Student Loans

Summary

Podcast Summary: More or Less – "Student Loans"
Host: Tim Harford (BBC Radio 4)
Date: August 15, 2014

---

Overview

This episode of More or Less explores student loan statistics in the UK, focusing on government forecasting, the cost to taxpayers, and the increasingly generous nature of the current student loan system. Additional segments briefly analyze claims about gold extraction from mobile phones, the prevalence of paedophilia among clergy and the general population, and give an introduction to machine learning, capping off with rapid-fire analysis of current statistical headlines.

---

Key Discussion Points & Insights

1. Student Loan System: Cost Projections and RAB Charge

[00:08–08:38]

• Background:
  The UK government’s revised forecast expects to recoup less from student loans than previously estimated, amounting to an unexpected additional taxpayer cost of £1.7 billion each year.
• RAB Charge:
  • Initially, the "Resource Accounting and Budgeting" (RAB) charge was forecast at 28p in the pound (government would recoup 72p per £1 lent).
  • By 2014, it had risen to 45p, meaning only 55p per £1 lent is expected to be paid back.
• Significance:
  Every 1p increase in the RAB charge equates to £100 million more that will not be recouped annually.
• Contributing Factors:
  • Poor economic recovery leads to lower graduate salaries, meaning less and slower loan repayment.
  • Unrealistic Assumptions in government forecasting (e.g., continued historical salary growth rates for graduates, and assuming all graduates' salaries would grow uniformly by 2% annually).
  • Repayment Threshold Increase: The threshold for starting loan repayments rose from £15,000 to £21,000, dramatically reducing the number of graduates making repayments ([04:26] Gavin Conlon).
  • Discount Rate: The Treasury values cash received in the future much less than cash now, using a 2.2% annual discount rate which makes future repayments seem less valuable ([07:05]).

Notable Quotes:

• "The higher than expected RAB charge means a more generous than expected system of student loans. And, well, not everyone's going to be unhappy about that." – Tim Harford ([02:48])

• "If the government wants to reduce the RAB charge, all they have to do is slow the rate of increase on the threshold or freeze it or reduce the level of threshold." – Gavin Conlon ([04:26])

• "If we were to use a different discount rate... say instead of 2.2%, if we used 1%... this reduces the RAB charge from 40% to about 25%." – Gavin Conlon ([07:53])

• Breakdown of Takeaways:

  • RAB charge increases mean loans are effectively more subsidized.
  • The point at which the new system costs more than the old system is a RAB charge of 48.6%. At 45%, it's getting close but not quite there ([06:09]).
  • If the economy improves, graduate salaries may recover, reducing future RAB estimates ([06:28]).

---

2. Communication of Financial Data on Radio

[09:04–11:22]

• Peter Day suggests that overprecision in broadcast financial updates confuses listeners.
• He proposes audio cues ("audio emoticons") for market movement, rather than overwhelming the audience with numbers.
  • E.g., a short sound for small drops, a fanfare for a big rise.

Notable Quotes:

• "The more precise you try to be in some media, such as radio, then... the more confusing you probably are." – Peter Day ([09:04])
• "Far better to do a sound for it, for example." – Peter Day ([09:34])

---

3. Gold from Mobile Phones vs. Mining

[12:00–15:48]

• Claim Investigated: "It takes a ton of ore to get one gram of gold, but you get the same amount from recycling 41 mobile phones."
• Findings:
  • Accurate in context: 1 ton of gold ore generally yields 1–2g gold ([12:44] Dr. Dave Holwell).
  • 1 ton of recycled mobile phones can yield about 300g of gold ([14:10]), making old phones far "richer" per ton than ore.
  • Economic Reality: For an individual, the gold value per mobile phone is only about 60p and is difficult to extract ([14:22]).
  • Globally, recycling cannot keep pace with the demand fulfilled by mining—mined gold production far surpasses what's available from recycling phones ([15:03]).

