a bunch of iterations from Ben's colab notebooks and ideas#15
Open
charliederr wants to merge 43 commits intotrueagi-io:mainfrom
Open
a bunch of iterations from Ben's colab notebooks and ideas#15charliederr wants to merge 43 commits intotrueagi-io:mainfrom
charliederr wants to merge 43 commits intotrueagi-io:mainfrom
Conversation
Added GNU General Public License version 2 to the LICENSE file, clarifying the terms under which the code is distributed.
…eight level for learning (standard much of the time, SB based where non-Gaussianity seems extreme)
The full training config had demotion_threshold=0.3 and promotion_threshold=0.7, which were too high for larger shell configurations (32 neurons). Sinkhorn transport with larger shells produces more diffuse distributions with lower off-diagonal transport mass (typically 0.10-0.16). Changes: - Lower demotion_threshold from 0.3 to 0.10 - Lower promotion_threshold from 0.7 to 0.25 - Scale shell distance cost by 0.15 for better eps compatibility - Track column usage history for shell dynamics - Compute demotions before registering new state (compare against history) Results: Shell demotions now 20 (was 0) while maintaining 97.76% accuracy. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Acknowledge inspiration from Benjamin Goertzel's PyTorch codebase.
…eight level for learning (standard much of the time, SB based where non-Gaussianity seems extreme)
The full training config had demotion_threshold=0.3 and promotion_threshold=0.7, which were too high for larger shell configurations (32 neurons). Sinkhorn transport with larger shells produces more diffuse distributions with lower off-diagonal transport mass (typically 0.10-0.16). Changes: - Lower demotion_threshold from 0.3 to 0.10 - Lower promotion_threshold from 0.7 to 0.25 - Scale shell distance cost by 0.15 for better eps compatibility - Track column usage history for shell dynamics - Compute demotions before registering new state (compare against history) Results: Shell demotions now 20 (was 0) while maintaining 97.76% accuracy. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…_coding' into experimental/causal_coding
A standalone benchmark package for evaluating continual learning algorithms. Works with any deep learning framework (PyTorch, JAX, TensorFlow, NumPy). Features: - Simple model protocol: predict() and train_on_task() - Built-in datasets: Split-MNIST, Permuted-MNIST, Split-CIFAR10/100 - Comprehensive metrics: accuracy matrix, forgetting, BWT, FWT - Statistical tests: paired t-test, Wilcoxon, bootstrap CI, Cohen's d - Reference baselines: Naive fine-tuning, EWC (pure NumPy) - Visualization: accuracy heatmaps, forgetting analysis, comparisons - PROTOCOL.md specification document Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update pyproject.toml dependencies: matplotlib -> plotly>=5.0.0 - Convert fabricpc/continual/utils.py plotting functions to plotly - Convert cl-benchmark visualization/plots.py to plotly - Update example comment to reflect plotly usage Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Consolidate duplicated Sinkhorn algorithm implementations from transweave.py and weight_causal.py into a new optimal_transport.py module. This reduces code duplication (~200 lines removed) and provides a single source of truth for optimal transport utilities. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace NumPy with JAX for performance-critical computations: - optimal_transport.py: Use jax.numpy and lax.fori_loop for Sinkhorn - weight_causal.py: JAX-accelerate kurtosis, multimodal gap, and correction - transweave.py: Add lax import for consistency Maintains NumPy fallback when JAX is unavailable. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Creates CausalLinear, TransWeaveLinear, and CausalTransWeaveLinear nodes that embed continual learning directly into FabricPC's forward_learning() method. Uses singleton registries (CausalGradientRegistry, TransWeaveRegistry) to maintain gradient history while keeping node methods pure/stateless. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
… math formatting.
…d can read Ben's PDF too
…erformant maybe?)
…n will also work?
…ameters for promotion and demotion instead of just setting them statically
to GPL licensing made previously are now officially null and void. An Addendum to the license will be inserted into main here.
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
2 participants
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.
There's a lot here, so happy to refactor whatever is needed and/or do this all in several smaller stages if needed.