SCHTs: A Semi-Structured Dynamic Sparse Training Framework for Hardware-Efficient Deep Learning
Combines dynamic sparse learning with N:M constraints to connect model training with hardware design.
ORIGINAL SCIENCE
Cannistraci–Hebb
Training
Intelligence learns to build its own structure. Connectivity shapes how information flows. CHT uses local network topology to guide connection regrowth as the network evolves during sparse training.
From connection strength
to connection structure.
Weights describe connection strength; topology describes which units are connected. CHT explores how these connections can be organized through a sparse training process that continually updates network structure.
Read the NeurIPS 2025 paperCannistraci–Hebb TrainingNeurIPS 2025 · FIG. 1
CHTsI II III
01
Sparse initialization
Begin with a sparse connectivity structure.
A guided reading of the original research figure · View source
ORIGINAL RESEARCH FIGURE
See structure evolve.
NeurIPS 2025 · Zhang et al. · Figure 1: the CHTs training process. Click to enlarge. Source
01Sparse initialization
02Connection removal
03Network percolation
04Topology-guided regrowth
05Resume training
One original paradigm.
Multiple paths to efficient AI.
- Dynamic sparse learningUpdate connections during training and study how sparse structure adapts to a task.
- Semi-structured trainingLearn connectivity under constraints such as N:M patterns supported by hardware.
- Spiking neural networksExplore the interaction between connectivity and event-driven computation.
CHT RESEARCH LIBRARY
From scientific foundations
to an expanding research frontier.
A collection of core CHT papers, public preprints, foundational studies and related sparse-learning research, checked against public sources. Publication status is stated for each entry.
Core CHT research
Read paper
Cannistraci-Hebb Channel-wise Dynamic Sparse Training of Convolutional Neural Networks with Contextual Modulation
Extends CHT to convolutional networks through channel-wise topology and contextual modulation.
Cannistraci-Hebb Training on Ultra-Sparse Spiking Neural Networks
Brings topology-guided dynamic sparse training to spiking neural networks.
Alignment-Enhanced Integration of Connectivity and Spectral Sparsity in Dynamic Sparse Training of LLM
Studies alignment between connectivity and spectral sparsity for parameter-efficient language-model training.
Cannistraci-Hebb Training with N:M Semi-Structured Sparsity for Pre-Training and Re-Training
Introduces CHTs24 and a dynamic sparse retraining pipeline for hardware-compatible patterns.
Brain network science modelling of sparse neural networks enables Transformers and LLMs to perform as fully connected
Introduces the CHTs soft rule for topology exploration and connectivity updates in Transformers and language models.
Epitopological learning and Cannistraci-Hebb network shape intelligence brain-inspired theory for ultra-sparse advantage in deep learning
Establishes ESML and the CHT framework for learning with ultra-sparse connectivity.
CHT preprints and research reports
Towards NVIDIA 1:4 Semi-Structured 75% Sparsity via Cannistraci–Hebb N:M Dynamic Sparse-to-Sparse Training
Explores N:M sparse-to-sparse training at higher sparsity with optimizer co-design.
Constructing Brain-Inspired Sparse Topologies for Energy-Efficient ANN-to-SNN Conversion via Cannistraci-Hebb Training
Studies how topologies learned through sparse training transfer into ANN-to-SNN conversion.
Cannistraci-Hebb Training of Convolutional Neural Networks
Explores CHT under convolutional weight sharing and receptive-field constraints; listed as an early research report.
Foundations and network science
Adaptive Cannistraci-Hebb Network Automata Modelling of Complex Networks for Path-based Link Prediction
Extends local-topology link-prediction rules that underpin structural learning.
Modelling Self-Organization in Complex Networks Via a Brain-Inspired Network Automata Theory Improves Link Reliability in Protein Interactomes
Studies local-topology-driven link growth and network self-organization.
Can local-community-paradigm and epitopological learning enhance our understanding of how local brain connectivity is able to process, learn and memorize chronic pain?
Investigates epitopological learning through the rewiring of brain networks.
Common neighbours and the local-community-paradigm for topological link prediction in bipartite networks
Extends the local-community paradigm to topological link prediction in bipartite networks.
From link-prediction in brain connectomes and protein interactomes to the local-community-paradigm in complex networks
Studies how local connectivity organization informs link prediction in complex networks.
Complementary sparse-learning research
A Dendritic-Inspired Network Science Generative Model for Topological Initialization of Connectivity in Sparse Artificial Neural Networks
A complementary study of topology initialization for sparse learning, distinct from CHT itself.
Sparse Spectral Training and Inference on Euclidean and Hyperbolic Neural Networks
Explores parameter-efficient training through spectral representations, complementary to connectivity sparsity.
Pivoting Factorization: A Compact Meta Low-Rank Representation of Sparsity for Efficient Inference in Large Language Models
Studies compact model representations for inference; its results are distinct from CHT performance.
Verified November 15, 2026 against public primary sources. Updated as the research evolves.