---

4. Paedophilia Statistics: Contextualizing the Pope’s Comment

[15:48–22:09]

• Claim Discussed: The Pope was quoted as saying 2% of Catholic priests are paedophiles.
• Expert Analysis:
  • Dr. James Cantor:
    Estimating prevalence is difficult; definitions and methodology matter. 2% may be a reasonable guess for society in general, but for strictly defined paedophilia (prepubescent attraction), it’s closer to or below 1% ([16:54]).
  • Dr. Michael Seto:
    Surveys find 3–5% of men admit to fantasizing about children, but persistent paedophilic interest is likely 1% or lower ([18:12]).
  • Prof. Philip Jenkins:
    A 2002 John Jay study found 4.2% of US priests had been 'plausibly accused' of abuse (including both prepubescent and adolescent victims). The majority of allegations were in a six-year period, suggesting data spikes ([20:11]).
  • The true prevalence is difficult to assess, and comparisons with the general population are fraught due to differing definitions and underreporting.

---

5. Introduction to Machine Learning

[22:09–26:43]

• Anthony Goldblum (Kaggle) introduces machine learning:
  • Computers learn from past patterns in data, moving beyond hand-coding rules.
  • Early example: Arthur Samuels’ checkers program that learned to play better over time ([23:05]).
  • "Deep learning" – New techniques that can find complex patterns in data that humans can't easily parse.
• Applications:
  • Recommendation algorithms for online retailers.
  • Automation and optimization in agriculture (e.g., prescriptive planting by Monsanto).
• Impact:
  Many machine learning systems are already a part of daily life, often working behind the scenes.

---

6. Short Statistical News Headlines

[26:43–end]

• Mortgage borrowing cap reduced under the Help to Buy scheme.
• No increase in pro-independence sentiment among Scots post-Commonwealth Games.
• Black bear population in North Carolina has increased fivefold in 35 years.

---

Memorable Moments & Quotes

• On RAB charge forecasts:
  "Every time one of these numbers comes out, we're talking about a fresh 30-year prediction that may or may not come true." – Ben Carter ([02:10])
• On overprecision in statistics reporting:
  "There must be a better way of doing... rises and falls of the stock market... than the way they are done at the moment." – Peter Day ([09:04])
• On gold recycling myth:
  "At that rate, you would chew through 7 billion mobile phones in just over 23 days." – Charlotte McDonald ([15:34])
• On machine learning limitations:
  "You would never be able to program a self-driving car with if then else statements because there are just too many configurations." – Anthony Goldblum ([24:16])

---

Timestamps of Key Segments

• Student Loans and RAB Charge: [00:08–08:38]
• Peter Day on Financial Data in Broadcasting: [09:04–11:22]
• Gold Mining vs. Phone Recycling: [12:00–15:48]
• Paedophilia Statistics: The Pope’s 2% Claim: [15:48–22:09]
• Introduction to Machine Learning: [22:09–26:43]
• Statistical News Headlines: [26:43–end]

---

Final Word:
This episode unpacks some of the biggest statistical debates in current UK education finance, debunking myths, exposing flawed assumptions, and giving listeners both a sharper eye for numbers and a sense of their sometimes surprising uncertainty. The overall tone is approachable, at times irreverent, but always driven by the pursuit of clarity and truth hiding inside the numbers.

Transcript

[A]

Thank you for downloading More or less from the BBC. This is the version first broadcast on Radio 4. Here's Tim Harford.

[B]

Hello and welcome to a brand new summer series of More or Less. This week, one of the greats of economic broadcasting advises us not to get lost in over precise statistics. The Pope is reported to have said that one Catholic priest in 50 is a pedophile. We look at the facts and we investigate the strange numbers behind the unglamorous business of recycling and the alluring world of gold. But first let's talk about a government policy that's costing 1.7 billion pounds more than expected every year. Well, Ben Carter's here to help. And Ben, this is about student loans.

[C]

Yes, Tim, while we were off air, the government released a report on the student loan system. And the main headline from that report was that the government are expecting to recoup less money from student loan repayments than they had initially forecast.

[B]

And there's history here. A few years ago the coalition government floated its new system of student loans where tuition fees would be much higher but students wouldn't have to pay back their loans until they were earning more. And the first prediction was that for every pound the government lent to students, they'd only get 72 pence back. That's a loss of 28 pence.

[C]

A now that 28 pence in the pound is known as the RAB or RAB charge, but it keeps rising. In 2011 the legislation was going through parliament with an estimated RAB charge of 30 to 32 pence in the pound.

