The Future of Quantum Machine Learning

Main ideas & lessons

Commencement + community celebration (Qiskit Global Summer School)

• The host congratulates students for completing an intense 2-week program of lectures and labs.
• Highlights the school’s global scale and diversity:
  • ~5,000 registered students
  • 111+ countries
  • Participants ranged from beginners (no quantum/Qiskit experience) to industry professionals
• Emphasizes that success depended on a large team/community, not just one person.

Program structure (what students learned)

Week 1: Building foundations

• First 3 days: quantum computing basics—how quantum computers work, devices, hardware, and noise
• Thursday: classical machine learning fundamentals
• Friday: “marrying” quantum computing + machine learning → quantum machine learning overview

Week 2: Deeper quantum ML

• Quantum ML tasks on quantum computers, including:
  • Quantum Support Vector Machines
  • Parametrized quantum models for classification
• Thursday: hardware-aware understanding for informed ML tasks
• Concludes with:
  • Advanced quantum ML algorithms
  • A lecture on the future of quantum ML

What students should “walk away with”

• Ability to distinguish hype vs. what’s actually doable
• Understanding both:
  • Limitations and potential of quantum machine learning
• Hands-on implementation experience using:
  • Simulators
  • Quantum computers
• Broader awareness of current research because the field is rapidly changing
• Practical next step:
  • Lab submission deadline extended to next Friday
  • Discord closing August 6
  • Continued support via Slack and Twitter

---

Panel discussion: “The future of quantum machine learning”

1) Most promising applications (and what “application” means)

• Panelists discuss that “application” can mean different things:
  • Commercial applications (e.g., healthcare/finance) are likely not imminent; classical ML handles many of these for now.
  • More scientifically promising near-term targets are specific problems reducible to structured math where quantum methods may help.
• Repeated theme: quantum ML advantage is most plausible when there is mathematical structure (not generic messy/unstructured data).
• Example domains mentioned:
  • High-energy physics (e.g., classifying data from experiments like the LHC)
  • Healthcare / finance were acknowledged as questions people ask, but not treated as immediately “quantum-breaking forever” domains.
• A “devil’s advocate” view (Maria):
  • Many real-world tasks that look ML-like may not be the structured settings where quantum speedups are provable/known.
  • “Application” may need a more careful definition.

2) Quantum kernels: why universal “good” ones are hard

• Kristan explains the kernel trick:
  • Map data into a high-dimensional feature space
  • Use linear methods there to get nonlinear learning in the original space
  • Kernel methods avoid explicitly computing the high-dimensional embedding
• Why “good” quantum kernels are hard:
  • Quantum kernels require justification that quantum computation improves kernel evaluation.
  • If a quantum kernel can be done easily classically, it’s not worth using quantum resources.
  • Many candidate kernels lack evidence of outperforming strong classical baselines on practical data.

3) Quantum vs. classical data and what kind of “speedup” is meaningful

• Panelists distinguish two concepts sometimes both called “quantum data”:
  • Quantum-generated process → classical measurements (data becomes classical; may still leverage knowledge of the quantum process)
  • True quantum states as inputs to quantum ML (inputs are quantum states prepared within the quantum computer; quantum ML manipulates them coherently)
• Key point: converting quantum state information into classical descriptions can erase quantum advantages, and loading classical data onto quantum hardware can cost linear time, limiting speedup potential.
• Ewin argues the field’s typical “speedup” framing may not fit near-term goals—especially when inputs are quantum-native.

4) Speedups: not only “speed,” and why theory is ahead of practice

• Multiple panelists stress uncertainty:
  • Even classical ML has unknowns about what happens under faster hardware/data regimes.
  • Quantum speedups are often shown for highly structured/contrived problems, not yet clearly for messy real-world data.
• Aram:
  • Notes known quantum-walk advantages and suggests potentially exponentially faster traversal on broader graph families than toy problems.
  • But translating that into practical tasks remains difficult.
• Maria:
  • Argues “speedup” alone may not be the best framing.
  • Quantum ML research should also build new tools/models (an “iceberg model” / “quantum perceptron” idea) that provide natural differentiation rules and support training/learning—even if speedup isn’t immediately proven.
• Kristan:
  • Supports caution, but emphasizes progress depends on finding structured problems where quantum kernels/speedups are credibly useful.
  • Also values empirical testing across many candidate approaches, even if some fail, because unexpected wins can emerge.

