain (and no one really understands how that operates either).

Like a brain, a deep neural network has layers of neurons — artificial ones that are figments of computer memory. When a neuron fires, it sends signals to connected neurons in the layer above. During deep learning, connections in the network are strengthened or weakened as needed to make the system better at sending signals from input data — the pixels of a photo of a dog, for instance — up through the layers to neurons associated with the right high-level concepts, such as “dog.” After a deep neural network has “learned” from thousands of sample dog photos, it can identify dogs in new photos as accurately as people can. The magic leap from special cases to general concepts during learning gives deep neural networks their power, just as it underlies human reasoning, creativity and the other faculties collectively termed “intelligence.” Experts wonder what it is about deep learning that enables generalization — and to what extent brains apprehend reality in the same way.

Last month, a YouTube video of a conference talk in Berlin, shared widely among artificial-intelligence researchers, offered a possible answer. In the talk, Naftali Tishby, a computer scientist and neuroscientist from the Hebrew University of Jerusalem, presented evidence in support of a new theory explaining how deep learning works. Tishby argues that deep neural networks learn according to a procedure called the “information bottleneck,” which he and two collaborators first described in purely theoretical terms in 1999. The idea is that a network rids noisy input data of extraneous details as if by squeezing the information through a bottleneck, retaining only the features most relevant to general concepts. Striking new computer experiments by Tishby and his student Ravid Shwartz-Ziv reveal how this squeezing procedure happens during deep learning, at least in the cases they studied.

Tishby’s findings have the AI community buzzing. “I believe that the information bottleneck idea could be very important in future deep neural network research,” said Alex Alemi of Google Research, who has already developed new approximation methods for applying an information bottleneck analysis to large deep neural networks. The bottleneck could serve “not only as a theoretical tool for understanding why our neural networks work as well as they do currently, but also as a tool for constructing new objectives and architectures of networks,” Alemi said.

Some researchers remain skeptical that the theory fully accounts for the success of deep learning, but Kyle Cranmer, a particle physicist at New York University who uses machine learning to analyze particle collisions at the Large Hadron Collider, said that as a general principle of learning, it “somehow smells right.”

Geoffrey Hinton, a pioneer of deep learning who works at Google and the University of Toronto, emailed Tishby after watching his Berlin talk. “It’s extremely interesting,” Hinton wrote. “I have to listen to it another 10,000 times to really understand it, but it’s very rare nowadays to hear a talk with a really original idea in it that may be the answer to a really major puzzle.”

According to Tishby, who views the information bottleneck as a fundamental principle behind learning, whether you’re an algorithm, a housefly, a conscious being, or a physics calculation of emergent behavior, that long-awaited answer “is that the most important part of learning is actually forgetting.”

The Bottleneck

Tishby began contemplating the information bottleneck around the time that other researchers were first mulling over deep neural networks, though neither concept had been named yet. It was the 1980s, and Tishby was thinking about how good humans are at speech recognition — a major challenge for AI at the time. Tishby realized that the crux of the issue was the question of relevance: What are the most relevant features of a spoken word, and how do we tease these out from the variables that accompany them, such as accents, mumbling and intonation? In general, when we face the sea of data that is reality, which signals do we keep?

“This notion of relevant information was mentioned many times in history but never formulated correctly,” Tishby said in an interview last month. “For many years people thought information theory wasn’t the right way to think about relevance, starting with misconceptions that go all the way to Shannon himself.”

Claude Shannon, the founder of information theory, in a sense liberated the study of information starting in the 1940s by allowing it to be considered in the abstract — as 1s and 0s with purely mathematical meaning. Shannon took the view that, as Tishby put it, “information is not about semantics.” But, Tishby argued, this isn’t true. Using information theory, he realized, “you can define ‘relevant’ in a precise sense.”

Imagine X is a complex data set, like the pixels of a dog photo, and Y is a simpler variable represented by those data, like the word “dog.” You can capture all the “relevant” information in X about Y by compressing X as much as you can without losing the ability to predict Y. In their 1999 paper, Tishby and co-authors Fernando Pereira, now at Google, and William Bialek, now at Princeton University, formulated this as a mathematical optimization problem. It was a fundamental idea with no killer application.

“I’ve been thinking along these lines in various contexts for 30 years,” Tishby said. “My only luck was that deep neural networks became so important.”

Eyeballs on Faces on People on Scenes

Though the concept behind deep neural networks had been kicked around for decades, their performance in tasks like speech and image recognition only took off in the early 2010s, due to improved training regimens and more powerful computer processors. Tishby recognized their potential connection to the information bottleneck principle in 2014 after reading a surprising paper by the physicists David Schwab and Pankaj Mehta.

The duo discovered that a deep-learning algorithm invented by Hinton called the “deep belief net” works, in a particular case, exactly like renormalization, a technique used in physics to zoom out on a physical system by coarse-graining over its details and calculating its overall state. When Schwab and Mehta applied the deep belief net to a model of a magnet at its “critical point,” where the system is fractal, or self-similar at every scale, they found that the network automatically used the renormalization-like procedure to discover the model’s state. It was a stunning indication that, as the biophysicist Ilya Nemenman said at the time, “extracting relevant features in the context of statistical physics and extracting relevant features in the context of deep learning are not just similar words, they are one and the same.”

The only problem is that, in general, the real world isn’t fractal. “The natural world is not ears on ears on ears on ears; it’s eyeballs on faces on people on scenes,” Cranmer said. “So I wouldn’t say [the renormalization procedure] is why deep learning on natural images is working so well.” But Tishby, who at the time was undergoing chemotherapy for pancreatic cancer, realized that both deep learning and the coarse-graining procedure could be encompassed by a broader idea. “Thinking about science and about the role of my old ideas was an important part of my healing and recovery,” he said.