[B]

But independent analysts had higher forecasts and the latest estimate for the RAB charge is 45 pence. And Ben, I think we're safe in saying that is a lot more than 28 pence. And this is where the 1.7 billion pound cost comes from because as a rule of thumb, every penny the RAB charge goes up is another hundred million pounds less recouped from a particular year's worth of students across the lifetime of their loans. So Ben, it's probably worth taking a step back and asking what's really going on here.

[C]

Yes, Tim. Student loans are repaid or not over the course of 30 years. And each RAB number refers to a New Year's student entry. It's the government's attempt to account for how much the loans to those new is going to end up costing. Every time one of these numbers comes out, we're talking about a fresh 30 year prediction that may or may not come true. And what we're Predicting is how heavily the student loan system will end up being subsidised when the RAB goes up. That's the government expecting that students won't pay back their loans or will repay a long time in the future, meaning the cash is worth less.

[B]

So the higher than expected RAB charge means a more generous than expected system of student loans. And, well, not everyone's going to be unhappy about that. But why are the forecasts rising so much now? Partly, it's just the slow recovery. Student loan repayments are linked to income, and the graduate job market in recent years has been slow. Low graduate incomes mean small loan repayments, which is bad.

[C]

But independent forecasters also say that biz, the Department for Business Innovation and Skills, made some questionable assumptions with the way it was forecasting future graduate incomes and therefore future repayments. Here's Bharam Bhekradna of the Higher Education Policy Institute.

[D]

There were two pretty significant assumptions, one of which looked to us to be implausible, the other of which was just a mistake. The implausible one was that graduate salaries would continue to rise at the same rate as they had risen in the past 30 years. Bearing in mind that 30 years previously, the graduate population was really quite different. It didn't include nurses, for example, it didn't include social workers. So it seemed to us that the assumption that it would carry on growing at the same rate as in the past, that itself was unlikely. But what was just wrong was the assumption that the increase would be evenly spread. So if there was, for example, a 2% increase per year, which is what they were assuming in graduate salaries, all graduates would benefit from 2% increase per year. Now, that hasn't happened in the past. There's no reason to think it'll happen in the future.

[B]

That all sounds fairly sophisticated, but there was actually a bigger and rather simpler problem too.

[E]

One of the points about the reforms of higher education in 2012 was the threshold for gradual repayment, and that's £21,000. And that increase from £15,000, that's very significant. There's a lot of people that are taken out of repayment essentially because of this higher threshold for repayment. If the government wants to reduce the RAB charge, all they have to do is slow the rate of increase on the threshold or freeze it or reduce the level of threshold.

[B]

That was Gavin Conlon, a partner at London Economics, a another forecaster who consistently believed that a RAB of 28 pence was too low. Gavin told us that biz just didn't adjust properly for the fact that the threshold for repayments had risen from £15,000 to £21,000. Unsurprisingly, that made a big difference to how much money was paid back. Now the new report from BIZ has admitted that the early forecasts weren't up to scratch. BIZ is sharpening up the way the RAB charge is calculated and that's one reason why the RAB charge is rising.

[C]

And interestingly, Tim, Biz has told us that if they had known then what they know now about earnings growth and the latest earnings growth forecasts, a more appropriate threshold for repayment would have been £19,500 rather than £21,000, which Gavin Conan reckons would be an extra £400 million a year that graduates were repaying.

[B]

So the government has quite dramatically changed its mind about how generous the student loan system is to graduates. But Ben, all this raises an Awkward now the RAB charges 45% in the pound. Might the new student loan system actually end up costing the taxpayer more than the old system where graduates borrowed much less, but more of them ended up having to repay quite quickly?

[C]

We're close, but we're not quite there. The crossover point is a RAB charge of 48.6. That's the moment at which new student loans cost the treasury more than the old ones did. Since the RAB has risen from 28 to 45 over the past four years, you, you'd think will cross that Rubicon soon.

[B]

But perhaps not. Gavin Conlon of London Economics thinks that the current RAB charge of 45 is likely to fall as the job market for graduates perks up. Graduates will get richer and that means they'll be paying more of their loans back more quickly.

[C]

But there's another odd issue here, Tim. All these forecasts for the RabCharge have to take a view about what a loan repayment in 10, 20 or even 30 years is worth.

