Our Team

We are a recently graduated student team with complementary strengths in artificial intelligence and theoretical computer science. Our collaboration brings together deep expertise in applied AI/ML research and rigorous algorithmic thinking. We are driven by curiosity, technical excellence, and a shared goal of solving complex, real-world problems through interdisciplinary innovation.

Project details

We propose an open-source framework that transforms raw observations {X₁,…,Xₙ}  - could be text, image, and audio streams- into a continuously evolving knowledge graph (KG) that supports causal, neuro-symbolic reasoning in the OpenCog Hyperon ecosystem. The system meets the RFP’s call for end-to-end tooling—from extraction through refinement to benchmarking—while remaining domain-agnostic and fully compatible with MeTTa and the MORK hypergraph backend.

An asynchronous ingestion layer normalises heterogenous inputs into a common event format ⟨id,time,payload,meta⟩. Modular adapters handle UTF-8 text, RGB images, and 16-kHz audio. Each adapter performs light pre-processing (tokenisation, patching, spectrogramming) and forwards the result to its modality-specific encoder. The adapters expose a uniform gRPC interface so new modalities (e.g., time-series sensors) can be snapped in without recompiling the pipeline.

Three self-supervised encoders—Transformer-based for text [1], masked-token ViT for images [2], and contrastive speech transformer for audio [3]—learn modality-specific embeddings without labels. The pre-text tasks (masked language modelling, masked patch prediction, future audio contrast) force each encoder to capture high-level semantics rather than surface patterns, delivering 768-D vectors zₖ with cosine-normalised magnitudes. A joint projection head maps all embeddings into one aligned latent space, enabling cross-modal comparison similar to CLIP-style losses [4]. This architecture is agnostic to the specific self-supervised recipe, satisfying the requirement to remain flexible to future representation advances.

A lightweight, incremental clustering module discretises incoming embeddings. New vectors are either merged into the nearest prototype if the cosine distance <τ (adaptive) or seeded as a fresh concept node otherwise. Each accepted assignment generates a MeTTa S-expression:

(concept :id 8743 :type Entity :label "violin")
(contextLink 8743 9021 :relation usedIn :confidence 0.87)

where 9021 may be a latent “classical_music” node inferred from co-occurrence statistics. Edges are hyperedges when n-ary relations are detected (e.g., (play Person Instrument Location)). Provenance, timestamps, and encoder uncertainty are stored as atom-level fields, enabling later integrity checks.

The symbolic layer (PLN + ECAN) consumes the graph to perform logical inference, similarity-guided retrieval, and planning. It can issue graph queries back to the neural side—for instance, “find an embedding most similar to the prototype for cat but closer to water_terrain” to hypothesise fishing_cat. Conversely, the neural layer consults the graph as a semantic prior, biasing its predictions toward entities already grounded with high-confidence links. This bidirectional flow realises the cognitive synergy advocated in neuro-symbolic literature [5], while maintaining clear audit trails for every symbol introduced.

A dedicated stream engine ingests atoms on commodity hardware. Every Δt seconds, a refinement job executes three passes: (i) duplicate detection and alias merging via incremental disjoint-set union; (ii) contradiction spotting using rule-based SAT templates [6]; (iii) redundancy pruning guided by a compression-coverage objective similar to PG-T [7]. Obsolete atoms decay by exponential ageing unless refreshed by new evidence. This keeps the KG compact, semantically rich, and internally consistent as mandated by the RFP’s quality goals.

Temporal edges carry lag metadata, enabling automatic extraction of candidate causal graphs. We implement an algorithm inspired by PCMCI+ [8] to propose causes links, which the symbolic layer verifies through constraint-based testing. Counterfactual queries are answered by running abduction on the causal subgraph, followed by forward simulation using learned conditional distributions. Benchmarks will include TETRAD synthetic sets and MetaQA multi-hop causal subsets, reporting AUROC and average intervention score to satisfy the “evaluate KG utility for reasoning” requirement.

The entire knowledge‐graph layer is built on MeTTa and backed by the high-performance MORK engine, with a simple API (REST/JSON-LD) for integration. It’s delivered as a modular, containerized package for straightforward deployment and maintenance. The code is organized into clear neural and symbolic components under an open-source license, thoroughly tested, and optimized to meet the RFP’s performance and scalability expectations.

We will report: extraction F₁ on Wikidata subsets, compression-coverage ratio, contradiction resolution accuracy, causal inference AUROC, and average query latency. These metrics map directly to the RFP’s emphasis on compactness, integrity, reasoning support, and performance. Public leaderboards and Jupyter notebooks will accompany each release for third-party replication.

[1] Devlin J. et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” NAACL-HLT 2019.
[2] He K. et al., “Masked Autoencoders Are Scalable Vision Learners,” CVPR 2022.
[3] Baevski A. et al., “Wav2Vec 2.0: A Framework for Self-Supervised Learning of Speech Representations,” NeurIPS 2020.
[4] Radford A. et al., “Learning Transferable Visual Models From Natural Language Supervision,” ICML 2021.
[5] Liu Z. et al., “Neuro-Symbolic Methods and Knowledge Graph Reasoning: A Survey,” AAAI 2023.
[6] Zhang J. et al., “Rule-Based Knowledge Graph Consistency Checking,” ISWC 2022.
[7] Bourhis P. et al., “Pruning Knowledge Graphs with Pattern-Graph Truncation,” WWW 2023.
[8] Runge J. et al., “Detecting Causal Associations in Large Nonlinear Time Series Datasets,” Science Advances 2022.

Background & Experience

Seyed Mohammad Seyed Javadi is a Master’s student in Computer Science at York University, specializing in theoretical computer science. He has published in top-tier conferences such as IJCAI and ICALP, and is a National Gold Medalist in the Iranian Computer Olympiad. His strengths include competitive programming, algorithm design, and formal theoretical analysis.

Amirhossein Mohammadi is a Master’s student in Artificial Intelligence at York University with over three years of focused research experience in AI and machine learning. His work spans model development, deep learning, and applied machine learning systems. Amir has hands-on experience building AI solutions and is passionate about advancing practical applications of intelligent systems.

Describe the particulars.

Crypto week in Toronto