PICASO

Probabilistic Conceptual Spaces for
Uncertainty-Aware Knowledge Representation

Shoaib Jameel & Ubaid Azam
University of Southampton

Phase I Results & Phase II Vision  |  Grant KGaO-100018

Alan Turing Institute Funded

Agenda

🎯

The Problem

Why AI needs to know what it doesn't know

☁️

How PICASO Works

Probability clouds & geometry

📊

Phase I Results

SOTA performance & uncertainty

🔮

Phase II Vision

Temporal, multi-source, inductive

1

The Problem

Why defence intelligence needs uncertainty-aware AI

The Problem We're Addressing

Intelligence analysts deal with vast, messy, conflicting information every day.

📡 Volume

Thousands of reports from multiple sources daily

⚠️ Ambiguity

"An armoured vehicle near the border" — what kind? whose?

🔀 Conflict

Source A says "Country X", Source B says "border of Y"

Click to toggle: Pins (old) vs. Clouds (PICASO)

Current view: Fixed Pins — every entity is a single point, no uncertainty

Current AI systems treat every fact as 100% certain. PICASO replaces rigid pins with probability clouds — capturing both what we know and how sure we are.

What is a Knowledge Graph?

A Knowledge Graph stores facts as triples: (Subject, Relation, Object). Click the nodes!

part of Battle of Stalingrad World War II instance of NATO Military Alliance country Praetorian Guard Roman Empire part of Battle of Kursk ? ← PICASO predicts this

Our dataset: 41,425 entities and 143,532 relations across 355 relation types

The Theory Behind PICASO: Gärdenfors' Conceptual Spaces

In 2000, cognitive scientist Peter Gärdenfors proposed that concepts live in geometric spaces — not as symbols or logic rules, but as regions in multi-dimensional quality spaces.

🧠 Core Idea

A conceptual space is built from quality dimensions (colour, size, temperature…). Concepts are convex regions in this space.

📐 Geometry = Meaning

Similarity = distance. Categories = regions. "Between" has literal geometric meaning. A point between "red" and "yellow" is "orange".

🌍 Three Levels of Representation

Sub-conceptual (dimensions) → Conceptual (geometric regions) → Symbolic (language/logic). PICASO operates at the conceptual level.

Conceptual space: concepts as geometric regions with quality dimensions

Key insight: If concepts are inherently geometric, then the right AI framework should reason geometrically — exactly what PICASO does with entities and relations in knowledge graphs.

From Conceptual Spaces to Geometric Representations

Gärdenfors' theory gives us a principled bridge between human-like concepts and mathematical representations:

Convex Regions

Natural concepts form convex sets: any point between two members also belongs to that category.

If a, b ∈ C ⇒ λa + (1−λ)b ∈ C
📏

Similarity = Distance

"How alike are two entities?" becomes how close are their points? Distance captures graded similarity.

sim(a, b) ∝ e−d(a,b)
🔄

Relations = Transforms

Relations like "part of" or "located in" are geometric operations: rotations, translations, scalings.

R: regionhregiont

Traditional KG Embeddings

Arbitrary vector spaces with no cognitive grounding. Entities are meaningless points.

vs

Conceptual Space Embeddings

Geometry is the meaning. Distances, regions, and transformations have cognitive interpretations.

Why Bayesian Geometric Learning Is Ideal for Conceptual Spaces

Conceptual spaces tell us what to represent. Bayesian geometry tells us how to learn it under uncertainty:

🎲 Incomplete Knowledge

We never observe the full conceptual space — only sparse facts. We need to represent what we know and what we don't.

☁️ Gaussians = Soft Convex Regions

A Gaussian N(μ, Σ) has high density at the prototype, fading outward — exactly how Gärdenfors describes concepts.

🔁 Learning = Posterior Updating

New evidence sharpens the Gaussian (more certain) or shifts it (revised belief). Naturally handles evolving knowledge.