In 2015, he and his student Noga Zaslavsky hypothesized that deep learning is an information bottleneck procedure that compresses noisy data as much as possible while preserving information about what the data represent. Tishby and Shwartz-Ziv’s new experiments with deep neural networks reveal how the bottleneck procedure actually plays out. In one case, the researchers used small networks that could be trained to label input data with a 1 or 0 (think “dog” or “no dog”) and gave their 282 neural connections random initial strengths. They then tracked what happened as the networks engaged in deep learning with 3,000 sample input data sets.

The basic algorithm used in the majority of deep-learning procedures to tweak neural connections in response to data is called “stochastic gradient descent”: Each time the training data are fed into the network, a cascade of firing activity sweeps upward through the layers of artificial neurons. When the signal reaches the top layer, the final firing pattern can be compared to the correct label for the image — 1 or 0, “dog” or “no dog.” Any differences between this firing pattern and the correct pattern are “back-propagated” down the layers, meaning that, like a teacher correcting an exam, the algorithm strengthens or weakens each connection to make the network layer better at producing the correct output signal. Over the course of training, common patterns in the training data become reflected in the strengths of the connections, and the network becomes expert at correctly labeling the data, such as by recognizing a dog, a word, or a 1.

In their experiments, Tishby and Shwartz-Ziv tracked how much information each layer of a deep neural network retained about the input data and how much information each one retained about the output label. The scientists found that, layer by layer, the networks converged to the information bottleneck theoretical bound: a theoretical limit derived in Tishby, Pereira and Bialek’s original paper that represents the absolute best the system can do at extracting relevant information. At the bound, the network has compressed the input as much as possible without sacrificing the ability to accurately predict its label.

Tishby and Shwartz-Ziv also made the intriguing discovery that deep learning proceeds in two phases: a short “fitting” phase, during which the network learns to label its training data, and a much longer “compression” phase, during which it becomes good at generalization, as measured by its performance at labeling new test data.

As a deep neural network tweaks its connections by stochastic gradient descent, at first the number of bits it stores about the input data stays roughly constant or increases slightly, as connections adjust to encode patterns in the input and the network gets good at fitting labels to it. Some experts have compared this phase to memorization.

Then learning switches to the compression phase. The network starts to shed information about the input data, keeping track of only the strongest features — those correlations that are most relevant to the output label. This happens because, in each iteration of stochastic gradient descent, more or less accidental correlations in the training data tell the network to do different things, dialing the strengths of its neural connections up and down in a random walk. This randomization is effectively the same as compressing the system’s representation of the input data. As an example, some photos of dogs might have houses in the background, while others don’t. As a network cycles through these training photos, it might “forget” the correlation between houses and dogs in some photos as other photos counteract it. It’s this forgetting of specifics, Tishby and Shwartz-Ziv argue, that enables the system to form general concepts. Indeed, their experiments revealed that deep neural networks ramp up their generalization performance during the compression phase, becoming better at labeling test data. (A deep neural network trained to recognize dogs in photos might be tested on new photos that may or may not include dogs, for instance.)

It remains to be seen whether the information bottleneck governs all deep-learning regimes, or whether there are other routes to generalization besides compression. Some AI experts see Tishby’s idea as one of many important theoretical insights about deep learning to have emerged recently. Andrew Saxe, an AI researcher and theoretical neuroscientist at Harvard University, noted that certain very large deep neural networks don’t seem to need a drawn-out compression phase in order to generalize well. Instead, researchers program in something called early stopping, which cuts training short to prevent the network from encoding too many correlations in the first place.

Tishby argues that the network models analyzed by Saxe and his colleagues differ from standard deep neural network architectures, but that nonetheless, the information bottleneck theoretical bound defines these networks’ generalization performance better than other methods. Questions about whether the bottleneck holds up for larger neural networks are partly addressed by Tishby and Shwartz-Ziv’s most recent experiments, not included in their preliminary paper, in which they train much larger, 330,000-connection-deep neural networks to recognize handwritten digits in the 60,000-image Modified National Institute of Standards and Technology database, a well-known benchmark for gauging the performance of deep-learning algorithms. The scientists saw the same convergence of the networks to the information bottleneck theoretical bound; they also observed the two distinct phases of deep learning, separated by an even sharper transition than in the smaller networks. “I’m completely convinced now that this is a general phenomenon,” Tishby said.

Humans and Machines

The mystery of how brains sift signals from our senses and elevate them to the level of our conscious awareness drove much of the early interest in deep neural networks among AI pioneers, who hoped to reverse-engineer the brain’s learning rules. AI practitioners have since largely abandoned that path in the mad dash for technological progress, instead slapping on bells and whistles that boost performance with little regard for biological plausibility. Still, as their thinking machines achieve ever greater feats — even stoking fears that AI could someday pose an existential threat — many researchers hope these explorations will uncover general insights about learning and intelligence.

Brenden Lake, an assistant professor of psychology and data science at New York University who studies similarities and differences in how humans and machines learn, said that Tishby’s findings represent “an important step towards opening the black box of neural networks,” but he stressed that the brain represents a much bigger, blacker black box. Our adult brains, which boast several hundred trillion connections between 86 billion neurons, in all likelihood employ a bag of tricks to enhance generalization, going beyond the basic image- and sound-recognition learning procedures that occur during infancy and that may in many ways resemble deep learning.

For instance, Lake said the fitting and compression phases that Tishby identified don’t seem to have analogues in the way children learn handwritten characters, which he studies. Children don’t need to see thousands of examples of a character and compress their mental representation over an extended period of time before they’re able to recognize other instances of that letter and write it themselves. In fact, they can learn from a single example. Lake and his colleagues’ models suggest the brain may deconstruct the new letter into a series of strokes — previously existing mental constructs — allowing the conception of the letter to be tacked onto an edifice of prior knowledge. “Rather than thinking of an image of a letter as a pattern of pixels and learning the concept as mapping those features” as in standard machine-learning algorithms, Lake explained, “instead I aim to build a simple causal model of the letter,” a shorter path to generalization.

