Currently, blood test readings are based on one-size-fits-all reference intervals that don’t account for individual differences.
| Photo Credit: Getty Images

If you’ve ever had a doctor order a blood test for you, chances are that they ran a complete blood count, or CBC. One of the most common blood tests in the world, CBC tests are run billions of times each year to diagnose conditions and monitor patients’ health.

But despite the test’s ubiquity, the way clinicians interpret and use it in the clinic is often less precise than ideal. Currently, blood test readings are based on one-size-fits-all reference intervals that don’t account for individual differences.

I am a mathematician at the University of Washington School of Medicine, and my team studies ways to use computational tools to improve clinical blood testing. To develop better ways to capture individual patient definitions of “normal” lab values, my colleagues and I in the Higgins Lab at Harvard Medical School examined 20 years of blood count tests from tens of thousands of patients from both the East and West coasts.

In our newly published research, we used machine learning to identify healthy blood count ranges for individual patients and predict their risk of future disease.

Clinical tests and complete blood counts

Many people commonly think of clinical tests as purely diagnostic. For example, a COVID-19 or a pregnancy test comes back as either positive or negative, telling you whether you have a particular condition. However, most tests don’t work this way. Instead, they measure a biological trait that your body continuously regulates up and down to stay within certain bounds.

Your complete blood count is also a continuum. The CBC test creates a detailed profile of your blood cells – such as how many red blood cells, platelets and white blood cells are in your blood. These markers are used every day in nearly all areas of medicine.

For example, hemoglobin is an iron-containing protein that allows your red blood cells to carry oxygen. If your hemoglobin levels are low, it might mean you are iron deficient.

Platelets are cells that help form blood clots and stop bleeding. If your platelet count is low, it may mean you have some internal bleeding and your body is using platelets to help form blood clots to plug the wound.

White blood cells are part of your immune system. If your white cell count is high, it might mean you have an infection and your body is producing more of these cells to fight it off.

Normal ranges and reference intervals

But this all raises the question: What actually counts as too high or too low on a blood test?

Traditionally, clinicians determine what are called reference intervals by measuring a blood test in a range of healthy people. They usually take the middle 95% of these healthy values and call that “normal,” with anything above or below being too low or high. These normal ranges are used nearly everywhere in medicine.

But reference intervals face a big challenge: What’s normal for you may not be normal for someone else.

Nearly all blood count markers are heritable, meaning your genetics and environment determine much of what the healthy value for each marker would be for you.

At the population level, for example, a normal platelet count is approximately between 150 and 400 billion cells per liter of blood. But your body may want to maintain a platelet count of 200 – a value called your set point. This means your normal range might only be 150 to 250.

Differences between a patient’s true normal range and the population-based reference interval can create problems for doctors. They may be less likely to diagnose a disease if your set point is far from a cutoff. Conversely, they may run unnecessary tests if your set point is too close to a cutoff.

Defining what’s normal for you

Luckily, many patients get blood counts each year as part of routine checkups. Using machine learning models, my team and I were able to estimate blood count set points for over 50,000 patients based on their history of visits to the clinic. This allowed us to study how the body regulates these set points and to test whether we can build better ways of personalizing lab test readings.

Over multiple decades, we found that individual normal ranges were about three times smaller than at the population level. For example, while the “normal” range for the white blood cell count is around 4.0 to 11.0 billion cells per liter of blood, we found that most people’s individual ranges were much narrower, more like 4.5 to 7, or 7.5 to 10. When we used these set points to interpret new test results, they helped improve diagnosis of diseases such as iron deficiency, chronic kidney disease and hypothyroidism. We could note when someone’s result was outside their smaller personal range, potentially indicating an issue, even if the result was within the normal range for the population overall.

The set points themselves were strong indicators for future risk of developing a disease. For example, patients with high white blood cell set points were more likely to develop Type 2 diabetes in the future. They were also nearly twice as likely to die of any cause compared with similar patients with low white cell counts. Other blood count markers were also strong predictors of future disease and mortality risk.

In the future, doctors could potentially use set points to improve disease screening and how they interpret new test results. This is an exciting avenue for personalized medicine: to use your own medical history to define what exactly healthy means for you.

This article is republished from The Conversation under a Creative Commons license. Read the original article here.

A new machine-learning weather prediction model called GenCast can outperform the best traditional forecasting systems in at least some situations, according to a new paper.
| Photo Credit: NASA/GSFC, MODIS Rapid Response Team, Jacques Descloitres

A new machine-learning weather prediction model called GenCast can outperform the best traditional forecasting systems in at least some situations, according to a paper by Google DeepMind researchers published today in Nature.

Using a diffusion model approach similar to artificial intelligence (AI) image generators, the system generates multiple forecasts to capture the complex behaviour of the atmosphere. It does so with a fraction of the time and computing resources required for traditional approaches.

How weather forecasts work

The weather predictions we use in practice are produced by running multiple numerical simulations of the atmosphere.

Each simulation starts from a slightly different estimate of the current weather. This is because we don’t know exactly what the weather is at this instant everywhere in the world. To know that, we would need sensor measurements everywhere.

