Feature/mu pc#17
Merged
matthewbehrend merged 17 commits intomainfrom Apr 17, 2026
Merged
Conversation
… config handling in existing nodes to remove mutable from argument default.
…eld node with learnable hopfield_strength parameter wrapped in softplus to ensure positive values.
…tion and tests; add CIFAR-100 data loader and demo for muPC; Remove redundant recurrency in StorkeyHopfield implementation. MNIST classifer performs equally with StorkeyHopfield replacing a hidden linear layer. Added recall test demos for StorkeyHopfield.
…ogic; add more demo examples to test mupc - still debugging via tests. Revise Storkey node design to remove pass through signal; add hopfield fewshot demo testing a network with a hopfield node under conditions of data scarcity and noise: results in stat-sig accuracy gain versus control MLP network. StorkeyHopfield node validated.
…g skip connections instead of depth heuristic. Extends muPC to arbitrary graphs, accounting for node in_degree. Changed default weight initialization of hopfield node to Xavier. Update docstrings and comments.
…pdate documentation and tests for clarity. muPC demo now working to > 16 layers. Fix hopfield node by restoring probe passthrough signal and applying blending of probe and projection according to hopfield strength. Hopfield demo now producing U shaped response to hopfield strength and stat-sig improvement over vanilla MLP in noisy few shot conditions.
…nd the variance of activation jacobians in the gradient flow. Add CIFAR-10 data loader and resnet18 demo; replace ReLU with Tanh activation in muPC mnist demo. Add comprehensive documentation in docs/user_guides/ that covers intro, tutorials, and API.
….helpers with jax_setup for improved environment configuration. Consolidated mupc ab experiment into resnet18 example. added JPC_fc_resent comparison of JPC to FabricPC implementations of muPC
…with improve variance scaling. Replace learning rate terminology with inference rate when referring to inference phase. Adjust mupc demo default values. Remove energy metrics from demo outputs for clarity.
…g on skip connections to avoid the vanishing signal with in_degree scaling. Improve variance scaling internal to transformer node. Refactor mupc_demo to use fully connection resnet and leverage new SkipConnection.
…niform scaling. Update mupc_demo to reflect changes in training epochs and inference rate. Integrate SkipConnection nodes into transformer_demo for improved architecture flexibility.
… muPC scaling logic - Introduced LinearResidual node combining linear transformation with identity skip connection. - Updated SlotSpec to include is_variance_scalable for per-slot scaling control. - Refactored muPC scaling algorithm to use edge-based scaling based on slot properties. - Enhanced graph construction to accommodate new scaling logic. - Updated mupc_demo to support new LinearResidual mode for reduced graph depth.
…he slot is a skip connection path in the network. muPC resnet depth counts nodes with skip connection slots. Updated mupc demo. Improved resnet18 demo. Refactored v1 transformer demo with new skip connection nodes and refactored transformer v1 node to internally implement mupc scaling.
…ion on muPC and new node types.
matthewbehrend
force-pushed
the
feature/muPC
branch
from
c798aef to
8d7678b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Introduces maximal-update parameterization (muPC) to FabricPC, including new node types (LinearResidual, SkipConnection, edge-based scaling, depth computation, and multi-GPU training support. Introduces associative memory StorkeyHopfield as a composable node in predictive coding networks.