Training deep learning models typically requires large-scale data, yet in the medical domain such data are often difficult to obtain due to privacy constraints, the rarity of certain diseases, and the high cost of acquisition. In this talk, I present one approach to this challenge: pretraining with synthetic data generated from domain knowledge. As concrete examples, I introduce the synthesis of electrocardiograms (ECG) and phonocardiograms (PCG). For ECG, each waveform component (P, Q, R, S, and T) is modeled with Gaussian functions; for PCG, synthetic signals are generated by combining S1 and S2 heart sounds with modulated noise. I show that pretraining a model on such synthetic data and then fine-tuning on a small amount of real data substantially improves classification performance compared to training on real data alone, and that this improvement becomes more pronounced as the size of the real dataset decreases. I will also touch on extensions such as self-supervised learning with synthetic data and a comparison between knowledge-driven simulators and learned generative models, and discuss the broader potential of domain knowledge as a data source for medical applications where real data are limited.

This work was featured in a RIKEN press release.
For details, please see the related link.

For a quantum theory of gravity to have a well-defined Hilbert space, the inner product between different states of open and closed universes must be positive semi-definite. Positivity however is not manifest in the low-energy effective theory and in fact imposes nontrivial constraints on the theory. Working in the Gravitational Path Integral (GPI) approach, we derive the general set of positivity constraints on the closed and open universe Hilbert spaces.

In the case of AdS gravity, open universe positivity in principle follows from CFT unitarity, however the holographic description of closed universes remains unclear. Strikingly, we exhibit positivity of closed universes across many theories and prove that open positivity implies closed positivity, showing that the CFT 'knows' about the closed universe hilbert space.

We then analyze positivity constraints on gravitational theories coupled to axions. We present a method to compute off-shell axion wormholes in AdS and flat space which we use to show that positivity is violated if the axion shift symmetry is exact.
In low-energy EFTs where these wormholes are perturbatively stable, to restore positivity the wormhole must have a non-perturbative instability due to instantons that breaks the shift symmetry.
Positivity then leads to a proof of a sharp version of the Axion Weak Gravity Conjecture A-WGC, including precise numerical constants.
For the QCD axion this provides a bound on the axion decay constant which has phenomenological and experimental consequences for axion searches. In string theory, positivity gives a bound on the coupling between the axion and the dilaton in the low energy effective action.

Stochastic differential games with a large number of players are notoriously challenging, both theoretically and numerically, typically when it comes to computing Nash equilibria. Yet, when many players interact somehow symmetrically by responding only to the average behavior of the others, the game can surprisingly become more tractable by taking the limit of an infinite number of players. This is in direct analogy with the so-called « mean-field theory » which simplifies the analysis of large systems of interacting particles in statistical physics. Introduced independently about two decades ago by Lasry and Lions (mathematics) and Caines, Huang and Malahamé (engineering), the theory of Mean-Field Games has since been greatly developed with various applications in engineering, economical, social and biological sciences. The goal of this short lecture is to introduce the key concepts, particularly the deep connections between game theory, Partial Differential Equations and stochastic analysis, and to showcase a few striking recent applications.

Stochastic thermodynamics, initiated three decades ago, aims at quantifying the fluctuations of physical observables in relation with thermodynamic quantities (such as heat or entropy production). A typical result is the Thermodynamic Uncertainty Relation, which states that a high entropy production is required to obtain a favorable noise-to-signal ratio for some observables (such as displacement of a molecular motor) in stationary out-of-equilibrium systems. It is especially relevant for nanoscale systems, where the fluctuations cannot be neglected. This includes biological systems (e.g. biological motors such as kinesin along a microtubule), electronic systems (transistor-based memories), chemical reactions, etc.

The talk will be both a tutorial on some basic results or applications, and a presentation of some recent results and perspectives.