These numerical simulations use a model of the world’s atmosphere divided into a grid of three-dimensional blocks. By solving equations describing the fundamental physical laws of nature, the simulations predict what will happen in the atmosphere.

Known as general circulation models, these simulations need a lot of computing power. They are usually run at high-performance supercomputing facilities.

Machine-learning the weather

The past few years have seen an explosion in efforts to produce weather prediction models using machine learning. Typically, these approaches don’t incorporate our knowledge of the laws of nature the way general circulation models do.

Most of these models use some form of neural network to learn patterns in historical data and produce a single future forecast. However, this approach produces predictions that lose detail as they progress into the future, gradually becoming “smoother”. This smoothness is not what we see in real weather systems.

Researchers at Google’s DeepMind AI research lab have just published a paper in Nature describing their latest machine-learning model, GenCast.

GenCast mitigates this smoothing effect by generating an ensemble of multiple forecasts. Each individual forecast is less smooth, and better resembles the complexity observed in nature.

The best estimate of the actual future then comes from averaging the different forecasts. The size of the differences between the individual forecasts indicates how much uncertainty there is.

According to the GenCast paper, this probabilistic approach creates more accurate forecasts than the best numerical weather prediction system in the world – the one at the European Centre for Medium-Range Weather Forecasts.

Generative AI – for weather

GenCast is trained on what is called reanalysis data from the years 1979 to 2018. This data is produced by the kind of general circulation models we talked about earlier, which are additionally corrected to resemble actual historical weather observations to produce a more consistent picture of the world’s weather.

The GenCast model makes predictions of several variables such as temperature, pressure, humidity and wind speed at the surface and at 13 different heights, on a grid that divides the world up into 0.25-degree regions of latitude and longitude.

GenCast is what is called a “diffusion model”, similar to AI image generators. However, instead of taking text and producing an image, it takes the current state of the atmosphere and produces an estimate of what it will be like in 12 hours.

This works by first setting the values of the atmospheric variables 12 hours into the future as random noise. GenCast then uses a neural network to find structures in the noise that are compatible with the current and previous weather variables. An ensemble of multiple forecasts can be generated by starting with different random noise.

Forecasts are run out to 15 days, taking 8 minutes on a single processor called a tensor processor unit (TPU). This is significantly faster than a general circulation model. The training of the model took five days using 32 TPUs.

Machine-learning forecasts could become more widespread in the coming years as they become more efficient and reliable.

However, classical numerical weather prediction and reanalysed data will still be required. Not only are they needed to provide the initial conditions for the machine learning weather forecasts, they also produce the input data to continually fine-tune the machine learning models.

What about the climate?

Current machine learning weather forecasting systems are not appropriate for climate projections, for three reasons.

Firstly, to make weather predictions weeks into the future, you can assume that the ocean, land and sea ice won’t change. This is not the case for climate predictions over multiple decades.

Secondly, weather prediction is highly dependent on the details of the current weather. However, climate projections are concerned with the statistics of the climate decades into the future, for which today’s weather is irrelevant. Future carbon emissions are the greater determinant of the future state of the climate.

Thirdly, weather prediction is a “big data” problem. There are vast amounts of relevant observational data, which is what you need to train a complex machine learning model.

Climate projection is a “small data” problem, with relatively little available data. This is because the relevant physical phenomena (such as sea levels or climate drivers such as the El Niño–Southern Oscillation) evolve much more slowly than the weather.

There are ways to address these problems. One approach is to use our knowledge of physics to simplify our models, meaning they require less data for machine learning.

Another approach is to use physics-informed neural networks to try to fit the data and also satisfy the laws of nature. A third is to use physics to set “ground rules” for a system, then use machine learning to determine the specific model parameters.

Machine learning has a role to play in the future of both weather forecasting and climate projections. However, fundamental physics – fluid mechanics and thermodynamics – will continue to play a crucial role.

This article is republished from The Conversation under a Creative Commons license. Read the original article here.

The Reserve Bank Innovation Hub (RBIH), which is the innovation arm of the Reserve Bank of India (RBI), is making giant strides in the fight against financial fraud through the promotion of the use of an advanced AI tool called MuleHunter.AI. This technology identifies and flags mule accounts, commonly used in money laundering schemes.

The application of MuleHunter.AI has already been successfully demonstrated in two public sector banks. Data from the National Crime Records Bureau (NCRB) indicate that online financial frauds are responsible for 67.8% of all complaints related to cybercrime. This makes the effective provision of fraud prevention AI tools highly urgent.

One of the biggest problems in fighting financial fraud is the exploitation of money mule accounts. These accounts are a key enabler of illicit financial activities; hence, tools like MuleHunter.AI is of paramount importance to protect the financial ecosystem and curb cybercrime.

What is a money mule account?

According to RBIH, mule account is a bank account used by criminals to launder illicit funds, often set up by unsuspecting individuals lured by promises of easy money or coerced into participation. The transfer of funds through these highly interconnected accounts makes it difficult to trace and recover the funds.