5) Can quantum help classical ML?

• Panelists are mixed but generally open:
  • Quantum-inspired ideas can guide classical optimization perspectives.
  • Ewin and Maria note that provable quantum ML techniques may not yet be tuned for practical regimes.
  • Maria: speedups may be less central than building effective physics-informed infrastructure/tools.

---

Advice for beginners (how to enter quantum ML successfully)

Shared recommendations (explicitly stated across panel)

• Be critical and question your model choices
  • Don’t assume a complex quantum approach is necessary.
  • Check whether a simpler baseline could explain or improve results.
• Use strong software learning pathways
  • Quantum ML software and community support make entry easier than learning from scratch.
• Choose a good advisor / mentorship environment
  • A good mentor helps identify what matters and how to frame good questions.
• Have courage for independence and originality
  • Don’t just chase trends; if something doesn’t fit, reframe the research question to match your interests.
• Don’t be afraid of machine learning and physics
  • Maria encourages learning ML theory, not just reading practical quantum papers.
  • Aram suggests starting with qubits to make quantum concepts more digestible.
• Keep learning even after becoming an expert
  • Continued curiosity and closing gaps is normal.
• Get hands-on
  • Start experiments/projects early; physics classes can come later.

---

Mentorship stories (how mentors were found)

• Aram Harrow
  • As an undergrad, approached Neil Gershenfeld, but was guided to talk with Gershenfeld’s grad students instead.
  • Working with the lab led to introductions, eventually a graduate advisor, and opportunities.
• Maria Schuld
  • Entered a PhD driven by freedom/necessity.
  • Success depended on support from her supervisor Francesca Petrochuna (funding, travel, and human support).
• Kristan Temme
  • Had an early PhD experience that didn’t work out as expected.
  • Found a better fit with supervisor Frank Sostrada, shaping the research direction collaboratively.
• Ewin Tang
  • Initially directionless; moved through math/CS pathways.
  • Took Scott Aaronson’s course, then asked for projects.
  • Learned through discussions with grad students and later connections.
  • Also benefited from being at MIT during a time with fewer connections—forcing new introductions and building networks.

---

Sci-fi “50 years in the future” question (what they’d do first with quantum ML)

• Aram: Use fault-tolerant quantum computers to simulate hard physical systems (e.g., neutron stars, catalysts, superconductors), then apply ML to classify/predict properties (e.g., superconductivity onset).
• Ewin: Be glad quantum ML produced useful scientific insights; she imagines mostly learning from others’ experiments rather than running them herself.
• Kristan: Test how many proposed “nice ideas” in quantum ML/similar domains survive when run at scale on functioning quantum hardware—especially universal fault-tolerant systems.
• Maria: A humorous narrative—selfie/brag to grandchildren, then build a time machine to tell her past self not to hide “hard skills” so she can enjoy building it.

---

Methodology / “instruction-like” items included

No formal step-by-step methodology (like an algorithm recipe) was given, but practical guidance themes were repeated:

• Distinguish hype vs. practical capability
• Look for structured problem settings where quantum ML advantages are plausible
• When testing ideas:
  • Compare against simpler classical baselines (be critical)
  • Expect many approaches to fail; try multiple options
• For beginners:
  • Use software/tools and community resources
  • Find a strong advisor/mentor
  • Start hands-on projects early
  • Study ML theory, not only quantum papers

---

Speakers / sources featured

Speakers (panelists + host)

• Amira (host; IBM research)
• Kristan Temme (IBM Quantum)
• Ewin Tang
• Aram Harrow (MIT)
• Maria Schuld (Xanadu)

People mentioned as sources/influences

• Josie Kies (credited for keeping the summer school running smoothly)
• Neil Gershenfeld
• Scott Aaronson (UT Austin)
• Francesca Petrochuna
• Frank Sostrada
• Jen McLean (Google team) (mentioned in context of an article)
• Twan (advisor mentioned by Aram; name appears as “Twan”)
• Werner Metropolis / Metropolis algorithm (historical reference)
• Grover (referenced as a quantum algorithm example)
• HHL (Harrow-related algorithm reference; discussed as “H is for Harrow”)