PICASO's Design Follows Directly

Entities = Gaussian regions in conceptual space
μ = prototype, σ² = boundary sharpness

Relations = Geometric transformations between regions
Rotation + scaling + translation with own uncertainty

Scoring = Geometric overlap of predicted & actual regions
KL divergence, Mahalanobis distance — conceptual space interpretations

Uncertainty = Intrinsic to the representation
Not bolted on afterwards — emerges from Gaussian geometry

PICASO is not just another embedding model — it is a cognitively grounded framework where Bayesian geometry meets conceptual space theory, giving us uncertainty-aware reasoning with a principled foundation.

2

How PICASO Works

Three key ideas, explained simply

Idea 1: "Fuzzy Clouds" Instead of Pins

Entities Uncertainty Relations Scores Weights

Traditional systems: fixed pin for every entity.
PICASO: a probability cloud — capturing both what we know and how sure we are.

Real-World Example

"Challenger 2 Tank" → Small, tight cloud (certain)

"Armoured Vehicle" → Large, diffuse cloud (could be anything)

q(e) = N(μe, σ²e)

μ = "best guess" location  |  σ² = "how unsure" we are.
Small σ² = confident. Large σ² = uncertain.

High

Idea 2: Speed Through Geometry

Bayesian methods are powerful but normally too slow. PICASO solves this with closed-form geometry:

Old Way: Sampling

🎯🎯🎯

Throw thousands of random darts and count hits. Slow

PICASO: Direct Geometry

📐

Use a formula to calculate the exact answer instantly. O(1) per triple

μpred = srRot(μh) + tr   ,   σ²pred = sr² ⊙ σ²h + σ²r

In plain English: To predict a missing fact, PICASO rotates, scales, and translates the head entity's cloud. The uncertainty from both entity and relation combine — all with a single formula, no simulations.

Idea 3: Multi-Faceted Geometric Scoring (MGS)

Not all relationships look the same. PICASO uses 5 geometric lenses simultaneously:

Hierarchical

"Frigate is a type of naval vessel"

containment

Symmetric

"UK is ally of France"

equality

Asymmetric

"General X commanded Unit Y"

direction

s(h, r, t) = w1·sgeo + w2·strans + w3·sKL + w4·sbil + w5·scpx

In plain English: PICASO combines 5 different "views" of how well a fact fits. The model learns the weights (w) to trust the right view for each relationship type. All computed in closed form.

sgeo: Variance-normalised distance strans: Uncertainty-weighted translation sKL: Distribution overlap sbil: Multiplicative interactions scpx: Complex-plane score
3

Phase I Results

State-of-the-art performance with intrinsic uncertainty

What We Proposed vs. What We Delivered

📑 Original Proposal

  • • Build defence KG from Wikidata
  • • Develop probabilistic embedding model
  • • Evaluate on defence dataset
  • • Uncertainty via Brier Score & RMSE
  • • One publication
  • • Code release

✅ What We Actually Delivered

  • ✓ KG + Wikipedia text corpus (160M+ chars)
  • ✓ Full engine + 5 geometric scorers
  • ✓ Eval on defence + FB15k-237 + WN18RR
  • ✓ Stratified calibration (no post-hoc needed)
  • Two publications (LREC + CIKM target)
  • ✓ Code + data + GUI + blog + workshop + invited talk
  • Triple classification + ablation + NBFNet comparison
  • Text-enhanced init with LLM (Llama 3.2)

Phase I exceeded all original objectives in a rapid 3-month timeline (Dec 2025 – Mar 2026), delivering a fully validated prototype with an analyst GUI.

Defence Dataset — Complete

0
Defence Entities
0
Relations
0
Characters of Text
63M+
Wikidata Entities Scanned

Covers: military units, battles, weapons systems, treaties, operations, intelligence agencies, personnel.

The automated pipeline is repeatable and can be adapted for future data sources and domains.