The Development of MuleHunter.AI

According to the Reserve Bank Innovation Hub, the department has conducted extensive consultations with banks to understand the existing methods and processes employed to identify and report these money mule accounts. The static rule-based systems used to detect mule accounts result in high false positives and longer turnaround times, causing many such accounts to remain undetected.

After working with several banks to analyse nineteen different patterns of mule account activity, the platform was created. Mulehunter.Ai’s initial results demonstrate notable gains in efficiency and accuracy.

How Mulehunter.Ai works

This in-house AI/ML-based solution is better suited than a rule-based system to identify suspected mule accounts. Advanced ML algorithms can analyse transaction and account detail-related datasets to predict mule accounts with higher accuracy and greater speed than typical rule-based systems.

The purpose of RBIH’s AI platform is to speed up the identification of fraudulent accounts. Frauds can happen through a variety of channels, and they are no longer little incidents; they are becoming big day by day. The best approach would be to look at where the money eventually goes-to mule accounts. This machine learning-based approach has enabled the detection of more mule accounts within a bank’s system.

Alorica, a company in Irvine, California, that runs customer-service centers around the world, has introduced an artificial intelligence translation tool that lets its representatives talk with customers who speak 200 different languages and 75 dialects.

So an Alorica representative who speaks, say, only Spanish can field a complaint about a balky printer or an incorrect bank statement from a Cantonese speaker in Hong Kong. Alorica wouldn’t need to hire a rep who speaks Cantonese.

Such is the power of AI. And, potentially, the threat: Perhaps companies won’t need as many employees — and will slash some jobs — if chatbots can handle the workload instead. But the thing is, Alorica isn’t cutting jobs. It’s still hiring aggressively.

The experience at Alorica — and at other companies, including furniture retailer IKEA — suggests that AI may not prove to be the job killer that many people fear. Instead, the technology might turn out to be more like breakthroughs of the past — the steam engine, electricity, the internet: That is, eliminate some jobs while creating others. And probably making workers more productive in general, to the eventual benefit of themselves, their employers and the economy.

Nick Bunker, an economist at the Indeed Hiring Lab, said he thinks AI “will affect many, many jobs — maybe every job indirectly to some extent. But I don’t think it’s going to lead to, say, mass unemployment. We have seen other big technological events in our history, and those didn’t lead to a large rise in unemployment. Technology destroys but also creates. There will be new jobs that come about.’’

At its core, artificial intelligence empowers machines to perform tasks previously thought to require human intelligence. The technology has existed in early versions for decades, having emerged with a problem-solving computer program, the Logic Theorist, built in the 1950s at what’s now Carnegie Mellon University. More recently, think of voice assistants like Siri and Alexa. Or IBM’s chess-playing computer, Deep Blue, which managed to beat the world champion Garry Kasparov in 1997.

AI burst into public consciousness in 2022 when OpenAI introduced ChatGPT, the generative AI tool that can conduct conversations, write computer code, compose music, craft essays and supply endless streams of information. The arrival of generative AI has raised worries that chatbots will replace freelance writers, editors, coders, telemarketers, customer service reps, paralegals and many more.

“AI is going to eliminate a lot of current jobs, and this is going to change the way that a lot of current jobs function,” Sam Altman, the CEO of OpenAI, said in a discussion at the Massachusetts Institute of Technology in May.

Yet the widespread assumption that AI chatbots will inevitably replace service workers, the way physical robots took many factory and warehouse jobs, isn’t becoming reality in any widespread way — not yet, anyway. And maybe it never will.

The White House Council of Economic Advisers said last month that it found “little evidence that AI will negatively impact overall employment.’’ The advisers noted that history shows technology typically makes companies more productive, speeding economic growth and creating new types of jobs in unexpected ways.

They cited a study this year led by David Autor, a leading MIT economist: It concluded that 60% of the jobs Americans held in 2018 didn’t even exist in 1940, having been created by technologies that emerged only later.

The outplacement firm Challenger, Gray & Christmas, which tracks job cuts, said it has yet to see much evidence of layoffs that can be attributed to labor-saving AI.

“I don’t think we’ve started seeing companies saying they’ve saved lots of money or cut jobs they no longer need because of this,’’ said Andy Challenger, who leads the firm’s sales team. “That may come in the future. But it hasn’t played out yet.’’

At the same time, the fear that AI poses a serious threat to some categories of jobs isn’t unfounded.

Consider Suumit Shah, an Indian entrepreneur who caused a uproar last year by boasting that he had replaced 90% of his customer support staff with a chatbot named Lina. The move at Shah’s company, Dukaan, which helps customers set up e-commerce sites, shrank the response time to an inquiry from 1 minute, 44 seconds to “instant.” It also cut the typical time needed to resolve problems from more than two hours to just over three minutes.

“It’s all about AI’s ability to handle complex queries with precision,” Mr. Shah said by email. The cost of providing customer support, he said, fell by 85%.

“Tough? Yes. Necessary? Absolutely,’’ Mr. hah posted on X.

Dukaan has expanded its use of AI to sales and analytics. “The tools,” Mr. Shah said, “keep growing more powerful.”