Singularities are locations where something is exceptional. In particular, singularities of differentiable maps are mathematical concepts corresponding to stationary points of functions and apparent contours of surfaces under projection onto the retina. These are unavoidable in general, but important to study the shape of spaces and behavior of maps. The theory for them was initiated by R. Thom in 1950's, and have been deeply studied by many researchers.

This two-day workshop will bring together leading experts from academia, industry, and national laboratories to explore the rapidly evolving frontiers of quantum computing applications and their integration with high-performance computing (HPC) platforms.
Hosted by RIKEN Quantum, the event will provide a forum for discussing recent advances, practical challenges, and future directions toward achieving utility-scale quantum computations and robust quantum–HPC hybrid workflows.

The workshop is primarily an in-person event, but a special session on quantum computing in chemistry and life sciences will also be accessible via Zoom.

Inequality in resources is widely thought to undermine cooperation in social dilemmas. Yet cooperation among unequals is ubiquitous: between senior and junior colleagues, firms of different sizes, nations with asymmetric stakes. Here, we offer a resolution to this puzzle and derive a novel prediction: if the returns from cooperation are shared in accordance with the individuals' strategic incentives, inequality enables and strengthens cooperation. We develop a strategic framework to systematically explore cooperation when the returns of a joint project can be shared unevenly. We characterise the optimal sharing rule, which we call resilient sharing, that can sustain cooperation in repeated interactions when no other rule can. Resilient sharing equalises incentives to defect across players, but is neither egalitarian nor proportional. Surprisingly, it typically rewards weaker partners beyond their relative contributions. We show that cooperation can be sustained through direct reciprocity in any environment whenever individual contributions are sufficiently unequal. Evolutionary simulations and a behavioural experiment confirm the central prediction: under resilient sharing, cooperation succeeds among unequal partners where it fails among equals. This suggests that cooperation is more likely to evolve and thrive when individuals can vary contributions and divide returns flexibly, pointing to the role of institutions and norms in harnessing inequality to stabilize cooperation.

We analyse how inequality in endowments and social structure jointly affect individuals' ability to cooperate. Individuals repeatedly invest in a local public good ("cooperation'') in an environment that is described by a distribution of endowments and a network of beneficiaries. We measure the cooperativeness of an environment by the minimum discount factor needed to sustain (any) cooperation in equilibrium. We characterise the endowment distribution that maximises cooperativeness for any given network and the corresponding minimum discount factor. The latter is shown to be inversely proportional to the maximal index of the graph describing the network. The corresponding dominant eigenvalue of the adjacency matrix characterises the most cooperative income distribution. Moreover, we show that if an environment maximises cooperativeness (over all income distributions and networks of a certain size), then the network is described by a nested split graph. We further show that this is the same class of graphs that maximise welfare for any given discount factor, and yet, the most cooperative graph need not be equal to the most efficient.

Quantum link invariants relate topology in 3 dimensions to mathematical physics and representation theory. They admit liftings to 4-dimensional structures, known as link homology.

We will explain how the skein relations for quantum invariants turn into homological structures at this higher level and how semisimple representation theory turns into non-semisimple representations and homological algebra upon categorification.

Currently there are hundreds of models describing inflation, a period of accelerated expansion in our universe. Each model lead to different imprints in cosmological observables, and for the purpose of testing the idea of inflation itself, it is essential to understand which predictions are model independent. This lead to the idea of cosmological bootstrap, a set of constraints from physical principles and symmetries alone.
In this talk I will give an overview on the cosmological bootstrap program. I will first explain how locality, unitarity and symmetry can constrain the kinematics of cosmological correlators. I will then talk about some recent progress on constructing positivity bounds on cosmology, which places constraints on the interactions of fields in inflation.

