Back in June 2022, Oak Ridge National Laboratory debuted Frontier—the world’s most powerful supercomputer. Frontier can perform a billion billion calculations per second. And yet there are computational problems that Frontier may never be able to solve in a reasonable amount of time.

Some of these problems are as simple as factoring a large number into primes. Others are among the most important facing Earth today, like quickly modeling complex molecules for drugs to treat emerging diseases, and developing more efficient materials for carbon capture or batteries.

However, in the next decade, we expect a new form of supercomputing to emerge unlike anything prior. Not only could it potentially tackle these problems, but we hope it’ll do so with a fraction of the cost, footprint, time, and energy. This new supercomputing paradigm will incorporate an entirely new computing architecture, one that mirrors the strange behavior of matter at the atomic level—quantum computing.

For decades, quantum computers have struggled to reach commercial viability. The quantum behaviors that power these computers are extremely sensitive to environmental noise, and difficult to scale to large enough machines to do useful calculations. But several key advances have been made in the last decade, with improvements in hardware as well as theoretical advances in how to handle noise. These advances have allowed quantum computers to finally reach a performance level where their classical counterparts are struggling to keep up, at least for some specific calculations.

For the first time, we here at IBM can see a path toward useful quantum computers, and we can begin imagining what the future of computing will look like. We don’t expect quantum computing to replace classical computing. Instead, quantum computers and classical computers will work together to run computations beyond what’s possible on either alone. Several supercomputer facilities around the world are already planning to incorporate quantum-computing hardware into their systems, including Germany’s Jupiter, Japan’s Fugaku, and Poland’s PSNC. While it has previously been called hybrid quantum-classical computing, and may go by other names, we call this vision quantum-centric supercomputing.

A Tale of Bits and Qubits

At the heart of our vision for a quantum-centric supercomputer is the quantum hardware, which we call a quantum processing unit (QPU). The power of the QPU to perform better than classical processing units in certain tasks comes from an operating principle that’s fundamentally different, one rooted in the physics of quantum mechanics.

In the standard or “classical” model of computation, we can reduce all information to strings of binary digits, bits for short, which can take on values of either 0 or 1. We can process that information using simple logic gates, like AND, OR, NOT, and NAND, which act on one or two bits at a time. The “state” of a classical computer is determined by the states of all its bits. So, if you have N bits, then the computer can be in just one of 2^N states.

But a quantum computer has access to a much richer repertoire of states during computation. A quantum computer also has bits. But instead of just 0 and 1, its quantum bits— qubits—via a quantum property known as superposition, represent 0, 1, or a linear combination of both. While a digital computer can be in just one of those 2^N states, a quantum computer can be in many logical states at once during the computation. And the superpositions the different qubits are in can be correlated with one another in a fundamental way, thanks to another quantum property known as entanglement. At the end of the computation, the qubit assumes just one state, chosen based on probabilities generated during the running of the quantum algorithm.

It’s not obvious how this computing paradigm can outperform the classical one. But in 1994, Peter Shor, a mathematician at MIT, discovered an algorithm that, using the quantum-computing paradigm, could divide large numbers into their prime factors exponentially faster than the best classical algorithm. Two years later, Lov Grover discovered a quantum algorithm that could find a particular entry in a database much faster than a classical one could.

Perhaps most importantly, since quantum computers follow the laws of quantum mechanics, they are the right tool for simulating the fundamentally quantum phenomena of our world, such as molecular interactions for drug discovery or materials design.

The Quantum-Centric Supercomputer’s Center

Before we can build a quantum-centric supercomputer, we have to make sure it’s capable of doing something useful. Building a capable enough QPU relies on constructing hardware that can re-create counterintuitive quantum behaviors.

Here at IBM, the basic building block of a quantum computation—the qubit—is made out of superconducting components. Each physical qubit consists of two superconducting plates, which act as a capacitor, wired to components called Josephson junctions, which act as a special lossless, nonlinear inductor.