“It’s like upgrading from a Corolla to a Tesla,” he said. “What used to take hours now takes minutes. And the accuracy is on a whole new level.”

Similarly, researchers at Harvard Business School, the German Institute for Economic Research and London’s Imperial College Business School found in a study last year that job postings for writers, coders and artists tumbled within eight months of the arrival of ChatGPT.

A 2023 study by researchers at Princeton University, the University of Pennsylvania and New York University concluded that telemarketers and teachers of English and foreign languages held the jobs most exposed to ChatGPT-like language models. But being exposed to AI doesn’t necessarily mean losing your job to it. AI can also do the drudge work, freeing up people to do more creative tasks.

The Swedish furniture retailer IKEA, for example, introduced a customer-service chatbot in 2021 to handle simple inquiries. Instead of cutting jobs, IKEA retrained 8,500 customer-service workers to handle such tasks as advising customers on interior design and fielding complicated customer calls.

Chatbots can also be deployed to make workers more efficient, complementing their work rather than eliminating it. A study by Erik Brynjolfsson of Stanford University and Danielle Li and Lindsey Raymond of MIT tracked 5,200 customer-support agents at a Fortune 500 company who used a generative AI-based assistant. The AI tool provided valuable suggestions for handling customers. It also supplied links to relevant internal documents.

Those who used the chatbot, the study found, proved 14% more productive than colleagues who didn’t. They handled more calls and completed them faster. The biggest productivity gains — 34% — came from the least-experienced, least-skilled workers.

At an Alorica call center in Albuquerque, New Mexico, one customer-service rep had been struggling to gain access to the information she needed to quickly handle calls. After Alorica trained her to use AI tools, her “handle time’’ — how long it takes to resolve customer calls — fell in four months by an average of 14 minutes a call to just over seven minutes.

Over a period of six months, the AI tools helped one group of 850 Alorica reps reduce their average handle time to six minutes, from just over eight minutes. They can now field 10 calls an hour instead of eight — an additional 16 calls in an eight-hour day.

Alorica agents can use AI tools to quickly access information about the customers who call in — to check their order history, say, or determine whether they had called earlier and hung up in frustration.

Suppose, said Mike Clifton, Alorica’s co-CEO, a customer complains that she received the wrong product. The agent can “hit replace, and the product will be there tomorrow,” he said. ” ‘Anything else I can help you with? No?’ Click. Done. Thirty seconds in and out.’’

Now the company is beginning to use its Real-time Voice Language Translation tool, which lets customers and Alorica agents speak and hear each other in their own languages.

“It allows (Alorica reps) to handle every call they get,” said Rene Paiz, a vice president of customer service. “I don’t have to hire externally’’ just to find someone who speaks a specific language.

Yet Alorica isn’t cutting jobs. It continues to seek hires — increasingly, those who are comfortable with new technology.

“We are still actively hiring,’’ Ms. Paiz says. “We have a lot that needs to be done out there.’’

Alorica, a company in Irvine, California, that runs customer-service centers around the world, has introduced an artificial intelligence translation tool that lets its representatives talk with customers who speak 200 different languages and 75 dialects.

So an Alorica representative who speaks, say, only Spanish can field a complaint about a balky printer or an incorrect bank statement from a Cantonese speaker in Hong Kong. Alorica wouldn’t need to hire a rep who speaks Cantonese.

Such is the power of AI. And, potentially, the threat: Perhaps companies won’t need as many employees — and will slash some jobs — if chatbots can handle the workload instead. But the thing is, Alorica isn’t cutting jobs. It’s still hiring aggressively.

The experience at Alorica — and at other companies, including furniture retailer IKEA — suggests that AI may not prove to be the job killer that many people fear. Instead, the technology might turn out to be more like breakthroughs of the past — the steam engine, electricity, the internet: That is, eliminate some jobs while creating others. And probably making workers more productive in general, to the eventual benefit of themselves, their employers and the economy.

Nick Bunker, an economist at the Indeed Hiring Lab, said he thinks AI “will affect many, many jobs — maybe every job indirectly to some extent. But I don’t think it’s going to lead to, say, mass unemployment. We have seen other big technological events in our history, and those didn’t lead to a large rise in unemployment. Technology destroys but also creates. There will be new jobs that come about.’’

At its core, artificial intelligence empowers machines to perform tasks previously thought to require human intelligence. The technology has existed in early versions for decades, having emerged with a problem-solving computer program, the Logic Theorist, built in the 1950s at what’s now Carnegie Mellon University. More recently, think of voice assistants like Siri and Alexa. Or IBM’s chess-playing computer, Deep Blue, which managed to beat the world champion Garry Kasparov in 1997.

AI burst into public consciousness in 2022 when OpenAI introduced ChatGPT, the generative AI tool that can conduct conversations, write computer code, compose music, craft essays and supply endless streams of information. The arrival of generative AI has raised worries that chatbots will replace freelance writers, editors, coders, telemarketers, customer service reps, paralegals and many more.

“AI is going to eliminate a lot of current jobs, and this is going to change the way that a lot of current jobs function,” Sam Altman, the CEO of OpenAI, said in a discussion at the Massachusetts Institute of Technology in May.