Such brainy ideas might hold lessons for the AI community, furthering the back-and-forth between the two fields. Tishby believes his information bottleneck theory will ultimately prove useful in both disciplines, even if it takes a more general form in human learning than in AI. One immediate insight that can be gleaned from the theory is a better understanding of which kinds of problems can be solved by real and artificial neural networks. “It gives a complete characterization of the problems that can be learned,” Tishby said. These are “problems where I can wipe out noise in the input without hurting my ability to classify. This is natural vision problems, speech recognition. These are also precisely the problems our brain can cope with.”

Meanwhile, both real and artificial neural networks stumble on problems in which every detail matters and minute differences can throw off the whole result. Most people can’t quickly multiply two large numbers in their heads, for instance. “We have a long class of problems like this, logical problems that are very sensitive to changes in one variable,” Tishby said. “Classifiability, discrete problems, cryptographic problems. I don’t think deep learning will ever help me break cryptographic codes.”

Generalizing — traversing the information bottleneck, perhaps — means leaving some details behind. This isn’t so good for doing algebra on the fly, but that’s not a brain’s main business. We’re looking for familiar faces in the crowd, order in chaos, salient signals in a noisy world.

This article was reprinted on

Read the whole story
14 hours ago
Warsaw, Poland
Share this story

Men's and women's brains react differently when helping others, study says

1 Share

In women,

part of the brain showed a greater response when sharing money, while in men, the same structure showed more activity when

they kept the cash for themselves, a small


published Monday in Nature Human Behavior found.

Women tend to be more altruistic than men, previous studies have shown.

As Philippe Tobler, co-author of the new study, sees it, "women put more subjective value on prosocial behavior and men find selfish behavior more valuable."

"However, it was unknown how this difference comes about at the level of the brain," Tobler, an associate professor of neuroeconomics and social neuroscience at University of Zurich, wrote in an email. "But in both genders, the dopamine system encodes value."

By "encode," he means the activity in our brain changes in proportion to the value we give social experiences.

Searching for answers for why women and men are not equally selfish, he and his colleagues focused on the dopamine system.

Dopamine, which plays a fundamental role in the brain's reward system, is released during moments of pleasure, yet it also helps us process our values. This mental ability transpires within the brain machinery known as the striatum. Latin for "striped," the striatum is threaded with fibers that receive and transmit signals from the cerebral cortex, the thalamus and other brain regions.

Tobler and his colleagues designed a series of experiments to test how dopamine might influence the behavior of men and women. Fifty-six male and female participants made choices between sharing a financial reward with others or keeping the money for themselves.

Given only a placebo before making decisions, women acted less selfishly than men, choosing to share their money with others. However, when their dopamine systems were disrupted after they received a drug called amisulpride, women acted more selfishly, while men became more generous. Amisulpride is an antipsychotic normally used to treat the symptoms of schizophrenia.

"Based on the opposing priorities of the genders, interfering with the dopamine system has opposing effects," Tobler said.

In a second experiment, the researchers used functional MRI to investigate changes in the brain while eight female and nine male participants made choices. Compared with the males, the striatum in females showed more activity when they made a prosocial decision.

According to Anne Z. Murphy, an associate professor of neuroscience at Georgia State University, other


has shown "that females are more prosocial. We find it more rewarding, and if you manipulate dopamine signaling in the brain, you can make females less prosocial and males less selfish." Murphy was not involved in the study.

Still, she said, the study brings "greater awareness to the fact that there are brain differences in male and females."

"It just shows, once again, that people can point to a biological basis for some of the characteristics that are prototypically male," Murphy said. These traits would include selfishness, self-promotion, generally, a hard-driving profile.

"Now, you can point to another biological basis for it," she added, "And rather than using this knowledge to divide us, maybe we can use this to help make society a better place."

For instance, she said, when women act in more altruistic ways, they shouldn't be regarded as less deserving than male colleagues who are more self-promoting.

Gender differences in the brain may not be due to structural differences -- for example, variations in region size or shape based on sex, noted the researchers. Gender differences in the brain could be functional. This would mean a flood of the very same neurotransmitter -- dopamine -- might cause a very different response in women than in men.

"It may be worth pointing out that the differences are likely to be learned," Tobler said.

Though male and female tendencies may be learned, Murphy said, these behaviors are not acquired in a single lifetime.

'Shaped by history'

Instead, these preferences develop over time based on the differing roles of females and males: "reproduction versus resource-gathering," Murphy said.

"You see similar behavior in rodents," she said, noting that female rats act in more altruistic ways than males. "It's evolutionarily conserved. It's shaped by history."

The study has implications for drug research, Tobler noted.

"Historically, medical drugs were often tested primarily on men and sometimes drugs have been found to be more effective in men than women," he wrote.

See the latest news and share your comments with CNN Health on Facebook and Twitter.

Murphy explained that "preclinical studies have shown that females require approximately twice the amount of morphine than males to produce the same level of analgesia."

All opiates that are metabolized in one specific way produce what is known as a "sexually dimorphic response," she added.

"People are starting to look at whether cannabinoids are sexually dimorphic. It's been suggested that cannabinoids are more effective in females than in males," she said, with a lot of preclinical data showing this is the case.

Read the whole story
14 hours ago
Warsaw, Poland
Share this story

Renewables will give more people access to electricity than coal, says IEA

1 Share

Around the world, more than a billion people still lack access to electricity.

This number is shrinking, down by one third since 2000, despite rising population levels, according to an International Energy Agency (IEA) special report on energy access, published today.

The report says that while coal has supplied nearly half of the progress from 2000 to date, its role is set to decline “dramatically”. This is because renewables are becoming cheaper and because the hardest-to-reach people are in remote, rural areas where off-grid solutions offer the lowest cost.

The report shows the number of people without access to electricity will shrink by another third by 2030, with 60% of these gains supplied by renewables. Furthermore, if the world commits to providing universal access by 2030, then renewables would bridge 90% of the remaining gap, the IEA says.

Recent progress

There have been spectacular gains in providing access to electricity this century, cutting the number without it from 1.7 billion in 2000 to 1.1 billion in 2016, the IEA says. Most of this progress has been in Asia, as the charts below show (blue, yellow and green lines and columns).