The current flowing across Josephson junctions is quantized—fixed to discrete values. The Josephson junctions ensure that only two of those values (or their superpositions) are realistically accessible. The qubit is encoded in two current levels, one representing a 0, the other a 1. But, as mentioned, the qubit can also exist in a superposition of the 0 and 1 states.

Because superconductors need frigid temperatures to maintain superconductivity, the qubits and some of their control circuitry are held inside a specialty liquid-helium fridge called a dilution refrigerator.

We change the qubit states and couple qubits together with quantum instructions, commonly known as gates. These are a series of specially crafted microwave waveforms. A QPU includes all of the hardware responsible for accepting a set of quantum instructions—called a quantum circuit—and returning a single output represented by a binary string. The QPU includes the qubits plus components that amplify signals, the control electronics, and the classical computation required for tasks such as holding the instructions in memory, accumulating and separating signals from noise, and creating single binary outputs. We etch components like qubits, resonators for readouts, output filters, and quantum buses into a superconducting layer deposited on top of a silicon chip.

But it’s a challenge trying to control qubits at the supersensitive quantum level. External noise, noise from the electronics, and cross talk between control signals for different qubits all destroy the fragile quantum properties of the qubits. Controlling these noise sources has been key in reaching the point where we can envision useful quantum-centric supercomputers.

Getting the Quantum Stuff up to Snuff

No one has yet conclusively demonstrated quantum advantage—that is, a quantum computer that outperforms the best classical one on a real-world relevant task. Demonstrating true quantum advantage would herald a new era of computing, where previously intractable tasks would now be within reach.

Before we can approach this grandiose goal, we have to set our sights a bit lower, to a target we call quantum utility. Quantum utility is the ability of quantum hardware to outperform brute-force classical calculations of a quantum circuit. In other words, it’s the point where quantum hardware is better at doing quantum computations than a traditional computer is.

This may sound underwhelming, but it is a necessary stepping-stone on the way to quantum advantage. In recent years, the quantum community has finally reached this threshold. Demonstrating quantum utility of our QPU, which we did in 2023, has convinced us that our quantum hardware is advanced enough to merit being built into a quantum-centric supercomputer. Achieving this milestone has taken a combination of advances, including both hardware and algorithmic improvements.

Since 2019, we’ve been incorporating advances in semiconductor fabrication to introduce 3D integration to our chips. This gave us access to qubits from a controller chip placed below the qubit plane to reduce the wiring on the chip, a potential source of noise. We also introduced readout multiplexing, which allows us to access the information from several qubits with a single wire, drastically reducing the amount of hardware we have to put in the dilution refrigerator.

In 2023, we implemented a new way to perform quantum gates—the steps of a program that change the value of the qubits—on our hardware, using components called tunable couplers. Previously, we prevented cross talk by fabricating the qubits that respond to different frequencies so that they wouldn’t react to microwave pulses meant for other qubits. But this made it too difficult for the qubits to perform the essential task of talking to one another, and it also made the processors slow. With tunable couplers, we don’t need the frequency-specific fabrication. Instead, we introduced a sort of “on-off” switch, using magnetic fields to decide whether or not a qubit should talk to another qubit. The result: We virtually eliminated cross-talk errors between qubits, allowing us to run much faster, more reliable gates.

As our hardware improved, we also demonstrated that we could deal with some noise using an error mitigation algorithm. Error mitigation can be done in many ways. In our case, we run quantum programs, analyze how the noise in our system changes the program outputs, and then create a noise model. Then we can use classical computing and our noise model to recover what a noise-free result would look like. The surrounding hardware and software of our quantum computer therefore includes classical computing capable of performing error mitigation, suppression, and eventually, error correction.

Alongside ever-improving hardware advances, we teamed up with the University of California, Berkeley, to demonstrate in 2023 that a quantum computer running our 127-qubit quantum chip, Eagle, could run circuits beyond the ability of brute-force classical simulation—that is, methods where the classical computer exactly simulates the quantum computer in order to run the circuit, reaching quantum utility. And we did so for a real condensed-matter physics problem—namely, finding the value of a property called magnetization for a system of simplified atoms with a structure that looked like the layout of our processors’ qubits.