[B]

At the moment, the government can borrow really cheaply, which suggests that there's no huge benefit to getting money right now rather than getting money in a few years time.

[C]

But the treasury takes a different view, a more cautious view, you might say.

[B]

The treasury rule of thumb is that for every year that goes by, money is worth 2.2% less. The technical term for that is the discount rate. That 2.2% a year is applied to any delayed cash payment, including student loan repayments. A pound in 30 years time is worth barely 50 pence today by that reckoning, since these student loans are going to be repaid slowly, that's one reason the RAB charge is high. With a discount rate of 2.2%, the treasury just doesn't value cash in 20 or 30 years very much.

[C]

And there's probably no right answer as to what the discount rate should be. But if it was lower, closer to the actual borrowing rate, then the government forecast would be to recoup a lot more of the student loans. Here's Gavin Conlon again.

[E]

If we were to use a different discount rate, and I'm not suggesting that we should, but if we were to use a different discount rate, just say instead of 2.2%, if we used 1%, which might be a bit closer to the cost of borrowing, this reduced the RAV charge from 40% to about 25%. So in money terms, that's about one and a half billion pounds per cohort, which is not a trivial amount of money.

[B]

Ok, Ben, so let's run through that again.

[C]

First, the Government has concluded that it's going to end up being more generous than expected to recent graduates to the tune of 1.7 billion pounds a year.

[B]

Second, it probably should have realised that in advance. It was certainly told as much by other forecasters.

[C]

Third, all of these numbers are very long term forecasts, highly uncertain and very sensitive to the choice of discount rate.

[B]

And finally, if the economy continues to pick up, we can expect the forecast of student loan costs to start falling. Well, here's hoping. Thank you, Ben. And from one of the BBC greats, Ben Carter, to another Peter Day, who's been presenting business and financial news for 40 years. Over the years he's spotted something that has made him think you've been musing about the way that numbers are presented and you think there's a problem.

[F]

I think there's a paradox that the more precise you try to be in some media, such as radio, then the less precise, the more confusing you probably are. And there must be a better way of doing, oh, rises and falls of the stock market or of the power against the dollar than the way they are done at the moment. And it may be sound.

[B]

So at the moment we just have this blur of numbers. The stock market rose 5.62% and it doesn't really mean anything.

[F]

Well, if it rose, that would be a huge jump, wouldn't it? And everybody would know that. So probably the newsroom would say jumped or something. But what if it goes down two points? How do you describe it? Did it fall two points? Did it go down two points? Probably a better word is eased, but that's an interpretive Thing. And that's quite a difficult thing to get right. Far better to do a sound for it, for example.

[B]

That would be two points off the footsie, maybe.

[F]

Yeah. And then when it went down 50 points and then when it went down 500 points. And that's quite a serious observation because I think that's much more graspable to anybody than is the way we try to describe rises and falls with numbers attached to them in radio at the moment.

[B]

It certainly has an impact. What about rising.

[F]

And maybe a fanfare if it went up that much. Trumpets and pat de foie gras and that kind of thing. There's a serious thing though running here. You know, when you used to go into the old stock exchange, when there was a floor of the stock exchange, you could instantly see how busy things were by the swarm of people physically round desks on the stock exchange floor. When the stock Exchange screen had 100 share prices on it and the red ones were down and the blue ones or green ones, I can't remember now, were up, you could see at a glance what was going on far clearer than anybody actually having to explain that in words. And I think we ought to in radio sometimes try to do the oral equivalent of a graph or a display or the swarming on the stock exchange floor.

[B]

The wonderful Peter Day with his proposal for a kind of audio emoticon. We may give it a try later in the program. And by the way, you should subscribe to Peter's BBC download in business. And you can also subscribe to this program more or less at our website, BBC.co.uk more or less. Now imagine if you had a gold mine in your pocket. Sound far fetched. Well, the possibility has been brought to our attention by loyal listener Terry Ballard, who emailed more or [e-mail] with a claim that caught his eye in a European Union document about waste.

[G]

It takes a ton of ore to get one gram of gold.

[B]

But you can get the same amount

[G]

from recycling the materials in 41 mobile phones. Is that right?