Population without electricity access, by region, 2000-2016. Source: IEA special report on energy access.

India has led the way, with 500 million gaining access to electricity. Sub-Saharan Africa now has the majority of people still without access, at 600 million, an increase over the past 15 years due to rising populations. Recently, this number peaked and started to fall (red line and columns).

Fuelling gains

The rate of progress has been accelerating, the IEA says, rising from 62 million people gaining electricity access each year during 2000-2012 to 103 million during 2012-2015.

Coal has been the main source of this new supply, generating 45% of the electricity used by people gaining access for the first time between 2000 and 2016 (purple pictograms in the chart, below).

There has also been a growing role for renewable sources of electricity, the IEA notes, with particularly rapid growth in decentralised off-grid access (dark green pictograms). From 2000-2012, renewables provided 28% of new access to electricity. This figure rose to 34% during 2012-2016.

Annual number of people gaining access to electricity by fuel type. Source: IEA special report on energy access.

There are regional differences in the sources of new electricity connections. In India, for example, coal generated 75% of new supplies, against 20% for renewables. (This pattern is expected to reverse, see below.)

Sub-Saharan Africa has had the most rapid recent improvement in providing electricity access, rising from 9m new connections per year during 2000-2012 to 26m per year during 2012-2016. Most of this acceleration is due to renewables, responsible for 70% of new access since 2012, whereas coal has not supplied any new connections in this period.

Future growth

Looking ahead, the IEA says the number of people without access to electricity will fall to around 700 million by 2030, under its central scenario.

Asia will reach close to 100% access to electricity by 2030 (lilac, yellow and green lines and columns, below) and India will meet its aim of universal access in the early 2020s (blue). The vast majority of the 700 million still without electricity in 2030 will be in sub-Saharan Africa.

Electricity access rate and population without electricity, by region, under the IEA’s central scenario to 2030. Source: IEA special report on energy access.

Note that this chart reflects the IEA’s central “New Policies Scenario”. This includes existing policies plus announced policies and intentions. It also reflects assumptions about the costs of different technologies and the rates of population and electricity demand growth.

Growing grid

Around the world, the share of new electricity access supplied by renewables will nearly double to 60%, up from 34% over the past five years (green, blue and yellow columns, below). This pattern is even more extreme in India, where the share of new electricity from renewables will triple to 60%

Coal’s role in providing electricity access “declines dramatically”, the IEA says, providing power to 16% of those who gain access over the next 14 years. This compares to 45% during 2000-2016.

Population gaining access and cumulative investments, by type, under the central scenario. Source: IEA special report on energy access.

Note that the IEA has been criticised for repeatedly underestimating the rate of growth of renewables, particularly solar. This makes its outlook, in which renewables supply most new electricity access, even more striking.

Role of renewables

If the world wants to meet the Sustainable Development Goal (SDG) of providing universal energy access for all by 2030, then 90% of the additional electricity connections over and above the IEA’s central scenario will come from renewables, its report suggests.

This reflects the fact that the hardest-to-reach populations are those least likely to benefit from grid expansion. For these people, decentralised systems, predominantly supplied by solar (yellow columns, below), offer the “lowest cost pathway” to electricity access.

Additional population gaining access and cumulative investments, by type, under the “Energy for All” scenario, compared to the central scenario. Source: IEA special report on energy access.

The report, for the first time, uses geospatial analysis, at a resolution of one square kilometre, to assess the most cost-effective ways to deliver electricity access to sub-Saharan Africa, whether through grid or off-grid solutions. This analysis takes into account existing and planned infrastructure, technology developments, local resources, population density and likely demand.

It is this new analysis that suggests decentralised renewables will be the cheapest way to provide electricity access for sub-Saharan Africa’s rural poor. Note that research suggests Africa could more than meet its electricity needs, with renewable sources alone.

The IEA puts the cost of providing electricity access to everyone on the planet at an additional $391bn over the period to 2030. This would nearly double total spending, adding to the $324bn already expected to be spent under the IEA’s central scenario.

The energy access-focused SDG also includes provision of clean cooking services. The IEA says this can best be met using liquefied petroleum gas (LPG). As a result, providing universal energy access would increase CO2 emissions by 70m tonnes. This would be more than offset by savings of 165MtCO2 equivalent due to reduced methane and nitrous oxide from biomass used for cooking. The report says:

Achieving universal energy access is not in conflict with achieving climate objectives. The relatively small increase in total primary energy demand and the central role of renewables in our Energy for All Case means that global energy-related carbon dioxide (CO2) emissions increase by just 70 million tonnes (Mt) relative to the New Policies Scenario in 2030 (0.2% of the global level).


The large numbers of people without access to electricity are a frequent point of contention in debates over how to address climate change.

Some proponents cite China and India’s reliance on coal to bring electricity to their populations. They argue that coal is cheap and must be part of the solution for the remaining 1.1 billion people that still lack access to electricity.

Not everyone agrees on how best to meet the needs of these people, who are mostly in sub-Saharan Africa. In a November 2016 interview, Dan Kammen, professor of energy at the University of California, Berkeley and a former science envoy to the US State Department, told Carbon Brief that coal has been given too much credit as a solution to extreme poverty in Africa.

Coal doesn’t even deliver the thing for which it’s really been touted for, and that is, bringing people out of poverty because somehow it’s this least-cost fossil fuel source…I really cringe a bit when I see people touting mega fossil fuel projects as the obvious, first thing to look at…Distributed clean energy, time and time again today, has proven to be better, cheaper, more socially and environmentally positive.

As a July 2017 World Bank blog explains: “In many rural areas in Africa, impacts on economic development of grid extension in the near term may be very modest, while off-grid technologies can be more cost-effective for meeting the most highly-valued basic household needs.”