Left: A quantum processing unit is more than just a chip. It includes the interconnects, amplifiers, and signal filtering. It also requires the classical hardware, including the room-temperature classical computers needed to receive and apply instructions and return outputs. Right: At the heart of an IBM quantum computer is a multilayer semiconductor chip etched with superconducting circuits. These circuits comprise the qubits used to perform calculations. Chips are divided into a layer with the qubits, a layer with resonators for readout, and multiple layers of wiring for input and output.

Error Correction to the Rescue

We were able to demonstrate the ability of our quantum hardware outperforming brute-force classical simulation without leveraging the most powerful area of quantum-computing theory: quantum error correction.

Unlike error mitigation, which deals with noise after a computation, quantum error correction can remove noise as it arises during the process. And it works for a more general kind of noise; you don’t need to figure out a specific noise model first. Plus, while error mitigation is limited in its ability to scale as the complexity of quantum circuits grows, error correction will continue to work at large scales.

Error Correction

But quantum error correction comes at a huge cost: It requires more qubits, more connectivity, and more gates. For every qubit you want to compute with, you may need many more to enable error correction. Recent advances in improving hardware and finding better error-correcting codes have allowed us to envision an error-corrected supercomputer that can make those costs worthwhile.

Quantum error-correcting schemes are a bit more involved than error correction in traditional binary computers. To work at all, these quantum schemes require that the hardware error rate is below a certain threshold. Since quantum error correction’s inception, theorists have devised new codes with more relaxed thresholds, while quantum-computer engineers have developed better-performing systems. But there hasn’t yet been a quantum computer capable of using error correction to perform large-scale calculations.

Meanwhile, error-correction theory has continued to advance. One promising finding by Moscow State University physicists Pavel Panteleev and Gleb Kalachev inspired us to pursue a new kind of error-correcting code for our systems. Their 2021 paper demonstrated the theoretical existence of “good codes,” codes where the number of extra qubits required to perform error correction scales more favorably.

This led to an explosion of research into a family of codes called quantum low-density parity check codes, or qLDPC codes. Earlier this year, our team published a qLDPC code with an error threshold high enough that we could conceivably implement it on near-term quantum computers; the amount of required connectivity between qubits was only slightly beyond what our hardware already supplies. This code would need only a tenth the number of qubits as previous methods to achieve error correction at the same level.

These theoretical developments allow us to envision an error-corrected quantum computer at experimentally accessible scales, provided we can connect enough quantum processing power together, and leverage classical computing as much as possible.

Hybrid Classical-Quantum Computers for the Win

To take advantage of error correction, and to reach large enough scales to solve human-relevant problems with quantum computers, we need to build larger QPUs or connect multiple QPUs together. We also need to incorporate classical computing with the quantum system.

Quantum-centric supercomputers will include thousands of error-corrected qubits to unlock the full power of quantum computers. Here’s how we’ll get there.

2024

Heron

→ 156 qubits

→ 5K gates before errors set in

2025

Flamingo

→ Introduce l-couplers between chips

→ Connect 7 chips for 7 x 156 = 1,092 qubits

→ 5K gates before errors set in

2027

Flamingo

→ l-couplers between chips

→ 7 x 156 = 1,092 qubits

→ Improved hardware and error mitigation

→ 10K gates before errors set in

2029

Starling

→ 200 qubits

→ l-, m-, and c-couplers combined

→ Error correction

→ 100M gates

2030

BlueJay

→ 2,000 qubits

→ Error correction

→ 1B gates

Last year, we released a machine we call the IBM Quantum System Two, which we can use to start prototyping error mitigation and error correction in a scalable quantum computing system. System Two relies on larger, modular cryostats, allowing us to place multiple quantum processors into a single refrigerator with short-range interconnects, and then combine multiple fridges into a bigger system, kind of like adding more racks to a traditional supercomputer.