Understanding Our Metrics: MRR

Mean Reciprocal Rank (MRR) measures how high the correct answer appears in a ranked list.

Example Query

Question: (Battle of Stalingrad, part of, ??)

PICASO ranks 14,541 possible answers and produces:

RankAnswer
#1World War IICorrect!
#2Eastern Front
#3Operation Barbarossa
...14,538 more entities

Reciprocal Rank = 1/1 = 1.0 (perfect!)

If the model was less sure...

RankAnswer
#1Eastern FrontWrong
#2Operation BarbarossaWrong
#3World War IICorrect

Reciprocal Rank = 1/3 = 0.33 (ranked 3rd)

MRR = average of reciprocal ranks across all test queries.
Higher is better. A score of 0.417 means the correct answer is typically ranked around 2nd–3rd out of 14,541.

Understanding Our Metrics: Hits@K

Hits@K asks: "Is the correct answer in the top K predictions?"

Example: PICASO's Top 10 Predictions for (NATO, instance of, ??)

#1 Military Alliance ✓ #2 Organisation #3 Defense Pact #4 Treaty Org #5 ... #10 ...

Correct answer is at #1Hits@1 = Yes, Hits@3 = Yes, Hits@10 = Yes

Hits@1

33.9%

Correct answer is the #1 prediction
out of 14,541 entities

Hits@3

45.7%

Correct answer in top 3
out of 14,541 entities

Hits@10

55.9%

Correct answer in top 10
out of 14,541 entities

Think of it like a search engine: Hits@10 = 55.9% means more than half the time, the right answer appears on the "first page" of results — out of 14,541 possibilities.

Benchmark: PICASO vs. Established Methods

Validated against 12 established baselines on two international benchmarks:

ModelFB15k-237WN18RR
MRRH@1H@3H@10MRRH@1H@3H@10
TransE.294.465.226.501
RotatE.338.241.375.533.476.428.492.571
TuckER.358.266.394.544.470.443.482.526
GIE (prev. best).362.271.401.552.491.452.505.575
NBFNet (deep reasoning).415.321.454.599.551.497.573.666
PICASO (ours).417.339.457.559.488.445.510.568
+15.2%
MRR improvement
over GIE
+25.1%
Hits@1 improvement
Correct first guess
0.417
Matches NBFNet
at O(1) cost
Uncertainty
NBFNet cannot

Triple Classification: Is This Fact True?

Given (NATO, instance of, Military Alliance) — can PICASO tell if it's true or false?

89.0%
Accuracy (FB15k-237)
+12.6 pts over R2D2+
0.949
ROC-AUC (FB15k-237)
+9.2 pts over R2D2+
86.1%
Accuracy (WN18RR)
+5.7 pts over baseline
ModelAccuracyPR-AUCROC-AUC
TransE67.3%0.7270.736
DistMult73.9%0.7800.803
R2D2+76.4%0.8650.857
PICASO89.0%0.9510.949

Operational meaning: Nearly 90% accuracy in detecting logically inconsistent facts. Acts as an automated triage filter — instantly flagging potentially poisoned intelligence for human review.

Matching Deep Reasoning Models — At Fraction of Cost

NBFNet is a heavy deep reasoning model that traverses entire subgraphs. PICASO matches it using simple geometric operations:

PICASO (Ours)

0.417

MRR on FB15k-237

O(1) per triple
Instant lookup + uncertainty

NBFNet (Deep Reasoning)

0.415

MRR on FB15k-237

O(|E| · |R|) per query
Expensive subgraph traversal

ModelMRRHits@1Hits@3Hits@10Uncertainty?
NBFNet0.4150.3210.4540.599No
PICASO0.4170.3390.4570.559Yes

PICASO beats NBFNet on MRR and Hits@1 while being orders of magnitude faster and uniquely providing calibrated uncertainty for risk-aware decisions.