Lecture I: Conventional approach: Repeatability, Heisenberg’s original uncertainty principle, and the SQL for gravitational-wave detection

The conventional approach to quantum measurement theory taken by von Neumann (1932), Dirac (1958), and Schrödinger (1935) assumes the "repeatability hypothesis" stating that if a physical quantity is measured twice in succession, then the same value is obtained each time, which is often quantitatively generalized to the "approximately repeatable hypothesis" stating that after a measurement of a physical quantity with error ε, the post-measurement deviation around the measured value is no larger than ε; this is equivalent to saying that the state after obtaining a measurement result with error ε becomes an ε-approximate eigenstate corresponding to that measurement result.

From the approximate repeatability hypothesis, one can derive "Heisenberg’s original formulation of the uncertainty principle," namely, that when position and momentum are approximately measured simultaneously, the product of their respective errors is at least ℏ/2 (Heisenberg 1927, Kennard 1927, Ozawa 2015), as well as the "standard quantum limit (SQL) for monitoring the free-mass position", which states that when the position of a free mass m is measured at a time interval τ, the result of the second measurement cannot be predicted with uncertainty smaller than (ℏτ/ m)^{1/2} (Caves 1985). The last result leads to a sensitivity limit for interferometric gravitational-wave detectors, and in the early 1980s it was therefore argued that gravitational waves of the expected strength could not be observed using interferometric detectors (Braginsky et al. 1980, Caves et al. 1980).

Artificial intelligence is having a transformative impact on health and scientific discovery. This presentation will trace the evolution from foundational breakthroughs to the sophisticated capabilities of today's large-scale AI models. It will explore how these advanced systems are creating new possibilities across the healthcare landscape, from accelerating therapeutic development to enhancing diagnostic processes and interpreting complex medical data. The session will also take a deeper look at the future possibilities for AI in health and explore the emerging role of agentic AI in scientific discovery. The core theme is the responsible development of AI to create tools that assist scientists, support healthcare professionals, and empower users.

Bio:
Dr Joseph Ledsam leads Google Health in Japan, where he works across AI research, digital health and health in Google products. He has led research in medical AI, genomics and drug discovery published in journals including Nature, Nature Medicine and Nature Methods. Before moving to Japan he worked as a medical doctor in the UK, and founded the Health Research and Genomics teams in Google DeepMind. He obtained his medical degree from The University of Leeds, UK, and was a research fellow at University College London during his clinical residency.

In this talk, I will explore quantum-improved black hole solutions within the framework of asymptotic safety. In this approach, the Newton coupling becomes scale-dependent, necessitating a meaningful identification between the energy scale and a corresponding physical (length) scale to derive observable consequences for black hole spacetimes. I will argue that the requirement of consistency with the first law of black hole thermodynamics provides a physically motivated criterion for this scale-setting, particularly near the event horizon. Applying this principle, we propose a specific identification scheme that leads to a regularized geometry capable of resolving the ring singularity of Kerr black holes.

We study the quantum aspects of the conformal gravity in four dimensions, specifically addressing a known discrepancy in beta functions between general quadratic curvature theories and conformal gravity, which corresponds to two scalar degrees of freedom. We demonstrate that this mismatch is resolved by carefully introducing gauge-fixing and ghost terms via the BRST symmetry, which effectively adds the two scalar modes. Drawing lessons from two-dimensional quantum gravity and Liouville theory, we proceed to integrate the four-dimensional trace anomaly to derive a consistent Liouville action, which is given by a free-field action for the conformal mode with a consistent conformal anomaly. We give the condition that the BRST transformation is anomaly free. Finally I would like to talk about some application of this theory.

Lecture II: Modern approach: Quantum instruments, POVMs, measuring processes, intersubjectivity, and value reproducibility

The modern approach to quantum measurement theory is based on the "realizability theorem" stating that a measurement is physically realizable if and only if its statistical properties are represented by a completely positive instrument, and this is also equivalent to saying that the measurement can be described by an interaction with a measuring apparatus (Ozawa 1984, 2004).

The conventional analysis of a measuring process determines the post-measurement object state by applying the "projection postulate" to the meter measurement in the post-measurement state that "entangles" the object and the apparatus, but the above result has been established without assuming the projection postulate altogether; rather we use only the classical Bayesian probability update rule (Ozawa 1984).

