Quantum + AI : a convergence story (2/4)

In our last article, we examined how quantum computers differ fundamentally from classical ones and the ways they might be complementary. But a critical question remains in the field : Can quantum computing shatter the barriers holding back standard machine learning (ML), from memory bottlenecks to computational limits ? To answer this question, let’s dive into what quantum technology truly brings to the table first, and then examine how it could transform the machine learning methods shaping industries today.

EP #2 : AI’s Practical Challenges & Potential for Quantum

A Scientific Perspective

Machine learning shines in big data processing, particularly in classification tasks — where the goal is to organize data into distinct categories. A common challenge, however, is that data isn’t always naturally structured in a way that makes classification straightforward. Often, it lacks a clear geometric representation, making it difficult to separate into meaningful groups. One effective approach is to map the data into a higher-dimensional space, where patterns become more apparent and classification boundaries easier to define.

In 2020, researchers proposed a novel embedding technique that leverages quantum computing to perform this transformation. Instead of relying on traditional methods, they used a quantum computer to map data into a specialized high-dimensional space known as a Hilbert space — a fundamental concept in quantum mechanics that describes all possible states of a quantum system. This process, known as quantum embedding, opens the door to entirely new ways of handling complex datasets.

What makes this so powerful ? Quantum embedding spaces grow exponentially due to entanglement, allowing computations that far exceed the capabilities of classical systems. For instance, a quantum computer with just 1000 qubits can manipulate a superposition of ²¹⁰⁰⁰ different states simultaneously — something classical computers would have to process one step at a time. This ability to explore vast solution spaces in parallel could revolutionize how machine learning models operate, making them exponentially more efficient for certain types of problems.

To illustrate quantum embedding concretely, consider a classical dataset that cannot be linearly separated in its original space. While classical kernel methods struggle with complex transformations, a quantum feature map can leverage quantum entanglement to implicitly map data into an exponentially large feature space. This allows even simple linear models to capture highly complex relationships in the original data. Early experiments with quantum support vector machines have shown promise for classification tasks with fewer training examples than their classical counterparts, though scaling these advantages to practical problems remains a significant research challenge.

Yet, while quantum embedding offers exciting theoretical promise for enhancing machine learning, the real question is how quantum computing aligns with the actual needs of businesses deploying AI today. What bottlenecks are companies facing ? Where do classical machine learning approaches fall short ? And most importantly, could quantum computing offer a genuine competitive advantage ? To explore these questions, we turned to an industry expert with firsthand experience in enterprise AI adoption across industries: Aurélien Coquard, VPE at Dataiku.

Beyond Gen AI Hype : How Industry Actually Uses AI

Artificial intelligence may dominate headlines with the latest breakthroughs in large language models and image generators, but in real-world industries, AI plays a far more diverse and practical role. While the public conversation fixates on generative models, companies are quietly harnessing AI in ways that rarely make the news — but create tangible value.

At Dataiku, a leading enterprise AI platform, this reality is clear in how their customers implement machine learning. “The public conversation about AI has become heavily centered around generative models, but that represents just one segment of a much broader ecosystem,” says Aurélien Coquard, VP of Engineering at Dataiku. “In industrial settings, companies are deploying various specialized models tailored to specific business problems, many of which have been creating value for years. What’s often overlooked is that most businesses don’t have access to the billions of data points needed for training large models. They’re working with limited, domain-specific datasets and need AI approaches that can deliver value without massive data requirements.”

Contrary to the hype, many of the most effective AI applications today rely on well-established machine learning techniques that predate the recent generative AI boom. Classical supervised learning models — such as decision trees, random forests, and gradient boosting machines — continue to dominate structured data analysis, particularly in industries like finance and manufacturing, where interpretability and reliability are critical. Time series forecasting remains indispensable for inventory management and capacity planning, while clustering algorithms, dimensionality reduction techniques, and association rule learning power applications from customer segmentation to fraud detection.

Beyond the models themselves, the backbone of industrial AI is built on feature engineering pipelines, model monitoring systems, and seamless integration with existing business processes. As Aurélien points out, “What’s particularly interesting is how often these ‘traditional’ approaches outperform newer, more complex models in production environments. The right model isn’t always the most sophisticated one — it’s the one that solves the business problem reliably while being maintainable.”

This contrast between AI’s media portrayal and its practical deployment raises a crucial question: If classical machine learning methods still drive most industrial AI applications, where does quantum computing fit in ? Could it complement or enhance these approaches, or does it remain a solution in search of a problem ? To explore this, we need to look beyond the lab and into the boardroom — where real-world adoption decisions are made.

Quantum-Enhanced ML : Finding More with Less

The industrial AI landscape presents intriguing opportunities for quantum computing — not as a replacement for classical machine learning, but as a powerful complement especially to find use-case specific optimal methods. Instead of overhauling existing AI infrastructure, quantum computing offers enhancements that could propose new alternative solutions while unlocking new levels of efficiency and insight in machine learning.

