Shuhua Jiang

72 posts

Shuhua Jiang

@JiangSH24

Founder @ICIM_UK | NCIM Fellow | CNHC Reg. | Longitudinal child health, real-world trajectories, AI as memory infrastructure

Katılım Aralık 2021

397 Takip Edilen42 Takipçiler

Sabitlenmiş Tweet

Shuhua Jiang@JiangSH24·31 Ara

Prediction 2026: Medical AI moves from "Chatbots" to "World Models". LLMs predict text, not Homeostasis. To solve Autism, we need System Dynamics, not probability. Introducing WCDHI: A biological digital twin architecture inspired by @ylecun’s JEPA #NewYear2026 #AI #DeepTech

English

282

Shuhua Jiang@JiangSH24·2 Mar

Meta-analysis is powerful. But without causal framing, it risks averaging apples and oranges. As Judea Pearl reminds us, statistical aggregation needs structure. We are building a structure-aware clinical evidence engine.

Judea Pearl@yudapearl

I have asked ChatGPT if "meta analysis " is still an active field in stat. The answer: Yes, when aggregating effect sizes under statistical assumptions. Then I asked: Can you talk about "effect sizes" under strictly statistical assumption. Ans.: No. Go figure. I once described "meta analysis" as an attempt to average apples and oranges to learn properties of bananas. But statisticians, no matter how grotesquely, will continue to practice whatever their textbooks celebrate. See ucla.in/2N7S0K9 @eliasbareinboim @aclong111 @PatientStormDoc @DrJBhattacharya

English

Shuhua Jiang retweetledi

Judea Pearl@yudapearl·2 Mar

English

2.2K

Shuhua Jiang@JiangSH24·13 Oca

Professor Pearl, thank you for the clarification — this is exactly the identifiability boundary we grapple with in clinical practice. When M1 and M2 are interventionally equivalent, even rich RCT data remains silent, and counterfactual queries may stay underdetermined without further structural assumptions. In our work, this is where we treat model structure — rather than just more data — as the primary source of information. We focus on explicit functional constraints, system dynamics, and mechanistic assumptions that can be falsified over time. Ultimately, when Level-2 saturates and Level-3 is not identifiable, the path forward is stronger causal modeling, not just stronger statistics. 🙏

English

Judea Pearl@yudapearl·11 Oca

@JiangSH24 One tiny correction, in the example cited we have P(death | do(Drug)) = 0.1 under both M1 and M2. RCT in itself cannot distinguish between the two models.

English

246

Judea Pearl@yudapearl·9 Oca

Interesting. What is the principle by which the weights of the hypothesized causal models are updated by each empirical datum?

Shuhua Jiang@JiangSH24

One missing piece in medical AI is how to act under causal uncertainty. Inspired by @yudapearl, we don’t collapse uncertainty into heuristics. We maintain multiple causal hypotheses and update their weights online through real-world feedback. Prediction asks what may happen. Causal bounds + stability ask when automation must stop.

English

3.3K

Shuhua Jiang@JiangSH24·11 Oca

Great question! We are actually closing the loop between the two. 🔄 1. Dynamic Modeling: We use Structural Causal Models (SCMs) to simulate counterfactual trajectories in our model state space (S1-S7). This allows us to "audit" the potential risk of an intervention before it's applied. 2. Empirical Validation: We don't just rely on theory. We leverage 5 years of Real-World Data (RWD) from our clinical frontline to perform Bayesian updating on the model weights. We are starting the validation this month. The goal is "Predictive Homeostasis": Our industrial background in Process Simulation helps us treat these clinical datasets not as static points, but as dynamic system responses to do(x) actions. Would love to discuss how we handle the "causal uncertainty" in these bio-trajectories soon! 🩺🧬

English

SelfMonitoringLoop@SelfMonitorLoop·10 Oca

I haven't done symbolic regression yet, nor do I work with bio data, consequently I'm very interested in your work! I've recently been sanity checking with trig/reflection identities to see what's reliable and slowly building up from there, but that approach has been falling flat. You've been giving me lots of ideas though! Regarding your interventional audit are you modeling those counterfactuals dynamically, or do you have actual interventional data to validate against?

English

Shuhua Jiang@JiangSH24·6 Oca

Predicting a pediatric emotional meltdown isn't a text-completion task. It's a non-linear stability problem. Traditional RNNs/Transformers fail here because they try to model the trajectory in low-dimensional, noisy observation space. To solve this, we’ve been experimenting with Koopman Operator Theory. The idea: Lift the non-linear dynamics of S1 (Metabolism) and S7 (Interoception) into an infinite-dimensional Functional Space. In this lifted space, the chaotic evolution of a child's internal state becomes linear—and thus, controllable. The biggest hurdle? Choosing the right Lifting Functions (Observables). We are currently testing Radial Basis Functions (RBFs) vs. Neural Operators to capture the 'Homeostatic Manifold. Has anyone here applied EDMD (Extended Dynamic Mode Decomposition) to highly non-stationary biological time-series? Would love to hear your thoughts on preventing spectral leakage. 🧵 #KoopmanOperator #DynamicalSystems #ControlTheory #NonlinearDynamics #DataDrivenScience

English

Shuhua Jiang@JiangSH24·11 Oca

Professor @yudapearl, your “half-Bayesian” position resonates deeply. Bayesian priors are invaluable for encoding knowledge and uncertainty. But without causal structure, probability alone has no grammar for intervention or counterfactuals. In our work, we use Bayesian reasoning to seed beliefs, and causal models to govern how those beliefs change under do-operations and give counterfactual queries well-defined semantics at the unit level. Seeing + Doing — not one without the other.