In further support of the benefits of off-grid systems, it says:

The major downside of off-grid solar is that the relatively low amount of supplied electricity limits what those systems can do for the productive use of electricity. However, electricity usage patterns in newly electrified areas in rural Africa are often such that solar is able to meet those demands. Even in grid-covered rural areas, households and micro-enterprises use electricity mostly for lighting, phone charging, and entertainment – which can easily be provided by solar panels.

Regardless of these details, today’s new IEA report shows that coal’s role in expanding electricity access is set to decline dramatically. Renewables, both on and off the grid, will provide most new connections, as the population without access falls by another third to 700 million.

If the world hopes to meet its goal of universal electricity access by 2030, then the IEA report suggests it is solar – not coal – that will bridge the gap.

Note on definitions

The IEA report defines electricity access as a minimum of 250 kilowatt hours (kWh) per rural household per year. This excludes the more than 23m “pico solar” units sold since 2010. The report explains:

People relying on ‘pico solar’ products, mainly solar lanterns which may include mobile phone chargers, are considered to be below the minimum threshold to count as having [electricity] access. Nevertheless, there are significant benefits for the poor associated with pico solar products.

You can see the range of solutions it considers in its report in the graphic, below.

Illustrative technology options for providing electricity access and the range of uses they can supply. Source: IEA special report on energy access.

The IEA says there is a “general paucity” of data on access to electricity. Its report is based on its own statistics, national statistical agencies, other publicly available data and a network of contacts in government, multilateral development banks and elsewhere.

Sharelines from this story

  • Renewables will help more people to access electricity than coal, says IEA

Read the whole story
1 day ago
Warsaw, Poland
Share this story

Americans are pack rats. Swedes have the solution: ‘Death cleaning.’

1 Share


Get breaking news alerts from The Washington Post

Turn on desktop notifications?

Yes Not now

Read the whole story
6 days ago
Warsaw, Poland
Share this story

In a revealing interview with Henry Blodget, Ray Dalio offers a radical solution to the threat of 'fake news' and details life inside Bridgewater

1 Share
Ray Dalio is the founder of Bridgewater, the world's largest hedge fund.A week ago, his team reached out to me to propose an interview. Mr. Dalio, they said, wanted to discuss the problem of "fake news" and "distorted news," which he believes is exemplified in a recent WSJ articleA week ago, his team reached out to me to propose an interview. Mr. Dalio, they said, wanted to discuss the problem of "fake news" and "distorted news," which he believes is exemplified about his company (and which he wrote about in a post on LinkedInA week ago, his team reached out to me to propose an interview. Mr. Dalio, they said, wanted to discuss the problem of "fake news" and "distorted news," which he believes is exemplified about his company (and which he wrote about in a this week). We agreed on some ground rules: We had control over what we published, but Dalio wanted to be sure we shared his views with our readers, not our view of his views. Thus, I suggested a Q&A.

I spoke with Dalio by phone for nearly two and a half hours. To me, the conversation was itself a case study of the Bridgewater discussion style, which includes lively debate and open disagreement.

Dalio has kept a low profile for most of his career at the helm of Bridgewater, which he founded in the kitchen of his apartment in Queens 40 years ago. He explained that until recently he wanted to be "radically transparent within Bridgewater and radically opaque to those outside it, but as the media are now trying to pry open Bridgewater," he wants to clarify what's going on. He also believes that the systems and culture he has developed at Bridgewater, once fully understood, would be useful well beyond the firm.

Highlights of the interview include:

    Dalio believes that Bridgewater's culture has been misrepresented in the media to the point that incorrect information is perceived as fact.He calls for an independent organization of journalists to regulate the media, asserting that a free press requires regulation just as much as financial markets do.Bridgewater's innovative management processes are designed to improve decision-making, conflict resolution, and personal development. These processes can sound strange and uncomfortable, but Dalio believes deeply in them. He also believes they can work for any organization and society at large.Dalio acknowledges that Bridgewater's culture is unusual and "kooky," but he also notes that it is unusually effective. After working within it, he says, many of Bridgewater's 1,500 employees would never work anywhere else.Dalio says Bridgewater has a turnover rate of about 30% for an employee's first two years, but says the employees who remain - those who like the culture and can handle it - are very loyal.He shared a personal email from an employee thanking him for Bridgewater's culture and process and Dalio's own "teachings," one of many such emails he says he regularly receives.On Bridgewater's unusual decision-making process, Dalio said,"I'm scared of one man, one vote because it suggests that everybody has an equal ability at making decisions, and I think that's dangerous .... I'm also scared of people with power making the decision."He offered details of his solution, whereby everyone has a vote at Bridgewater, but the votes of those with expertise on a topic - as determined by a proprietary "believability" rating - have greater weight.Dalio invited me to take some of Bridgewater's employee-personality-and-capability tests to see what they revealed about me. I eagerly accepted this invitation and will take these tests soon.

The transcript below is edited for length and clarity.

Henry Blodget: Ray, you recently said the Wall Street Journal intentionally distorted the truth about Bridgewater. What do you think the Wall Street Journal intentionally distorted?

Ray Dalio: I think they wanted to create a picture of a weird, oppressive place in which weird things are going on, rather than a place in which there's a lot of hard work and high standards and an unusual culture that works very effectively. The media has the power to create an entrenched perception of reality that's incorrect. Many times people will be quiet about that. I wanted to clarify what Bridgewater is like.

Blodget: I want to be sure I understand the distortions you're referring to. You've been public about Bridgewater's culture, which embraces concepts like "radical truth" and "radical transparency." You yourself have described this culture as "kooky" and "unusual." You have described it as the intellectual and office equivalent of Navy SEALs - where lots of people can't handle it and drop out and only a small tough team of already scrutinized and screened people can handle it and those who can handle it think it's amazing.