Yet the widespread assumption that AI chatbots will inevitably replace service workers, the way physical robots took many factory and warehouse jobs, isn’t becoming reality in any widespread way — not yet, anyway. And maybe it never will.

The White House Council of Economic Advisers said last month that it found “little evidence that AI will negatively impact overall employment.’’ The advisers noted that history shows technology typically makes companies more productive, speeding economic growth and creating new types of jobs in unexpected ways.

They cited a study this year led by David Autor, a leading MIT economist: It concluded that 60% of the jobs Americans held in 2018 didn’t even exist in 1940, having been created by technologies that emerged only later.

The outplacement firm Challenger, Gray & Christmas, which tracks job cuts, said it has yet to see much evidence of layoffs that can be attributed to labor-saving AI.

“I don’t think we’ve started seeing companies saying they’ve saved lots of money or cut jobs they no longer need because of this,’’ said Andy Challenger, who leads the firm’s sales team. “That may come in the future. But it hasn’t played out yet.’’

At the same time, the fear that AI poses a serious threat to some categories of jobs isn’t unfounded.

Consider Suumit Shah, an Indian entrepreneur who caused a uproar last year by boasting that he had replaced 90% of his customer support staff with a chatbot named Lina. The move at Shah’s company, Dukaan, which helps customers set up e-commerce sites, shrank the response time to an inquiry from 1 minute, 44 seconds to “instant.” It also cut the typical time needed to resolve problems from more than two hours to just over three minutes.

“It’s all about AI’s ability to handle complex queries with precision,” Mr. Shah said by email. The cost of providing customer support, he said, fell by 85%.

“Tough? Yes. Necessary? Absolutely,’’ Mr. hah posted on X.

Dukaan has expanded its use of AI to sales and analytics. “The tools,” Mr. Shah said, “keep growing more powerful.”

“It’s like upgrading from a Corolla to a Tesla,” he said. “What used to take hours now takes minutes. And the accuracy is on a whole new level.”

Similarly, researchers at Harvard Business School, the German Institute for Economic Research and London’s Imperial College Business School found in a study last year that job postings for writers, coders and artists tumbled within eight months of the arrival of ChatGPT.

A 2023 study by researchers at Princeton University, the University of Pennsylvania and New York University concluded that telemarketers and teachers of English and foreign languages held the jobs most exposed to ChatGPT-like language models. But being exposed to AI doesn’t necessarily mean losing your job to it. AI can also do the drudge work, freeing up people to do more creative tasks.

The Swedish furniture retailer IKEA, for example, introduced a customer-service chatbot in 2021 to handle simple inquiries. Instead of cutting jobs, IKEA retrained 8,500 customer-service workers to handle such tasks as advising customers on interior design and fielding complicated customer calls.

Chatbots can also be deployed to make workers more efficient, complementing their work rather than eliminating it. A study by Erik Brynjolfsson of Stanford University and Danielle Li and Lindsey Raymond of MIT tracked 5,200 customer-support agents at a Fortune 500 company who used a generative AI-based assistant. The AI tool provided valuable suggestions for handling customers. It also supplied links to relevant internal documents.

Those who used the chatbot, the study found, proved 14% more productive than colleagues who didn’t. They handled more calls and completed them faster. The biggest productivity gains — 34% — came from the least-experienced, least-skilled workers.

At an Alorica call center in Albuquerque, New Mexico, one customer-service rep had been struggling to gain access to the information she needed to quickly handle calls. After Alorica trained her to use AI tools, her “handle time’’ — how long it takes to resolve customer calls — fell in four months by an average of 14 minutes a call to just over seven minutes.

Over a period of six months, the AI tools helped one group of 850 Alorica reps reduce their average handle time to six minutes, from just over eight minutes. They can now field 10 calls an hour instead of eight — an additional 16 calls in an eight-hour day.

Alorica agents can use AI tools to quickly access information about the customers who call in — to check their order history, say, or determine whether they had called earlier and hung up in frustration.

Suppose, said Mike Clifton, Alorica’s co-CEO, a customer complains that she received the wrong product. The agent can “hit replace, and the product will be there tomorrow,” he said. ” ‘Anything else I can help you with? No?’ Click. Done. Thirty seconds in and out.’’

Now the company is beginning to use its Real-time Voice Language Translation tool, which lets customers and Alorica agents speak and hear each other in their own languages.

“It allows (Alorica reps) to handle every call they get,” said Rene Paiz, a vice president of customer service. “I don’t have to hire externally’’ just to find someone who speaks a specific language.

Yet Alorica isn’t cutting jobs. It continues to seek hires — increasingly, those who are comfortable with new technology.

“We are still actively hiring,’’ Ms. Paiz says. “We have a lot that needs to be done out there.’’

Reserve Bank of India Governor Shaktikanta Das. File
| Photo Credit: PTI

RBI Governor Shaktikanta Das on Friday emphasised on the need to eliminate biases in algorithms as the use of artificial intelligence (AI) and machine learning (ML) is on the rise.