What Makes PICASO Work? (Ablation Study)

We removed each component one at a time to measure its contribution:

VariantMRRHits@1Hits@3Hits@10
Full PICASO.417.339.457.559
w/o MGS (5 scoring functions).355 (-14.9%).272 (-19.7%).386.519
w/o Rotation transforms.379 (-9.2%).288.428.541
w/o Semantic Types.425.364.449.539 (-3.6%)

Biggest impact: MGS

Removing the 5 geometric scoring functions causes -14.9% MRR drop. This is the core engine.

Rotation matters

Phase-based rotation captures directional relationships — -9.2% without it.

Gaussians = uncertainty

Replacing clouds with pins keeps accuracy but eliminates all uncertainty capabilities.

"Honest" Confidence: Knowing What It Doesn't Know

Critical for defence: when PICASO is confident, it's usually right. When uncertain, it warns you.

High Confidence

Query: (Battle of Stalingrad, part of, ??)

Prediction: World War II   ✅

Confidence: 92%

Low Confidence

Query: (Obscure Unit X, country, ??)

Prediction: Germany   ⚠️

Confidence: 34% — verify!

uncertainty = Σ σ²head + Σ σ²tail

In plain English: Uncertainty = how "spread out" both entity clouds are. If both are tight, we're confident. If either is diffuse, we flag it.
On WN18RR: confident predictions achieve 59% accuracy, uncertain ones drop to 15% — a 4x difference. No post-hoc calibration needed.

Defence Domain Case Study

Applied PICASO to our custom Defence-Wikidata knowledge graph:

68,826
Defence Entities
143,532
Triples
355
Relation Types

Selective Prediction: Act on What's Reliable

45.1%
High Confidence
(top 25%)
52.4%
Medium Confidence
(middle 50%)
26.3%
Low Confidence
(bottom 25%)

Decision-makers can act on confident predictions and defer uncertain ones for human review — critical in defence where false positives are unacceptable.

Text-Enhanced Entity Initialisation

20,808 entities from Wikipedia  |  20,607 generated by Llama 3.2 (3B, local)

The Prompt

You are a factual encyclopedia writer. Combine the given facts into a coherent 2-3 sentence summary. Use ONLY the facts provided.

Name: Treaty of Cherasco   Type: treaty
Description: peace treaty ending the War of the Mantuan Succession
- part of: War of the Mantuan Succession   - country: Duchy of Savoy

Output

"The Treaty of Cherasco was a peace treaty that ended the War of the Mantuan Succession. It involved the Duchy of Savoy..."

Text ⟶ Sentence-Transformer ⟶ μe init

Replaces random init with semantically meaningful Gaussian mean embeddings.

41,411
Entities with text
~1hr
LLM generation

Improved uncertainty calibrationalmost no hallucination. All processing runs locally.

Live Demonstrations

We built a fully operational GUI and packaged the model as a reusable library:

GUI Demo: Interactive query interface with uncertainty dashboard

Library Demo: PICASO as an easy-to-use Python package

A tangible interface brings PICASO from theory to a working applied prototype. The packaged library enables defence SMEs and researchers to integrate uncertainty-aware KG reasoning directly.

What Surprised Us in Our Results

💡 Geometry Beats Deep Reasoning

We did not expect a closed-form geometric model to match NBFNet on FB15k-237. MGS alone accounts for 14.9% MRR contribution.

💡 Calibration Came for Free

We expected to need post-hoc calibration (Platt scaling). Instead, PICASO's confidence is naturally calibrated. No post-processing needed.

💡 Triple Classification Dominance

The 12.6-point accuracy gap was unexpected. Distributional overlap provides a fundamentally richer signal than point-based scoring.

💡 LLM Text Init Improved Calibration

Llama 3.2 descriptions didn't just improve accuracy — they notably improved uncertainty calibration, with almost no hallucination.

Bottom line: Probabilistic geometry is not just theoretically elegant — it is a practically superior paradigm for knowledge representation in high-stakes settings.