Along with the System Two release, we also detailed a 10-year plan for realizing our vision. Much of the early hardware work on that road map has to do with interconnects. We’re still developing the interconnects required to connect quantum chips into larger chips like Lego blocks, which we call m-couplers. We’re also developing interconnects to transfer quantum information between more distant chips, called l-couplers. We hope to prototype both m- and l-couplers by the end of this year. We’re also developing on-chip couplers that link qubits on the same chip that are more distant than their nearest neighbors—a requirement of our newly developed error-correction code. We plan to deliver this c-coupler by the end of 2026. In the meantime, we’ll be improving error mitigation so that by 2028, we can run a quantum program across seven parallel quantum chips, each chip capable of performing up to 15,000 accurate gates before the errors set in, on 156 qubits.

We’re also continuing to advance error correction. Our theorists are always looking for codes that require fewer extra qubits for more error-correcting power and allow for higher error thresholds. We must also determine the best way to run operations on information that’s encoded into the error-correcting code, and then decode that information in real time. We hope to demonstrate those by the end of 2028. That way, in 2029, we can debut our first quantum computer incorporating both error mitigation and error correction that can run up to 100 million gates until the errors take hold, on 200 qubits. Further advances in error correction will allow us to run a billion gates on 2,000 qubits by 2033.

Knitting Together a Quantum-Centric Supercomputer

The ability to mitigate and correct errors removes a major roadblock in the way of full-scale quantum computing. But we still don’t think it’ll be enough to tackle the largest, most valuable problems. For that reason, we’ve also introduced a new way of running algorithms, where multiple quantum circuits and distributed classical computing are woven together into a quantum-centric supercomputer.

Many envision the “quantum computer” as a single QPU, working on its own to run programs with billions of operations on millions of physical qubits. Instead, we envision computers incorporating multiple QPUs, running quantum circuits in parallel with distributed classical computers.

Combining the strengths of quantum and classical

Quantum-centric supercomputing leverages quantum and classical resources in parallelized workloads to run computations larger than what was possible before. A quantum-centric supercomputer is a system optimized to orchestrate work across the quantum computers and advanced classical compute clusters in the same data center.

Recent work has demonstrated techniques that let us run quantum circuits much more efficiently by incorporating classical computing with quantum processing. These techniques, called circuit knitting, break down a single quantum-computing problem into multiple quantum-computing problems and then run them in parallel on quantum processors. And then a combination of quantum and classical computers knit the circuit results together for the final answer.

Another technique uses the classical computer to run all but the core, intrinsically quantum part of the calculation. It is this last vision that we believe will realize quantum advantage first.

Therefore, a quantum computer doesn’t just include one quantum processor, its control electronics, and its dilution refrigerator—it also includes the classical processing required to perform error correction, and error mitigation.

We haven’t realized a fully integrated quantum-centric supercomputer yet. But we’re laying the groundwork with System Two, and Qiskit, our full-stack quantum-computing software for running large quantum workloads. We are building middleware capable of managing circuit knitting, and of provisioning the appropriate computing resources when and where they’re required. The next step is to mature our hardware and software infrastructure so that quantum and classical can extend one another to do things beyond the capabilities of either.

Today’s quantum computers are now scientific tools capable of running programs beyond the brute-force ability of classical simulation, at least when simulating certain quantum systems. But we must continue improving both our quantum and classical infrastructure so that, combined, it’s capable of speeding up solutions for problems relevant to humanity. With that in mind, we hope that the broader computing community will continue researching new algorithms incorporating circuit knitting, parallelized quantum circuits, and error mitigation in order to find use cases that can benefit from quantum in the near term.

And we look forward to a day when the Top 500 list of most powerful supercomputers will include machines that have quantum processors at their hearts.

Andrew Ng has serious street cred in artificial intelligence. He pioneered the use of graphics processing units (GPUs) to train deep learning models in the late 2000s with his students at Stanford University, cofounded Google Brain in 2011, and then served for three years as chief scientist for Baidu, where he helped build the Chinese tech giant’s AI group. So when he says he has identified the next big shift in artificial intelligence, people listen. And that’s what he told IEEE Spectrum in an exclusive Q&A.