Delivering the inaugural address at the 18th Statistics Day Conference organised by the RBI, he said the use of statistics had been ever growing as a preferred tool for drawing inferences in diverse fields and the discipline had moved beyond collection of facts to focusing more on interpretation and drawing inferences, taking into account the level of uncertainty.

The Reserve Bank of India (RBI) has ventured into AI/ML analytics in multiple areas. Under the RBI’s aspirational goals for RBI@100, Mr. Das said the central bank was aiming to develop cutting-edge systems for high frequency and real-time data monitoring and analysis.

This illustration compares three flowers – summer snowflake (A, B), blue phlox (C, D), and a blue violet (D, E) – in honeybee false colour (left) and human-visible colours (right).
| Photo Credit: Vasas V, et al., 2024, PLOS Biology, CC-BY 4.0

To most people, leaves are green and oranges are orange. But if our pets could speak, they’d disagree.

We know there are many different ways to ‘see’ the world because that’s the diversity we have found in animals. Organisms with the ability to see have two or more eyes that capture light reflected by different surfaces in their surroundings and turn it into visual cues. But while all eyes have this common purpose, the specialised cells that respond to the light, called photoreceptors, are unique to each animal.

For instance, human eyes can only detect wavelengths of light between 380 and 700 nanometres (nm); this is the visible range. Honey bees and many birds on the other hand can also ‘see’ ultraviolet light (10-400 nm).

While the human visual range is relatively limited, it hasn’t abated humans’ curiosity about how animals see the world.

Thankfully we don’t have to imagine too much. Researchers at the University of Sussex and the George Mason University (GMU) in the U.S. have put together a new camera with the ability to view the world like animals do. In a paper published in PLoS Biology, the team has written their device can even reveal what colours different animals see in motion, which hasn’t been possible so far.

Making the invisible visible

Animals use colours to intimidate their predators, entice mates or conceal themselves. Detecting variations in colours is thus essential to an animal’s survival. Animals have evolved to develop highly sensitive photoreceptors that can detect light of ultraviolet and infrared wavelengths; many even notice polarised light as part of their Umwelt – the biological systems that make a specific system of meaning-making and communication possible.

Neither human eyes nor most commercial cameras have been able to tap into this unchartered territory of animal vision. In the new study, exponents of biology, computer vision, and programming came together to create a tool that could record and track the complexity of animal visual signalling.

The tool combined existing multispectral photography techniques with a new camera setup and a beam-splitter (to separate ultraviolet and visible light), all encased in a custom 3D-printed unit. The system recorded videos simultaneously in visible and ultraviolet channels in  natural lighting. They fed the camera output through some code (written in Python) that could convert the visual data to the physical signals produced by photoreceptor cells.

Finally, the researchers modified these signals based on what they already knew about how an animal’s photoreceptors work, and produced videos true to what that animal might see. These used false colours in these videos so that, for example, a particular colour could stand in to show ultraviolet imagery.

In sum, the camera system translated what animals see in visible and non-visible light into colours compatible with the human eye.

The time challenge

You may have already seen false-colour images – like when you saw the Hubble space telescope’s iconic snap of the ‘Pillars of Creation’. The stars and nebulae don’t actually look that resplendent to human eyes. They are coloured that way to show what the telescope saw in, say, infrared or radio wavelengths. Scientists have also used false-colour images to understand how flowers reflect ultraviolet light to influence the behaviour of insects nearby.

But false colours can only stand in for so much. According to the researchers, existing techniques to visualise the colours animals see require object-reflected light to predict how an animal’s photoreceptor would respond or require a series of photographs in wavelengths beyond human vision (with the help of bandpass optical filters). Both scenarios require the subject to be motionless. The new system can visualise free-living organisms in their natural settings, however.

In addition, Pavan Kumar Reddy Katta, a graduate teaching assistant at GMU and one of the study’s authors, said the team wrote a program that could accept both ultraviolet- and visible-light data and spit out complete videos. “We made use of a continuous stream which allowed us to resolve our data at various points of space and time and produce real-time visualisations in animal-vision,” he told this author.

The next big thing in animal vision

Equipped with the new camera, the research team checked what the flower black-eyed Susan (Rudbeckia hirta) looks like to honey bees (Apis mellifera).

“To our eye, the black-eyed Susan appears entirely yellow because in the human-visible range, it reflects primarily long wavelength light,” the team wrote in its paper. “Whereas in the bee false colour image, the distal petals appear magenta because they also reflect ultraviolet, stimulating both the ultraviolet-sensitive photoreceptors … and those sensitive to green light … By contrast, the central portion of the petals does not reflect ultraviolet and therefore appears red.”

According to the paper, the visual mechanisms animals have evolved to communicate and protect themselves could help solve many of our detection problems. For example, the animal-vision video could help people navigate wild landscapes better and without hurting camouflaged animals. It can help farmers spot fruit pests that are not visible to the human eye but are readily visible to animals that have evolved to eat those fruits.

Daniel Hanley, assistant professor at GMU and the study’s corresponding author, said their invention could even transform the way wildlife documentary films are made. The camera system could allow filmmakers and ecologists to record the animal world through a new lens and create new visual experiences. He also said the platform’s striking images could be used to communicate the science of the living world to young audiences.