Dalio: You're painting something that is in the right direction but an exaggeration. Twenty percent of our people in the first year don't like our culture and leave. Another 10% don't work out. So in the first two years, 30% of the population goes, and then 70% stays. From that point forward, we have hardly any loss. Let me explain what our culture is based on. I think the greatest tragedy of mankind is that people have ideas and opinions in their heads but don't have a process for properly examining these ideas to find out what's true. That creates a world of distortions. That's relevant to what we do, and I think it's relevant to all decision making. So when I say I believe in radical truth and radical transparency, all I mean is we take things that ordinarily people would hide, and we put them on the table, particularly mistakes, problems, and weaknesses. We put those on the table, and we look at them together. We don't hide them. That's what I mean by radical truth. I mean accepting reality. So Bridgewater's culture is not anywhere near as extreme as you're describing.

Blodget: You've talked about how all meetings at Bridgewater are recorded, and anyone can watch the tapes and see how other people talk about them. You've been frustrated that this is often portrayed as oppressive controlling Big Brother behavior. Can you give us some examples of people who have been helped and improved by the culture? People who, as you have said, would not want to work anywhere else?

Dalio: I can read you an e-mail I got yesterday from one our employees. I get them all the time. Do you want me to read that?

Blodget: You bet.

Dalio: "Ray, I wanted to thank you personally so much for your generous support. I'm always so touched that you continue to think of me and so many other people at this time of year. I'm always at a loss of what to give back to you as you've given so much. I want you to know how much Bridgewater, your principles, and our way of being have meant to me and have helped me, especially through this past year. I've had one of the toughest years of my life personally, and I can't tell you how grateful I am to have the instilling and learning of values from this company as well as amazing people who have been there to help and to support me. Everyone will have tough times. I know that, and your principles have helped to guide and helped me make sense of things. I know I can handle anything that life throws at me, and I'm a stronger person for having worked here for so many years. I can't tell you how grateful I am for that. The principles that I rehearsed in my head often, over the past year: Have trust in truth, and you have nothing to fear from truth as well as everyone gets what they deserve out of life. I know it was the combination of the principles I have lived with now and are just part of me that have truly rounded things out for me through the rough waters. I truly mean what I say and have said to my parents often over the past six months. Thank God for Bridgewater and you, Ray. While I'm sure you do not realize the extent of my gratitude to you, please know that I'm eternally grateful for you, all of your teachings and for this amazing company that I have been honored to be part of in the past 12 years, and I hope will be my home for the rest of my career. I will also do everything I can to make it and keep it great. I love you, Ray. I hope you have a wonderful Christmas."

I get a lot of those.

Blodget: That's a great letter. And I can say, as a leader of a company, it's wonderful to get notes like that.

Dalio: We are successful because we have those kinds of relationships with each other that go way beyond the job. I have a saying that the whole purpose of what we do is meaningful work and meaningful relationships, and they support each other. That's why I was frustrated by the Wall Street Journal's characterization of Bridgewater and the process they used. I described these things to them, and said I was happy to have them speak to people like [the employee who wrote the letter] and see our employee satisfaction surveys. But there was no hearing it, right? Instead, they characterized our decision-making process as something weird like "turning Ray's brain into a computer." [ Editor's note: These words were those of a Bridgewater employee quoted in the WSJ article, not the characterization made by the reporter.] They also made it seem like everyone's crying in the bathroom all the time. I'm sure somebody has cried in the bathroom. But it's not like people run around all day crying in bathrooms. I mean, listen to [the employee.]. You can watch so many other people here who hug each other and, you know what, Henry, are in love with each other in a sense and in love with the mission, and that's very powerful.

Blodget: Let me read you some of the facts the Wall Street Journal reported that I think are creating the impression of "weirdness" - the impression that you see as distorted. Now, if we were sitting together on a set, Ray, I would smile and wink at you before I read these things, because a lot of them do sound pretty wacky. First, the Wall Street Journal says Bridgewater is developing a systematic management software system that is sometimes called "The Book of the Future," "The One Thing," or "Principles OS, or just PriOS." This system is described as GPS-style directions for how staff should spend their days and make decisions. Bridgewater employees rate each other all the time on various attributes, using apps. Bridgewater employees each have a "baseball card" that includes their ratings on different criteria, and anyone can access these cards. Every employee takes written tests that measure various attributes, and every employee gets ranked in a "stratum." You, according the Wall Street Journal, are in the highest stratum. You are also considered, based on all of these tests, the most believable and open employee at the company. You've written an extensive set of Principles, and Bridgewater is now developing apps that direct employees to the appropriate section of these Principles that will have the answer to their questions. In the amazing employee letter you just read, the employee referred to your "teachings," and the writer said, "I love you, Ray." Now, you have to admit, all this sounds a little wacky, right, Ray? I mean, can you see where people from the outside would look in and say, "Huh. That is kind of strange and out there"?

Dalio: Okay. Well, first of all, I'd like to be clear that a number of people find me intolerable, and they don't hesitate to say so - which they and I cherish. We wouldn't be successful if we didn't have independent thinkers who can argue and then resolve their arguments. And, yes, I can see that our culture appears odd. Yes, we collect a lot of information. Yes, there are personality tests, there's performance reviews. There are opinions. There's a lot of information that is gathered about people that they participate in. Then we go beyond that. We collectively say, okay, how do we know who is good at what, based on objectively information. "Stratum," for example, is a measure of how well you can view yourself and your ideas objectively. In other words, we as group say, "Okay, now we're looking at the data on people the way we would look at the data on anything else." And, "Okay, given all of those different perspectives, how should we make decisions?"

Blodget: Are there things outside of trying to get to the truth about a particular investment or trade that you're devoting this process to at Bridgewater?

Dalio: Everything.

Blodget: Is there always a majority after the discussion? Is it just a simple vote?

Dalio: No. We have a process of what we call "believability weighting" votes. I won't go through it at length, but each person assesses each other person's credibility on different dimensions, because people are strong and weak in different things. Some people have subject matter expertise in one thing. Some people tend to be creative but not reliable, and others are reliable but not creative. Everybody has different dimensions. What we try to do is to keep that in mind, and we're very clear about it. It's very upfront. It's very data based as to why people have different strengths and weaknesses and what their rankings are. It's like a democracy of determining what people's believability is in different dimensions. And then, when we take a vote, you get two numbers. You get the average number, equal weighted, and you get believability-weighted votes. So you look at those two, and usually they're in alignment. If they're not in alignment, we do another round of discussion and voting, and then we go with believability-weighted vote, generally speaking. It's a little bit more complicated to explain how exactly it works.