Ng’s current efforts are focused on his company Landing AI, which built a platform called LandingLens to help manufacturers improve visual inspection with computer vision. He has also become something of an evangelist for what he calls the data-centric AI movement, which he says can yield “small data” solutions to big issues in AI, including model efficiency, accuracy, and bias.

Andrew Ng on...

• What’s next for really big models
• The career advice he didn’t listen to
• Defining the data-centric AI movement
• Synthetic data
• Why Landing AI asks its customers to do the work

The great advances in deep learning over the past decade or so have been powered by ever-bigger models crunching ever-bigger amounts of data. Some people argue that that’s an unsustainable trajectory. Do you agree that it can’t go on that way?

Andrew Ng: This is a big question. We’ve seen foundation models in NLP [natural language processing]. I’m excited about NLP models getting even bigger, and also about the potential of building foundation models in computer vision. I think there’s lots of signal to still be exploited in video: We have not been able to build foundation models yet for video because of compute bandwidth and the cost of processing video, as opposed to tokenized text. So I think that this engine of scaling up deep learning algorithms, which has been running for something like 15 years now, still has steam in it. Having said that, it only applies to certain problems, and there’s a set of other problems that need small data solutions.

When you say you want a foundation model for computer vision, what do you mean by that?

Ng: This is a term coined by Percy Liang and some of my friends at Stanford to refer to very large models, trained on very large data sets, that can be tuned for specific applications. For example, GPT-3 is an example of a foundation model [for NLP]. Foundation models offer a lot of promise as a new paradigm in developing machine learning applications, but also challenges in terms of making sure that they’re reasonably fair and free from bias, especially if many of us will be building on top of them.

What needs to happen for someone to build a foundation model for video?

Ng: I think there is a scalability problem. The compute power needed to process the large volume of images for video is significant, and I think that’s why foundation models have arisen first in NLP. Many researchers are working on this, and I think we’re seeing early signs of such models being developed in computer vision. But I’m confident that if a semiconductor maker gave us 10 times more processor power, we could easily find 10 times more video to build such models for vision.

Having said that, a lot of what’s happened over the past decade is that deep learning has happened in consumer-facing companies that have large user bases, sometimes billions of users, and therefore very large data sets. While that paradigm of machine learning has driven a lot of economic value in consumer software, I find that that recipe of scale doesn’t work for other industries.

Back to top

It’s funny to hear you say that, because your early work was at a consumer-facing company with millions of users.

Ng: Over a decade ago, when I proposed starting the Google Brain project to use Google’s compute infrastructure to build very large neural networks, it was a controversial step. One very senior person pulled me aside and warned me that starting Google Brain would be bad for my career. I think he felt that the action couldn’t just be in scaling up, and that I should instead focus on architecture innovation.

“In many industries where giant data sets simply don’t exist, I think the focus has to shift from big data to good data. Having 50 thoughtfully engineered examples can be sufficient to explain to the neural network what you want it to learn.”
—Andrew Ng, CEO & Founder, Landing AI

I remember when my students and I published the first NeurIPS workshop paper advocating using CUDA, a platform for processing on GPUs, for deep learning—a different senior person in AI sat me down and said, “CUDA is really complicated to program. As a programming paradigm, this seems like too much work.” I did manage to convince him; the other person I did not convince.

I expect they’re both convinced now.

Ng: I think so, yes.

Over the past year as I’ve been speaking to people about the data-centric AI movement, I’ve been getting flashbacks to when I was speaking to people about deep learning and scalability 10 or 15 years ago. In the past year, I’ve been getting the same mix of “there’s nothing new here” and “this seems like the wrong direction.”

Back to top

How do you define data-centric AI, and why do you consider it a movement?

Ng: Data-centric AI is the discipline of systematically engineering the data needed to successfully build an AI system. For an AI system, you have to implement some algorithm, say a neural network, in code and then train it on your data set. The dominant paradigm over the last decade was to download the data set while you focus on improving the code. Thanks to that paradigm, over the last decade deep learning networks have improved significantly, to the point where for a lot of applications the code—the neural network architecture—is basically a solved problem. So for many practical applications, it’s now more productive to hold the neural network architecture fixed, and instead find ways to improve the data.