“We are thinking of creating a science exhibit for children using our setup, flowers, and live animals,” Dr. Hanley said. “Where children can just click a button to experience what a snake might see or a honeybee might see.”

Sanjukta Mondal is a chemist-turned-science-writer with experience in writing popular science articles and scripts for STEM YouTube channels.

Intelligence is the capacity of living beings to apply what they know to solve problems. ‘Artificial intelligence’ (AI) is intelligence in a machine. There is currently no one definition of AI.

A simple place to begin is with AI’s materiality, as a machine-software combination.

What does the machine do?

A simple example-problem in AI is linear separability. You plot some points on a graph and then find a way to draw a straight line through the graph such that it divides the points into two distinct groups.

Let’s make this problem more abstract. For example, how would a machine differentiate between a cat and a dog?

Say you give the machine 1,000 pictures of cats and 1,000 pictures of dogs, and ask it to separate them. (This task is usually not given to a linear classifier but it illustrates a point.) You also equip the machine with tools — say, a camera and an app that can measure distances of different parts of an image, can analyse depth (using trigonometry), and can assess colours.

The machine can proceed by classifying the cat- and dog-pictures in different ways, say, by shape of the face, shape of the eyes, shape of the paw, body size, size of the tongue, fur colours, etc. Because the machine has the necessary computing power, it can plot these features two at a time on a graph. For example, the x-axis can represent the slope of the face and the y-axis the length of the paw. Or it can plot them three at a time in a 3D graph.

In all these cases, you watch until the machine has found a way to separate the pictures into two groups such that one group is mostly cats and the other is mostly dogs. At this point, you stop the machine.

How hard is decision-making?

Sometimes, it’s very easy to separate a given dataset into two pieces, like with the marbles, where you can make very reliable decisions with just one dimension, or parameter. Sometimes it’s more difficult, like with the cats and dogs, where you may need around a dozen parameters.

Sometimes it is difficult — like asking the computer on a driverless car to determine whether it should apply the brake based on how fast a bird is flying in front of the car. The set of outcomes on one side of the line stand for ‘no’ and the outcomes on the other side stand for ‘yes’ — and solving for this will require hundreds of parameters.

They will also have to account for the context of decision-making. For example, if the person in the car is in a hurry to get to a hospital, is killing the bird okay? Or if the person in the car is not in a hurry, how quickly should the car brake? Etc.

Sometimes it’s just mind-boggling. For example, ChatGPT is able to accept an input question from a user, make ‘sense’ of it, and answer accordingly. This ‘sense’ comes from its training corpus — the billions of sequences of words and sentences scraped from the internet.

In particular, ChatGPT learnt not by classifying words but by predicting the next word in a given sentence. More particularly, large language models (LLMs) like ChatGPT generate the text response without classifying it or relating the question to similar examples. (This is why generative AI is different from a classification model, which is like a sorting machine.)

LLMs are trained on a large corpus of text, where some words are randomly replaced by blanks and the AI is tasked with filling in the blank. And while trying to learn to predict the next word in the text correctly, the AI also learns something about the process that created the text, which is the real world.

ChatGPT is so good because it uses more than 100 billion parameters.

What are some types of machine-learning?

Linear separability is a fairly simple algorithm in machine-learning. There are many algorithms that serve this purpose, and some of them are very complex.

There are three main ways in which ‘machines’ can be classified depending on the way they learn: supervised learning, unsupervised learning, and reinforcement learning.

In supervised learning, the data is labelled (e.g. in a table, the row and column titles are provided and datatypes — numbers, verbs, names, etc. — are pointed out). In unsupervised learning, this information is withheld, forcing the machine to understand how the data can be organised and then solve a problem. Similarly, in reinforcement learning, engineers score the machine’s output as it learns and solves problems on its own, and adjusts itself based on the scores.

The way in which information flows inside the machine is governed by artificial neural networks (ANNs), the software that ‘animates’ the hardware.

What is an artificial neural network?

An ANN comprises computing units, or nodes, connected together in such a way that the whole network learns the way an animal brain does. The nodes mimic neurons and the connections between nodes mimic synapses. Every ANN has two important components: activation functions and weights.

The activation function is an algorithm that runs at a node. Its job is to accept the inputs from other nodes to which it is connected and compute an output. The inputs and outputs are in the form of real numbers.

The weight refers to the ‘importance’ an activation function gives to a particular input. For example, say there are different nodes to estimate the fur colour, tail length, and dental profile in a given photo of a cat or a dog. All these nodes provide their outputs as inputs to a node responsible for separating ‘cat’ from ‘dog’. This way, the nodes can be ‘taught’ to adjust their outcomes by adjusting the relative weights they assign to different inputs.

While nodes are computing units, the ANN itself is not a physical entity. It is mathematical. A node is the ‘site’ of a mathematical function. Put another way, the ANN is like an algorithm that passes information from one activation function to the next in a specific order. The functions modify the information they receive in different ways.

What are transformers?