Blodget: It must be really complicated because that's pretty complicated.

Dalio: Here's what I'm saying. Here's the overarching issue. We have an idea meritocracy, and it has just worked unbelievably well. It is a real idea meritocracy in which there's this radical truth and radical transparency, and it's not understood. When you're faced with a choice, you have one of three choices that you can have. You can have those with power decide. You can have one man, one vote. Or you can have believability-weighted decision-making. I'm scared of one man, one vote because it suggests that everybody has an equal ability at making decisions, and I think that's dangerous. I'm also scared of people with power making the decision. How do you know that that's meritocratic? How do you self-correct that? I would like it to be that everybody knows that person's knowledge on the subject, and we can draw upon those differences and actually have the more knowledgeable people have more weight in the decision making. That's my motivation. That's my dilemma, and that's what I work to solve. It's not perfect. I'm not saying it's perfect. But now you understand my thinking. That's where it's coming from.

Blodget: Do you ever have situations where you go through all that, and the vote comes out 51% - 49% or something awfully close?

Dalio: Yes. It constitutes a small percentage of the outcomes, but it happens. Then what we do is we try to get past it by discussing it and voting again. And then if we're still stuck with it, we'll just go with it. I really believe in people here. That process is what keeps the independent thinkers here. I must operate by that system. I think the President of the United States must operate by rules. I think our judicial system must operate by the rules. You have to operate by the rules of the system, and if you don't, if you pull rank, then you lose all your credibility. I have never overruled a decision.

Blodget: So, to be clear, what frustrates you about the Wall Street Journal article is mainly that they don't see Bridgewater's culture the way you see it?

Dalio: What I'm saying about the Wall Street Journal is I don't think they were trying to find out what's true. Some people seek to understand, and some people seek to portray what they want to portray. I don't know whether you're trying to understand or whether you have a picture that you want to portray. I came into this discussion with the general assumption that you were trying to do was to understand.

Blodget: That's what I'm trying to do. So let me stipulate for the record: Bridgewater's system obviously works for Bridgewater - and wow does it work for Bridgewater. You have 1,500 folks who can presumably work anywhere, and 95% of them say this system is helping them develop, and they like it.

Dalio: I'm saying something more. I'm saying our system would work for most of the world.

Blodget: That's another question I'm going to ask you. But first, I just want to have us agree that it is true to say that some people hear the facts and the reality of Bridgewater and regard it as, using your words, kooky, unusual and great…

Dalio: I may have used the word, "kooky." I will retract the word "kooky." What I meant by that is unusual. It's unusual, and it has produced unusually successful results - including people not wanting to work anywhere else because they love that we have an idea meritocracy.

Blodget: My question is, Ray, isn't the rich media ecosystem we have in which so many different ideas and opinions and views about reality are shared - Isn't that the "idea meritocracy" you're trying to build at Bridgewater? And isn't that fundamentally a good thing?

Dalio: Let me reply to your question with a question. Do you really believe that most of the media is trying to find out what is true, or do you believe that they are primarily trying to find facts to support their existing views?

Blodget: I believe the some journalists and media organizations are trying to find the truth. I also believe that each journalist and media organization brings a particular perspective to their work, just as each Bridgewater employee brings his or her own ideas and beliefs. But I believe the media ecosystem as a whole - with all of the diverse outlets and perspectives - combined with this amazing thing called the internet in which anybody can publish their own view - is a great tool for helping us get to the truth overall.

Dalio: Wow. We have a different perspective. And now I'm fearful about having this interview with you. Your tone suggests you're you're coming with a preconception. You're not trying to pull out and put down on paper my perspective. You're cross-examining me in a particular way that seems like what you want to do is characterize rather than seeking to understand and convey.

Foto: Business Insider

Blodget: I actually, respectfully, in the spirit of - what do you call it at Bridgewater? "Thoughtful disagreement"? - disagree.

Dalio: Well, I hope so because I think if you look at the statistics and you deal with fake media - and fake media and distorted media is a continuum - the vast majority of the population says, "I don't know what to believe." There are no checks and balances in quality control. I treasure the fact there's media freedom, but with that goes responsibility. I think that there should be a self-regulatory organization and that they should start to think about standards. Because I think a lot of people say, "I don't know how to read what is true versus somebody else's interpretation."

Blodget: If you think the current media system is broken, what is the perfect system?

Dalio: I don't think there is such a thing as the perfect system. I do think people need to recognize that a lot of journalists want to write a story a certain way because the story will be better or the portrayal will be better, or at least recognize that whenever you're looking at something, you're seeing it through somebody's eyes who may actually not be the person who is the most insightful. Like the saying goes, don't believe everything you read. Second, the motivations are not as pure or the fact checking as not as pure as some people might believe. I think that that's a threat because people don't know what's true. You have to put yourself in my shoes as a person who believes in radical truth and in being radically transparent, not in seeing things through somebody else's eyes. With the media, we don't know what's true, and we don't have radical transparency because we're seeing everything through somebody else's eyes. There's no other industry that has as much power and as much freedom and as little quality control. I can't imagine how anyone could not think that's a problem.

Blodget: Isn't the check and balance in media the existence of a free media? In which other outlets can say "Hey, that thing that they wrote was complete crap. Here's the truth"?

Dalio: Quote me on this. That sounds like saying, "Isn't the checks and balances on the financial system a free financial system?" No, I don't think individual media outlets will regulate. There are such things as self-regulatory organizations that will look at the members of the industry and their behavior and establish standards of behavior.

Blodget: So this would be a self-regulatory organization? Or a government regulator? And what would the regulator do, exactly? For example, we just had a controversial, contentious election. Different media outlets had very different views and portrayals of it. What would the regulator do in this situation?