When I started speaking about this, there were many practitioners who, completely appropriately, raised their hands and said, “Yes, we’ve been doing this for 20 years.” This is the time to take the things that some individuals have been doing intuitively and make it a systematic engineering discipline.

The data-centric AI movement is much bigger than one company or group of researchers. My collaborators and I organized a data-centric AI workshop at NeurIPS, and I was really delighted at the number of authors and presenters that showed up.

You often talk about companies or institutions that have only a small amount of data to work with. How can data-centric AI help them?

Ng: You hear a lot about vision systems built with millions of images—I once built a face recognition system using 350 million images. Architectures built for hundreds of millions of images don’t work with only 50 images. But it turns out, if you have 50 really good examples, you can build something valuable, like a defect-inspection system. In many industries where giant data sets simply don’t exist, I think the focus has to shift from big data to good data. Having 50 thoughtfully engineered examples can be sufficient to explain to the neural network what you want it to learn.

When you talk about training a model with just 50 images, does that really mean you’re taking an existing model that was trained on a very large data set and fine-tuning it? Or do you mean a brand new model that’s designed to learn only from that small data set?

Ng: Let me describe what Landing AI does. When doing visual inspection for manufacturers, we often use our own flavor of RetinaNet. It is a pretrained model. Having said that, the pretraining is a small piece of the puzzle. What’s a bigger piece of the puzzle is providing tools that enable the manufacturer to pick the right set of images [to use for fine-tuning] and label them in a consistent way. There’s a very practical problem we’ve seen spanning vision, NLP, and speech, where even human annotators don’t agree on the appropriate label. For big data applications, the common response has been: If the data is noisy, let’s just get a lot of data and the algorithm will average over it. But if you can develop tools that flag where the data’s inconsistent and give you a very targeted way to improve the consistency of the data, that turns out to be a more efficient way to get a high-performing system.

“Collecting more data often helps, but if you try to collect more data for everything, that can be a very expensive activity.”
—Andrew Ng

For example, if you have 10,000 images where 30 images are of one class, and those 30 images are labeled inconsistently, one of the things we do is build tools to draw your attention to the subset of data that’s inconsistent. So you can very quickly relabel those images to be more consistent, and this leads to improvement in performance.

Could this focus on high-quality data help with bias in data sets? If you’re able to curate the data more before training?

Ng: Very much so. Many researchers have pointed out that biased data is one factor among many leading to biased systems. There have been many thoughtful efforts to engineer the data. At the NeurIPS workshop, Olga Russakovsky gave a really nice talk on this. At the main NeurIPS conference, I also really enjoyed Mary Gray’s presentation, which touched on how data-centric AI is one piece of the solution, but not the entire solution. New tools like Datasheets for Datasets also seem like an important piece of the puzzle.

One of the powerful tools that data-centric AI gives us is the ability to engineer a subset of the data. Imagine training a machine-learning system and finding that its performance is okay for most of the data set, but its performance is biased for just a subset of the data. If you try to change the whole neural network architecture to improve the performance on just that subset, it’s quite difficult. But if you can engineer a subset of the data you can address the problem in a much more targeted way.

When you talk about engineering the data, what do you mean exactly?

Ng: In AI, data cleaning is important, but the way the data has been cleaned has often been in very manual ways. In computer vision, someone may visualize images through a Jupyter notebook and maybe spot the problem, and maybe fix it. But I’m excited about tools that allow you to have a very large data set, tools that draw your attention quickly and efficiently to the subset of data where, say, the labels are noisy. Or to quickly bring your attention to the one class among 100 classes where it would benefit you to collect more data. Collecting more data often helps, but if you try to collect more data for everything, that can be a very expensive activity.

For example, I once figured out that a speech-recognition system was performing poorly when there was car noise in the background. Knowing that allowed me to collect more data with car noise in the background, rather than trying to collect more data for everything, which would have been expensive and slow.