[B]

Is that right? Terry asks, what, do we all have tiny little gold mines in our handbags or pockets? More or less? It's Charlotte McDonald has been prising apart her old mobile phones and shopping online for a private jet. Hello, Charlotte. Is it true our mobile phones a treasure trove.

[A]

Okay, well, let's start with the first part for comparison, that it takes a ton of ore to get one gram of gold. Now, most people's idea of gold mining is probably prospectors finding nuggets of gold but according to Dr. Dave Hulwell, who's an economic geologist from the University of Leicester, the reality of modern mining is different.

[G]

So in one of these gold rich ore deposits, you're looking at concentrations in there of one or two parts per million, and that's the equivalent of grams per ton. So this is where gold is actually really enriched in the rock. So if you were to take one of these rocks from a typical gold mine operating at 1 gram per tonne, if you took a rock from there and smashed it into a million pieces, if one of those pieces is gold and the other 999,999 of them isn't gold, it's still economic to extract that gold.

[B]

So the first part of the EU claim is roughly accurate, that there is a gram of gold or maybe 2 grams of gold in every ton of ore. But digging up and processing all that ore seems like far too much physical labour for the statistical types that work on more or less. So it'd be far easier to get our hands on 41 old mobiles. Charlotte, would they really yield a gram of gold?

[A]

Well, phones do contain tiny amounts of gold, mostly in the SIM card. Now, the EU figure comes from the unit, and the UN got it from a material technology company in Brussels called Umicore. And Umicore said, yes, the number checks out. In fact, they reckon mobile phones are an even richer source. About 1 gram of gold for every 35 mobile phones recycled. Or to look at it another way, the company says one tonne of old phones yields about 300 grams of gold.

[B]

So one tonne of ore gives you one gram of gold, but a tonne of mobile phones gives you 300 grams. This is looking better and better, Charlotte. So should we all hand in our notice and set up a recycling empire or something like that?

[A]

Well, think about it this way. At current gold prices, the amount in your mobile phone is worth about 60p. It's pretty difficult to get it out. So it's not really something that's going to make you or me rich.

[B]

I just want to go back to that comparison. We heard earlier that a ton of ore yields a gram or two of gold, but a tonne of phones yields 300 grams. Does that mean that recycling might compete with mining as a source of the world's gold?

[A]

OK, there are just over 80 million mobile phone subscriptions in the UK and almost 7 billion worldwide, so there are certainly plenty of phones around. But according to Dr. Dave Holwell, recycling just can't compete with mining.

[G]

If we look at the amount of gold that's produced from gold mining every year, the amount of gold produced is around about 2,700 tonnes per year. So that's 2.7 billion billion grams of gold being produced. And that's the equivalent of 7 million grams per day, which I think if we're using the 41 mobile phones per gram, that's the equivalent amount of gold is being produced by gold mining, as there is in 300 million mobile phones per day.

[A]

So there you go. At that rate, you would chew through 7 billion mobile phones in just over 23 days. So in the long run, precious metal recovery is at best a marginal share of the global gold market.

[B]

Which, sadly for Charlotte, but luckily for me and our listeners, means the more or Less production team will have to stick to their day jobs. Thank you, BBC wage slave Charlotte MacDonald. You're listening to More or Less, and I'm Tim Harford and if, like our loyal listener Terry Ballard, you have a statistic you'd like us to take a look at, please drop us an email at more or lessbc.co.uk earlier this summer, the Pope himself got into the statistical game. He was quoted in a somewhat disputed interview with an Italian newspaper as saying that, according to reliable data, 2% of Catholic clergymen are paedophiles. Needless to say, these comments were reported worldwide and caused a bit of a storm. But we have a few questions. Is the claim true? How would we know? What does it mean to say that someone is a paedophile and is 2% higher or lower than the male population at large? One man who can give us an idea is Dr. James Cantor, a psychologist and sexual behaviour scientist at the University of Toronto.

[H]

2% is a probably reasonable guess for the number of paedophiles in society in general. Because paedophilia is so secretive and that so few people are willing to admit it, there really is no meaningful way to get a reliable estimate. The other problem is exactly what cutoff one uses in trying to define pedophilia. It's very common, for example, for regular men to be attracted to 18 year olds and 20 year olds. Nobody would bat an eyelash. It's not unusual for a typical 16 year old to be attractive to many men, and the younger we go, the fewer and fewer men are attracted to that age group. So if we say that a pedophile is somebody attracted to 14 and under, either early puberty or pre puberty, then we're going to get a larger estimate than if we use a very strict definition and say paedophilia refers only to the attraction to prepubescent children, which is probably much lower than a percent.