Foto: Business Insider

Dalio: First of all, the question is, does the industry have a problem or not? Consumers believe there is a problem, and it's a pervasive problem and it's probably coming to a head. We don't know what truth is anymore. You or other media people can say there's no problem, but you're losing your credibility. There's distortion, and it's hurting our society. And as a result of that, there will be forces that, one way or another, are going to naturally bring that into equilibrium. Pendulums swing from one extreme to another extreme. As a result I would think that a self-regulatory organization would be probably the best path. I wouldn't want a government regulator because it would threaten what I treasure as the free media. Who would decide on what's quality? I think that good media people could do what other self-regulatory organizations do - say, okay, we're going to have standards. The American Motion Picture Association represented the industry in creating rating for movies. I think the industry can probably self-correct without regulation, but I don't believe an individual outlet can regulate itself.

Blodget: I would suggest that the media spends a lot of time effectively regulating itself because media organizations criticize other media organizations constantly and talk about how everybody else is getting it wrong.

Dalio: No, I'm not talking about that.

Blodget: But…

Dalio: I'm not talking about that. You want to understand me, right?

Blodget: Absolutely.

Dalio: I think the media argues all the time with itself. That's not what I'm talking about. I'm talking about something in which there's a standard of behaving. There's no regulatory body. There's no judge. There's no means of assessing whether the truth was handled in a quality way. We're stuck with, "The Wall Street Journal said this." "I said this." How do we know what's true? We're just left with two people screaming at each other.

Blodget: My perspective is there is a self-regulating system in which everyone can share their view, which is exactly what we're doing right now.

Dalio: I thought this was an interview in which you wanted to convey what I think. Is it an interview in which you want to convey what I think? Or is it an interview in which you want to convey what you think?

Blodget: It's an interview that I see as part of an overall process to understand more deeply a very interesting situation at Bridgewater and also something that you've identified as a major issue in society, which is distrust in the news media and distortion.

Dalio: Okay, so let me then be clearer in terms of your last question. Now, what you've asked, I think, is the question of, is it a self-correcting system? The answer is I don't believe it's an effective self-correcting system because I think it is in the nature of individuals and individual entities not to self-correct.

Blodget: Is what we have done today, is this what happens at Bridgewater when you are trying to get to the truth when people have different perspectives on things, and they challenge each other and ask questions of each other? Is this the process?

Dalio: It's partially the process. The way it works is that it would be a group of people who are having this conversation. I thought that this was mostly an interview to get out my perspective, but in any case, let's suppose that we have two different perspectives and instead of this being an interview that this is a disagreement. Then what we would do is have a voting process. Then you would be asked to judge whether the parties were open-minded as well as assertive. In other words, what's expected of everybody is that they're trying to find out what's true. It's even more valuable and more admired to change your opinion than to be right. But anyway, at the end of it, there needs to be a resolution of a disagreement. That resolution happens in a voting way. In order to get past disagreements, you just can't have one person with power decide. In other words, so just because I'm a boss, it would be terrible if I then said, "Okay, we're gonna go do this." That's why, after that thoughtful disagreement, there has to be a process of an idea meritocracy. That means okay, now you have to vote, not that the decision resides with power. And then you vote and move beyond it.

Blodget: Let's say that we have a group of journalists who are interested in the truth. Let's say that they try to write the truth in a fair way, and our President Elect, Donald Trump, doesn't like it or believe it. Have they done their job?

Dalio: Henry, I think the basic problem is that everybody thinks they know what the truth is, and sometimes they're even distorting the truth to make their arguments. I'm sure Donald Trump will think that he has the truth, and some journalist is arguing that he has truth, and somebody else is arguing that they have the truth. And in fact it's even worse than that because they're so hell bent on their arguments that they will distort the truth consciously. They'll manipulate the facts to support their arguments because they're so hung up in the fight. That's where the problem is, so we argue all the time. We don't have ways of resolving the arguments in idea meritocratic way.

Blodget: Given that we're having a philosophical discussion about the media, one of the problems the media struggles with is that we have to work within the attention spans and schedules of the people who are consuming media. And although I could talk about these topics all day, I think most people will have only a few minutes that they can devote to figuring out which portrayal of Bridgewater's culture is true.

Dalio: Totally. We suffer from the same problem. Therefore, I empathize. And that is what actually produced the tools we are developing at Bridgewater - the technologies that people can get at particular points so that the technology can help them sort through decisions. I'm trying to find a way so that anybody in the process can literally push a button and get an answer or see the relevant part of a tape. Forget about what the technology is. Just understand the motivation behind it. If you have the power to see things through somebody else's eyes, it's like going from black and white to color or two dimensions to three dimensions. It's shocking, and we have systems that do that. This is what's cool, and that's what I'm talking about.

Blodget: When I went out with Trump on the road early in his campaign, just talking to a lot of the folks who were coming to the rallies, it was the same thing. It's like, wow, okay. I respect and like these people, they have just a different view of this person than a lot of the people that I talk to every day. And knowing that is very helpful.

Dalio: And then how do you get past that? In other words, how do you see it through their eyes? How do they see it through your eyes? And then, at the end of the day, rather than arguing, how do you get past that and make a decision in an idea meritocratic way? That is what has made Bridgewater successful.

Blodget: Thanks, Ray. And I'm looking forward to taking Bridgewater's personality tests!

Foto: Business Insider

Read the whole story
9 days ago
Warsaw, Poland
Share this story


3 Comments and 16 Shares
It's like I've always said--people just need more common sense. But not the kind of common sense that lets them figure out that they're being condescended to by someone who thinks they're stupid, because then I'll be in trouble.
Read the whole story
10 days ago
Warsaw, Poland
Share this story
2 public comments
10 days ago
it me
Earth, Sol system, Western spiral arm
10 days ago
Feelings may not be able to achieve much, but they provide purpose. #VaguelyPhilosophical
Moses Lake, WA
Next Page of Stories