Back to top

What about using synthetic data, is that often a good solution?

Ng: I think synthetic data is an important tool in the tool chest of data-centric AI. At the NeurIPS workshop, Anima Anandkumar gave a great talk that touched on synthetic data. I think there are important uses of synthetic data that go beyond just being a preprocessing step for increasing the data set for a learning algorithm. I’d love to see more tools to let developers use synthetic data generation as part of the closed loop of iterative machine learning development.

Do you mean that synthetic data would allow you to try the model on more data sets?

Ng: Not really. Here’s an example. Let’s say you’re trying to detect defects in a smartphone casing. There are many different types of defects on smartphones. It could be a scratch, a dent, pit marks, discoloration of the material, other types of blemishes. If you train the model and then find through error analysis that it’s doing well overall but it’s performing poorly on pit marks, then synthetic data generation allows you to address the problem in a more targeted way. You could generate more data just for the pit-mark category.

“In the consumer software Internet, we could train a handful of machine-learning models to serve a billion users. In manufacturing, you might have 10,000 manufacturers building 10,000 custom AI models.”
—Andrew Ng

Synthetic data generation is a very powerful tool, but there are many simpler tools that I will often try first. Such as data augmentation, improving labeling consistency, or just asking a factory to collect more data.

Back to top

To make these issues more concrete, can you walk me through an example? When a company approaches Landing AI and says it has a problem with visual inspection, how do you onboard them and work toward deployment?

Ng: When a customer approaches us we usually have a conversation about their inspection problem and look at a few images to verify that the problem is feasible with computer vision. Assuming it is, we ask them to upload the data to the LandingLens platform. We often advise them on the methodology of data-centric AI and help them label the data.

One of the foci of Landing AI is to empower manufacturing companies to do the machine learning work themselves. A lot of our work is making sure the software is fast and easy to use. Through the iterative process of machine learning development, we advise customers on things like how to train models on the platform, when and how to improve the labeling of data so the performance of the model improves. Our training and software supports them all the way through deploying the trained model to an edge device in the factory.

How do you deal with changing needs? If products change or lighting conditions change in the factory, can the model keep up?

Ng: It varies by manufacturer. There is data drift in many contexts. But there are some manufacturers that have been running the same manufacturing line for 20 years now with few changes, so they don’t expect changes in the next five years. Those stable environments make things easier. For other manufacturers, we provide tools to flag when there’s a significant data-drift issue. I find it really important to empower manufacturing customers to correct data, retrain, and update the model. Because if something changes and it’s 3 a.m. in the United States, I want them to be able to adapt their learning algorithm right away to maintain operations.

In the consumer software Internet, we could train a handful of machine-learning models to serve a billion users. In manufacturing, you might have 10,000 manufacturers building 10,000 custom AI models. The challenge is, how do you do that without Landing AI having to hire 10,000 machine learning specialists?

So you’re saying that to make it scale, you have to empower customers to do a lot of the training and other work.

Ng: Yes, exactly! This is an industry-wide problem in AI, not just in manufacturing. Look at health care. Every hospital has its own slightly different format for electronic health records. How can every hospital train its own custom AI model? Expecting every hospital’s IT personnel to invent new neural-network architectures is unrealistic. The only way out of this dilemma is to build tools that empower the customers to build their own models by giving them tools to engineer the data and express their domain knowledge. That’s what Landing AI is executing in computer vision, and the field of AI needs other teams to execute this in other domains.

Is there anything else you think it’s important for people to understand about the work you’re doing or the data-centric AI movement?

Ng: In the last decade, the biggest shift in AI was a shift to deep learning. I think it’s quite possible that in this decade the biggest shift will be to data-centric AI. With the maturity of today’s neural network architectures, I think for a lot of the practical applications the bottleneck will be whether we can efficiently get the data we need to develop systems that work well. The data-centric AI movement has tremendous energy and momentum across the whole community. I hope more researchers and developers will jump in and work on it.

Back to top

This article appears in the April 2022 print issue as “Andrew Ng, AI Minimalist.”