[B]

Another figure that's been quoted in some of the reporting surrounding the Pope's comments is 5%. If 5% of men are paedophiles, then no wonder we hear so many reports of paedophiles in The Catholic Church, 1980s TV personalities, politicians, and all the rest. Well, the number comes from a book written by Dr. Michael Seto. He's a clinical and forensic psychologist at the Royal Ottawa Healthcare Group. So how did he arrive at that figure?

[I]

There were a number of surveys that have been completed of men in the general population asking questions about their sexual interest in children, for example, whether they had sexual fantasies about children, whether they'd had sexual thoughts about children, whether they'd engaged in sexual behavior involving children. And typically, those surveys were finding estimates such as 3, 4, or 5% of men admitted to having those kinds of thoughts, fantasies, or behavior. Now, what I said in my book was that 5% was probably the upper limit estimate for pedophilic sexual interests, because what those surveys don't include are questions about the intensity of those thoughts or fantasies, whether they were repeated or not. So someone might say yes because they once had a fantasy. But our understanding of pedophilia would be that that person recurringly has sexual thoughts and fantasies about children. And so I suggested in the book that was an upper limit. The better estimate of the prevalence of paedophilia would be lower than that. What I've said in my presentations on the topic is that if we pull together the different studies using different methodologies and different samples from North America and Europe, that a 1% prevalence seems reasonable based on those converging data.

[B]

5%, then looks like a high number. But no wonder there are conflicting claims, because there are various different ways in which we can class people as being paedophiles. Some news reports describe sex with anyone under the age of consent as paedophilia. Researchers often say it's really about attraction to prepubescent children. And of course, there's a big difference between thoughts and actions, between paedophilic desires or fantasies and actual abuse. Now, we don't know what definition or methodology the Pope had in mind, but we can take a guess that the number came from a famous study commissioned by the Catholic Church in the United States. Professor Philip Jenkins from the Institute of Studies of Religion at Baylor University in Texas has written about sexual abuse and the Catholic clergy.

[J]

In 2002, there was a big crisis within the U.S. catholic Church. And they appointed a major, very sweeping, very searching study by John Jay College of Criminal justice in New York City, which is a highly respected institution. What they did was they looked at all clergy who served in the United States between 1950 and 2002, and they found that 4.2% of those clergy had been plausibly accused of abuse. One of the problems with the study or one of the findings that throws a little bit of doubt on what the Pope was saying. The charges were not equally spread over time. There were some periods that were much more likely to produce abusive clergy than others. And one of the real oddities that we still have to explain is that 40% of all those plausible allegations over a 50 year period occurred in just six years between 1975 and 1980. So it looks as if we have, like a very unusual period there that massively inflated the overall figures.

[B]

So the John Jay College study came up with a figure of 4.2%. But their research includes not just abuse against prepubescent children, but also adolescents. If you look at the strict definition of paedophilia, then the number of paedophiles identified by this survey is actually closer to 1 or 2%. And it's that that makes Dr. Phillips think that the Pope chose his words carefully.

[J]

If he was using a different word, like, you know, abusive clergy, then I think he would be going for a higher figure. So I think he is speaking correctly. One question I would have is whether he is speaking historically based on the John Jay study or of events today. And of course, I would add one other thing, which is the United States is not a microcosm of the whole world, and conditions might be very different in European churches, in African churches, or

[B]

indeed the rest of the world. Obviously, there are still problems with the comparison. In short, it's impossible to fact check the comments attributed to the Pope. But what we can say is that there are other studies out there which reach common conclusions about the male population at large. Paedophilia is, alas, not the incredibly rare problem we'd want. You're listening to More or Less with me, Tim Harford. Now, machine learning is what all the cool kids are talking about. It's something to do with computers learning from data, which sounds like our kind of thing. Google uses it to translate web pages. Microsoft have used it to produce their Kinect gaming systems. But I've been wondering, well, what exactly is machine learning? I spoke to a chap called Anthony Goldblum, and as he explained, the story begins where all good stories begin. Trying to win at a board game.