English

Judea Pearl@yudapearl·11 Oca

@PoliticsIntlLaw @ngdonghung @JiangSH24 @soboleffspaces My other half is wedded to commonsense logic -- a rare commodity that is slowly making its way into certain scientific circles. [Can't elaborate, for fear of being labeled "genocide enabler"].

English

129

Judea Pearl@yudapearl·10 Oca

I am still only "half Bayesian", for reasons explained here: ucla.in/2nZN7IH and made concrete here: x.com/yudapearl/stat… @ngdonghung @JiangSH24 @soboleffspaces

Dong Nguyen@DongNguyeb

@yudapearl By the way, you need to stop describe yourself as a half Bayes. That causes a lot of mis understand.

English

11.5K

Shuhua Jiang@JiangSH24·11 Oca

English

Shuhua Jiang@JiangSH24·11 Oca

Professor @yudapearl, thank you for this Ladder of Causation challenge. Although the Average Causal Effect is zero, the individual outcome is highly informative. Under M1, P(death | do(Drug)) = 0. Under M2, P(death | do(Drug)) = 0.1. Observing a death after do(Drug) therefore rules out M1 as the generating mechanism and identifies this patient as belonging to the harmed sub-population of M2. This is a Level-3 (counterfactual) update on the unit-level response function, not a population-level Bayesian averaging.

English

289

Judea Pearl@yudapearl·10 Oca

@JiangSH24 Consider two models, with equal (50%) priors: M1- a sugar tablet, M2 - a drug that cures 10% and kills 10%. RCT finds no causal effect, confirming both models. Now we see a patient taking the drug by choice and dies. How do we update our priors?

English

417

Shuhua Jiang@JiangSH24·10 Oca

You’re pointing at the exact failure mode we worry about too: stable but wrong. To avoid that, we don’t treat S1–S7 as a frozen dictionary. We validate the lift along three axes: 1. Structural anchoring (systems biology + control theory, not blind feature lifts) 2. Interventional audit (do-interventions to falsify causal links) 3. Residual monitoring (unexplained variance = leakage, not noise to be fit) The Koopman layer is allowed to be stable only after causal consistency holds. Curious if you’ve explored invariant discovery or symbolic regression under bio-noise.

English

SelfMonitoringLoop@SelfMonitorLoop·9 Oca

I’ve played with EDMD a bit. Stability only means something relative to the observables. If the lift or dictionary is misspecified you can get dynamics that are super stable but also just stably wrong. I’ve been using rolling spectral baselines plus delta deviation as a leakage detector, but I don’t trust it unless it agrees across multiple observables / lift choices. No idea if my observations hold up to biological noise levels though.

English

Shuhua Jiang@JiangSH24·8 Oca

In our framework, functional stability signals (e.g. loss of recovery speed / critical slowing down) act as a guardian, not an optimizer. If causal intervals widen or stability degrades, autonomous intervention halts. We don’t need more “confident” AI. We need AI that knows when it no longer has the right to act. #CausalAI #NeurosymbolicAI #ControlTheory #MedicalAI

English

256

Shuhua Jiang@JiangSH24·8 Oca

English

3.6K

Shuhua Jiang@JiangSH24·7 Oca

We deliberately removed learning from our MVP. In clinical systems, instability amplified by learning is often more dangerous than not learning at all. For now, we freeze the gains, audit every causal edge, and log every intervention. Intelligence can wait. Stability cannot.

English

Shuhua Jiang@JiangSH24·6 Oca

At this stage, our “digital twin” is not a high-fidelity physiological simulator. It is a low-resolution twin of homeostatic trajectories. Prediction comes later. Control and causality come first. We detect Critical Slowing Down (CSD) statistically via autocorrelation — not through opaque learning. This is Architectural Discipline. 🧬 #WorldModels #CausalAI #SystemDynamics #DigitalTwin #DeepTech

English

Shuhua Jiang@JiangSH24·6 Oca

World Models don’t start with intelligence. They start with stability. 🌍 In WCDHI v1.0, our MVP is deliberately white-box, causal, and auditable. We avoid the Black-Box Trap by design: • No deep nets • No latent-space guessing • No stochastic hallucinations We define a discrete state space (S1–S7), apply expert-prior causal DAGs, and use fixed-gain PID to regulate intervention dynamics, not just symptoms. 🧵

English

Shuhua Jiang@JiangSH24·5 Oca

One thing I find missing in the current JEPA discussion is a notion of stability, rather than prediction accuracy. In biological systems, even latent-space prediction becomes unreliable near bifurcation points. Clinically, what matters more is the loss of recovery speed — i.e. critical slowing down. Dimension-contrastive objectives (VICReg / SIGReg) are especially interesting here, not only for preventing representation collapse, but for preserving causal separability across physiological subsystems. Prediction tells us what might happen. Stability tells us when prediction itself is about to fail.