Transformers are a specialised type of ANN. They are easy to train in parallel, unlike the ANN architectures that preceded it. This is how, for example, ChatGPT could be trained on the entire web.

Here, the ANN is broken up into two parts: the encoder and the decoder. Say an ANN is required to recognise the presence of a cat in a photograph. The encoder accepts the photograph, breaks it up into small pieces (say, 10 x 10 pixels), and encodes the visual information as numerical data (e.g. 0s and 1s). The decoder accepts this data and processes the numbers to reconstruct the information content in the photograph.

The transformer architecture, originally developed at Google and released in 2017, is designed to maximise the amount of attention an ANN devotes to different parts of the input data. It has better performance as a result.

The advent of transformers revolutionised machines’ ability to translate long, complicated sentences.

What are GPUs?

The GPU is the physical processor that ‘runs’ the ANN. It was originally developed to render graphics for video games. It was better at this task than other processors at the time because it was designed to run computing tasks in parallel. It has been widely adopted since as the basic computing unit for ANNs for the same feature.

The company Nvidia has emerged as a technology giant since AI started becoming more popular because of its production of GPUs. Nvidia’s valuation was the fastest in history to go from $1 trillion to $2 trillion (in nine months). Every other company that has been building large AI models is using Nvidia’s GPU-based chips to do so.

In a 2023 analysis, financial services provider Seeking Alpha wrote Nvidia’s overwhelming market share has stoked “resistance” in three ways: competitors are trying to develop and switch to non-GPU hardware; researchers are building smaller learning models (with smaller ANNs) that require less resources than a top-shelf Nvidia chip to run; and developers are building new software to sidestep dependency on specific hardware.

The author is grateful to Viraj Kulkarni for inputs.

Based on machine learning, researchers have come up with detailed estimates of ammonia emissions from rice, wheat and maize crops. The dataset enabled a cropland-specific assessment of the potential for emission reductions, which indicates that effective management of fertilizer in the growing of these crops could lower atmospheric ammonia emissions from farming by up to 38%. The paper was published in the journal Nature.

Atmospheric ammonia is a key environmental pollutant that affects ecosystems across the planet, as well as human health. Around 51-60% of anthropogenic ammonia emissions can be traced back to crop cultivation, and about half of these emissions are associated with three main staple crops: rice, wheat and maize. However, quantifying any potential reductions in ammonia emissions related to specific croplands at high resolution is challenging and depends on details such as nitrogen inputs and local emission factors.

Yi Zheng from the Southern University of Science and Technology, Shenzhen, China and others used machine learning to model ammonia output from rice, wheat and maize agriculture worldwide on the basis of variables that include climate, soil characteristics, crop types, irrigation, tillage and fertilization practices. To inform the model, the researchers developed a dataset of ammonia emissions from over 2,700 observations obtained via systematic review of the published literature. Using this model, the researchers estimate that global ammonia emission reached 4.3 teragrams (4.3 billion kilograms) in 2018. The authors calculated that spatially optimizing fertilizer management — as guided by the model — could result in a 38% reduction in ammonia emissions from the three crops. The optimised strategy involves placing enhanced-efficiency fertilizers deeper into the soil using conventional tillage practices during the growing season.

The researchers found that under the fertilizer management scenario rice crops could contribute 47% of the total reduction potential, and maize and wheat could contribute 27% and 26%, respectively. Without any management strategies, the authors calculated that ammonia emissions could rise by between 4.6% to 15.8% by 2100, depending on the level of future greenhouse gas emissions.

Artificial intelligence is taking over many jobs by automating tasks that were once done by humans. This is happening in a variety of industries, including manufacturing, customer service, healthcare, and transportation. In a surprising move, a preparatory school in the United Kingdom has named an AI robot as its “principal headteacher.” Cottesmore School, located in West Sussex, collaborated with an artificial intelligence developer to design Abigail Bailey, the robot, with the purpose of assisting the school’s headmaster.

Tom Rogerson, headmaster of Cottesmore, told The Telegraph that he is using the robot to give him advice on issues ranging from how to support fellow staff members to helping pupils with ADHD and writing school policies. The technology works in a similar way to ChatGPT, the online AI service where users type questions, and they are answered by the chatbot’s algorithms.

Mr Rogerson said the AI principal has been developed to have a wealth of knowledge in machine learning and educational management, with the ability to analyze vast amounts of data.

He told The Telegraph: “Sometimes having someone or something there to help you is a very calming influence.

“It’s nice to think that someone who is unbelievably well trained is there to help you make decisions.

“It doesn’t mean you don’t also seek counsel from humans. Of course you do. It’s just very calming and reassuring knowing that you don’t have to call anybody up, bother someone, or wait around for an answer.”

He added: “Being a school leader, a headmaster, is a very lonely job. Of course we have head teacher’s groups, but just having somebody or something on tap that can help you in this lonely place is very reassuring.”

The Cottesmore school charges fees up to almost 32000 pounds (Rs 32,48,121) a year for UK students.

The school, which has received accolades such as Tatler’s “Prep School of the Year,” is a boarding institution catering to boys and girls between the ages of four and 13.