Publications & Open Resources

PUBLISHED   LREC 2026

Lightweight Cross-Lingual Federated Prompt Tuning for Low-Resource Languages

Privacy-preserving models for defence data. 90% GPU memory reduction.

PUBLISHED   ACL Main Track 2026

Long Context Modeling with Ranked Memory-Augmented Retrieval

ERMAR framework — dynamic relevance scoring & pointwise re-ranking. SOTA on long-context benchmarks.

IN PREPARATION   CIKM 2026

PROCS: Uncertainty-Aware Knowledge Representation via Conceptual Spaces

Core PICASO engine — strengthened with peer review feedback for resubmission.

Upcoming Talks

Apr 2026: Cardiff University Invited Talk  |  May 2026: LREC 2026 Presentation

💻 Open-Source Code

Full codebase + docs on GitHub

📦 Dataset Release

41K entities, 143K triples, 160M chars

🖥️ Analyst GUI

Interactive prototype + demo videos

📦 Packaged Library

pip-installable for integration

🌐 Project Blog

ubaidazam.github.io/PICASO

🎤 Workshop

School research presentation

All outputs freely available for the community to use and extend.

4

Phase II Vision

From static uncertainty to living intelligence

Extension Proposal: Three Research Directions

Building on the PICASO foundation to address three critical limitations:

WP4: Temporal KG

Facts change over time. "Country X is led by Person Y" has a validity period.

Goal: Time-conditioned Gaussian distributions that evolve

📄

WP5: Multi-Source Fusion

Text-init showed value of text. Now fuse structured + textual evidence end-to-end.

Goal: Text-to-distribution encoders + Bayesian fusion

🔍

WP6: Inductive Reasoning

New entities appear constantly in intelligence. Can't retrain every time.

Goal: Infer about unseen entities from neighbourhood structure

Together, these advance PICASO from a static, transductive system to a temporally-aware, text-grounded, and inductive reasoning framework — addressing real operational needs in defence intelligence.

Why These Directions?

⏰ Temporal: Intelligence is time-sensitive

A military alliance in 1940 differs from 2024. Temporal distributions let us ask: "Was this true at time T?"

📄 Fusion: Evidence comes in many forms

Analysts read reports, cables, news — not triples. End-to-end fusion lets the model absorb textual evidence continuously.

🔍 Inductive: The world doesn't stand still

New entities appear constantly. Inductive reasoning enables instant inference without retraining.

Building on Strong Foundations

CapabilityDeliveredExtension Adds
Temporal reasoningSnapshot-based→ Time-aware distributions
Text integrationEntity initialisation→ End-to-end fusion
New entitiesTransductive→ Inductive (zero-shot)
Evidence fusionStructure + text-init→ Multi-source joint
Operational useWorking GUI→ Production-ready

Publication Targets

• AAAI / NeurIPS 2027 (full paper)

• EMNLP 2026 (text fusion)

• Open-source framework release

Project Impact Summary

🎯 Best-in-Class Performance

State-of-the-art on FB15k-237 (MRR 0.417). Matches NBFNet at fraction of cost. 89% triple classification.

📊 Genuine Uncertainty Calibration

Natively calibrated without post-hoc adjustment. 1.7x accuracy gap between confident and uncertain predictions.

🧬 Text-Enhanced Initialisation

LLM-generated descriptions improve uncertainty calibration. Almost no hallucination in generated text.

🖥️ Operational Prototype

Full GUI with file upload, manual query, and uncertainty dashboard. Ready for analyst evaluation.

0.417
MRR (SOTA)
89.0%
Triple Class. Acc
O(1)
Inference Cost
68K+
Defence Entities
Modelled
3
Publications

Thank You

Questions & Discussion

Shoaib Jameel — M.S.Jameel@southampton.ac.uk

Ubaid Azam — u.azam@soton.ac.uk

Grant KGaO-100018 University of Southampton
github.com/ubaidazam/PICASO · ubaidazam.github.io/PICASO