Extracting Signal Where Classical Models See Noise

The true promise of quantum computing in AI extends beyond sheer computational speed. Its real strength lies in its ability to navigate complex parameter spaces and uncover patterns that classical models often fail to detect. Traditional machine learning struggles when faced with high-dimensional, noisy data, often requiring vast datasets to extract meaningful correlations. Quantum computing, however, has the potential to process intricate probability distributions more efficiently, revealing subtle relationships that would otherwise remain hidden, in particular when data is sparse as in most industrial use-cases.

This capability could be particularly transformative in fields where critical signals are buried under layers of noise. In financial modeling, where market trends are often obscured by fluctuations and randomness, quantum methods could help identify patterns that classical approaches overlook. In medical diagnostics, where early disease biomarkers may be faint and difficult to detect, quantum algorithms could enhance predictive accuracy, leading to earlier and more reliable diagnoses. Similarly, in materials science, where the properties of new compounds emerge from complex quantum interactions, quantum computing could accelerate the discovery of novel materials with groundbreaking applications.

Quantum Kernel Methods: Bridging Quantum and Classical AI

One of the most promising near-term applications of quantum computing in machine learning is quantum kernel methods. These leverage quantum computers as specialized processors to calculate similarity measures (kernels) between data points — tasks that classical systems struggle with at scale. Crucially, quantum kernels integrate seamlessly into established machine learning frameworks, such as support vector machines, allowing businesses to explore quantum advantages without abandoning their existing AI workflows.

Quantum kernel methods establish a natural connection between quantum computing and established classical algorithms. By performing quantum feature mapping on quantum hardware while leveraging classical ML frameworks for the actual learning, these hybrid approaches offer a pragmatic path to quantum advantage. This methodology has been pioneered in Schuld and Killoran’s research on quantum machine learning in feature Hilbert spaces, demonstrating how quantum and classical techniques can complement each other effectively.

Empirical Validation : The ML Path Forward

The quantum advantage in machine learning will likely follow the path of classical ML itself — driven more by empirical results than theoretical guarantees. Neural networks flourished despite initial theoretical gaps, and quantum ML may follow a similar trajectory.

In machine learning, practical success often precedes theoretical understanding. This pattern has been well-documented in the development of deep learning, where empirical breakthroughs drove the field forward while theoretical explanations followed later, as explored in Bengio et al.’s foundational work on representation learning.

For quantum ML, this means :

• Developing rigorous benchmarks to compare quantum and classical methods fairly
• Identifying specific problem domains where quantum features match task requirements
• Developing hybrid approaches that combine classical and quantum components
• Prioritizing real-world validation over theoretical quantum speedups

The path forward isn’t about quantum computers replacing classical AI infrastructure, but rather finding the specific niches where quantum processing offers genuine advantages in accuracy, data efficiency, or insight generation. As quantum hardware continues to mature, these empirical demonstrations will likely emerge first in specialized domains before potentially expanding to broader AI applications.

While this episode explored how quantum computing might enhance traditional machine learning applications, our next installment will dive into perhaps the most visible domain of AI today : generative models and natural language processing. As large language models continue to transform how we interact with technology, quantum computing presents tantalizing possibilities for enhancing their capabilities — from more efficient training to potentially novel architectures that could overcome current limitations. Join us as we explore how quantum-AI convergence might reshape the future of generative AI.

Jean, Aurélien & Floriane

• Aurélien Coquard is an engineering leader with deep expertise in data science and artificial intelligence. He is currently Vice President of Engineering at Dataiku, a leading enterprise AI platform that enables organizations to build, deploy, and manage AI-driven solutions at scale. Prior to joining Dataiku, Aurélien held key technical leadership roles at Ivalua and Systran, where he played a pivotal role in advancing AI-powered analytics and language processing technologies. He graduated from École Polytechnique.
• Jean Senellart is a seasoned technology leader with over two decades of experience at the intersection of artificial intelligence, natural language processing, and machine translation. Currently serving as the Chief Product Officer at Quandela, a leading quantum computing company, Jean brings his extensive expertise in AI to the emerging field of quantum technologies. Prior to his role at Quandela, Jean held various leadership positions at Systran, including Chief Executive Officer, Global Chief Technology Officer, and Chief Scientist. His work at Systran was instrumental in advancing the field of machine translation and applying AI techniques to language processing.
• Floriane de Maupeou is an Investment Director at Serena Data Ventures, the deep tech arm of Serena Capital, where she focuses on data infrastructure, AI, and quantum technologies. She also serves as Head of Investor Relations at Le Lab Quantique, a French nonprofit dedicated to advancing the global quantum ecosystem.

About Serena

Serena is a leading European VC fund with €1B AUM, investing from Seed to Series A. Founded in 2008, it backs ambitious founders in AI, SaaS, Climate Tech, and Impact, offering Europe’s largest operational support team and a 550+ member startup community. Serena has invested in 100+ startups, including Dataiku, Malt, The Fork, and Electra.

—

The Serena team is excited to share with you this content which is provided for information purposes only. It’s meant solely for non-commercial, personal use and shouldn’t be considered as the basis for any decision or action of any nature. Entrepreneurs or any other user of this study should base their decision solely on their own case analysis.

Serena is not liable and expressly declines any liability for any damages whatsoever, resulting from the use of all information provided within this document. There is no guarantee that the information contained on this page is correct, complete or up-to-date. If you notice any irregularity or erroneous information, please let us know at communication@serena.vc.