Yann LeCun@ylecun

I think you missed the main ideas. - The basic premise of JEPA is that training by reconstructio/prediction in input space is evil (or counterproductive). The details are almost always unpredictable. Hence prediction must take place in representation space, where unpredictable details are eliminated. - The main issue with JEPA is how to prevent collapse (in the absence of reconstruction loss). There are two classes of methods: (1) EMA: Using weights in target encoder that are an exponential moving average (EMA) of the weights in other encoder (I-JEPA, V-JEPA, DINO, BYOL). (2) Infomax: Using a regularizer that attempts to maximize the information content of the representation (e.g. over a batch). There are two sets of methods for that: (2a) sample-contrastive methods: that want to make each representation vector different from the others (Siamese nets, DrLIM, SimCLR, etc). They tend to not work well in high dimension, to require large batches, and hard negative mining (2b) dimension-contrastive methods: that want to make each variable independent from the others (Barlow Twins, VICReg, SIGReg/ LeJEPA, MMCR, MCR2....) Bottom line: A. SSL by reconstruction/prediction doesn't work for high-dim, continuous, noisy data B. EMA sucks: no loss function being minimized, requirement for weightmsharing.... C. Sample-contrastive informax doesn't scale to high dimension D. My money is on dimension-contrastive methods like SIGReg/LeJEPA

English

104

Shuhua Jiang@JiangSH24·3 Oca

Week 1 of 2026: The Paradigm Shift. 🚀 A foundational week for ICIM (UK) and the Causal AI movement in medicine: ✅ Architecture: Unveiled WCDHI v1.0 (Moving from LLMs to World Models). ✅ Validation: Confirmed theoretical alignment with the Causal AI community (including interaction with Prof. @yudapearl). ✅ Methodology: Defined the Bio-Adaptive PID control loop for homeostasis. ✅ Execution: Initiating ICIM (UK) registration process. We are not building a "Feature"; we are building an "Architecture". Rest up. Next week: We tackle the "Guardian" Safety Layer. 🛡️ #BuildInPublic #CausalAI #TechBio #SystemDynamics

English

Shuhua Jiang@JiangSH24·2 Oca

People ask: "How do you apply Industrial Simulation to Medicine?" 🏭🧬 It’s not just about World Models; it's about Control Theory. In my 17 years in Industry, we used PID Controllers to stabilise complex chemical reactions. In our model, we apply Bio-Adaptive PID to Child Homeostasis: 🔶 P (Proportional): Managing acute symptoms (Meltdown Control). 🟩 I (Integral): Clearing metabolic debt (Inflammation Reduction). 🟪 D (Derivative): Predicting the trend via Interoception (Trend Prediction). Biology is the ultimate engineering challenge. #DeepTech #ControlTheory #Autism #EngineeringMedicine

English

Shuhua Jiang@JiangSH24·1 Oca

Honored by the acknowledgment from Professor @yudapearl! Your work on Causality is the bedrock of our architecture. We are committed to proving that in medicine, 'Why' matters more than 'Likely'. We will continue to validate this Causal World Model with clinical data in 2026.

Shuhua Jiang@JiangSH24

Why World Models? Medicine is a Causal Loop, not a sequence of tokens Our model simulates the feedback between S1 (Metabolism) & S7 (Interoception) Integrating @yudapearl’s Causal Inference, we detect physiological "Tipping Points" before a meltdown Simulation > Prediction🧬

English

415

Shuhua Jiang@JiangSH24·31 Ara

Professor @yudapearl said: 'Start by solving local problems.' We took this advice to heart. Here is our Causal Architecture (WCDHI) for the specific local problem of Autism Metabolic Meltdowns. 👇 (Linking my thread below) x.com/JiangSH24/stat…

Judea Pearl@yudapearl

Just back from vacation, glad to see new revolutionary ideas for causation. I'll be studying it soon but, off hand, I am skeptical of a global solution that do not start by solving local problems. I've the quote from Democritus in "Causality" (2000), but I leveraged it to solve concrete local problems: testable implications, identification, generalizability, explanation, etc, one by one, principle by principle. Will try to find the principle behind this one.

English

Keşfet

@eliasbareinboim @aclong111 @PatientStormDoc @DrJBhattacharya @yudapearl @PoliticsIntlLaw @ngdonghung @soboleffspaces