We introduce the "intersubjectivity theorem" that states that, when multiple observers simultaneously and statistically correctly measure the same physical quantity, they obtain the same measurement value and the "value reproducibility theorem" that states that a statistically correct measurement correctly reproduces the value of the physical quantity immediately before the measurement (Ozawa 2025).

The above three theorems essentially solves the so-called measurement problem, since we eliminate the collapse of the wave function and we establish the reality of the the pre-measurement value of the measured observable to be copied to the meter value and to be recorded by the observer.

Lecture III: Measurement error, disturbance, the universally valid reformulation of Heisenberg’s uncertainty principle, and a quantitative generalization of the Wigner–Araki–Yanase theorem

Definitions of measurement error and disturbance are introduced (Ozawa 2002, 2019) and it is shown that there exists a solvable model for a physically realizable measurement that serves as a counterexample both to Heisenberg’s uncertainty principle in the conventional formulation and to the SQL (Ozawa 1988, 1989, 2002). Thus, those limits are no more considered as universal limits. In fact, the above counter example to SQL was found in 1988 using the idea of contractive state measurements by Yuen (1983) and the LIGO was started in 1994 to succeed in the gravitational wave detection in 2015 as announced in 2016.

New formulations are then proved for the uncertainty principle concerning the errors in the approximate simultaneous measurement of two physical quantities, called the "joint error relation" (Ozawa 2003b, 2004), and for the uncertainty principle concerning the error and disturbance associated with the measurement of a single physical quantity, called the "error-disturbance relation" (Ozawa 2003a). From the error-disturbance relation, a quantitative relation for measurement error under an additive conservation law is proved (Ozawa 2002a, 2003b), generalizing the "Wigner–Araki–Yanase theorem" (Wigner 1952, Araki-Yanase 1960), which states that a physical quantity not commuting with a conserved quantity cannot be measured accurately by a measurement interaction satisfying an additive conservation law. The above relation also derives limits for realizing quantum computing and operations under conservation laws (Ozawa 2002b), the results later developed as the resource theory of asymmetry.

We introduce a novel characterization of phase transitions based on hypothesis testing. In our formulation, a phase transition is defined as the breakdown of statistical indistinguishability under vanishing parameter perturbations in the thermodynamic limit. This perspective provides a general, order-parameter-free framework that does not rely on model-specific insights or learning procedures. We show that conventional approaches, such as those based on the Binder parameter, can be reinterpreted as special cases within this framework. As a concrete realization, we employ a distribution-free two-sample run test and demonstrate that the critical point of the two-dimensional Ising model is accurately identified without prior knowledge of the order parameter.

Lecture IV: Instruments in classical mechanics, quantum field theory, and cognitive science

In algebraic quantum field theory, measurements describable by interactions between the field and the measuring apparatus are characterized by the class of completely positive instruments that satisfy the condition called the normal extension property (NEP) (Okamura-Ozawa 2016).

In classical mechanics, traditionally only non-invasive measurements—those with trivial interaction—were considered admissible, for the observability of the trajectory of motion. Here, however, the full class of measurements realizable by classical-mechanical interactions is characterized in terms of instruments with NEP for the basis of the study of invasive measurements of classical systems.

Cognitive processes are also represented by completely positive instruments, along with the long-standing paradigm provided by von Helmholtz, who described a sensation-perception process as a sort of measuring interaction and referred to it as an unconscious inference. This framework is used to show the compatibility of the question order effect and the response replicability effect (Ozawa-Khrennikov 2019), which failed to be explained in an earlier approach using only projective measurement models. It is shown that there exists an instrument model, realizing both the question order effect and the response replicability effect, that is also capable of almost faithfully reproducing public-opinion survey data such as the well-known Clinton-Gore survey by Gallup in 1997 (Ozawa-Khrennikov 2021).