[K]

Machine learning really just means learning off past patterns in order to learn something about the data that you have in order to apply that learning to future versions of the same outcome.

[B]

And it's called machine learning because the computer is doing it.

[K]

Exactly. When I explain machine learning, I like to reference the first use of machine learning that I'm aware of, which was a guy called Arthur Samuels who wanted to build a machine that he could play checkers against. And he started off using things like if then else statements. If you're in this situation, then do this, otherwise do this, that. On a checkers board, there's far too many possible board configurations for an if then else set of statements to be able to be well programmed.

[B]

So you can't basically just type some rules in and program the computer to follow the rules. You've got to do something else.

[K]

No, there are just too many possible rules. So what he did was he built a machine that the more games of checkers the machine played, the better it got because it would learn board configurations that correlated with victory and board configurations that would correlate with a loss. And so the more games the checkers machine played, the better it got. It eventually was able to beat not only Arthur Samuels, but a checkers grandmaster. And similarly, you know, in today's world, we see examples like the self driving car by Google. You would never be able to program a self driving car with if then else statements because there are just too many configurations. So a machine learning approach means that the more the car drives, the more patterns it learns, the more reliably it's able to drive.

[B]

Anthony Goldblum was so impressed by the prospect of using machine learning to solve real world problems that he founded a website called Kaggle, which we've featured before on more or less. Kaggle hosts competitions where people compete to solve statistical problems involving big data sets. And as Anthony explains, the competitions are increasingly being won by computer techniques where humans can barely begin to understand how the data are structured.

[K]

There's a new set of techniques coming that we've seen win a couple of Kaggle competitions called deep learning. And the whole promise of deep learning is to try and get away from having to rely on a human to creatively come up with different ways to slice and dice the data. Now, so far, we've only seen deep learning be effective in situations where a human can't intuitively understand the data. So things like Geoffrey Hinton, one of the pioneers of deep learning, won a Kaggle competition to take chemical compounds and predict which drugs were going to be toxic and which weren't. And the data on the chemical compounds are not very human readable. The same goes for pixel representations of images as well as audio data.

[B]

When do you think we're going to see the implications of this affect daily life? Or maybe it's already affecting our daily lives?

[K]

I think it already is. Kaggle's been involved in building the recommendation algorithm, for instance, for neimanmarcus.com's.

[B]

This is a department store.

[K]

Yeah, yeah, Department store in the US. If you go on to neimanmarcus.com and you've got a history of past patterns of what you've purchased, you'll be recommended items that you may want to buy in the future. Shoppers at Amazon have probably seen similar. So there are already ways in which this sort of technology is being used. Also behind the scenes. Monsanto, for instance, just bought a company called the Climate Corporation that helps do what they call prescriptive planting. Predict, you know, what sort of seed you should plant, given the weather, given the soil, given all sorts of different readings. So although we as consumers don't experience the output of Monsanto and the Climate Corporation in our everyday lives, it is inadvertently affecting us in terms of lower prices, potentially for fruit and vegetables.

[B]

Anthony Goldblum of Kaggle. That's almost all we have time for this week. But before we go, a few statistical headlines presented with help from Peter Day. Britain's biggest mortgage lender, Lloyds Banking Group, has cut the amount house buyers can borrow under the Government's Help to buy scheme by 70%, from £500,000 to £150,000. Scottish success at the Commonwealth Games hasn't boosted support for the nation's independence campaign. A poll published on the day the Games finished showed support for independence down by 1 percentage point, to 40%. And the number of black bears currently residing in North Carolina habitats is on the rise. According to recent figures, in the last 35 years, the number of bears has risen from just 2,000 to 10,000. Please get in touch with your questions and comments via more or lessbc.co.uk Our website address is BBC.co.uk more or less, where you can download as many episodes as you like and listen to us incessantly in a slightly obsessive fashion, or we'll be back on air next week. Goodbye.

[A]

More or Less was presented by Tim Harford, the Financial Times undercover economist. The producer was James Fletcher and the programme is made in association with the Open University. You can download many more programmes for free from the Radio 4 website. You can find [e-mail] Radio 4.