CHANGELOG.md

ExplainableAI.jl

Version v0.8.0

This release removes the automatic reexport of heatmapping functionality. Users are now required to manually load VisionHeatmaps.jl and/or TextHeatmaps.jl.

This reduces the maintenance burden for new heatmapping features and the amount of dependencies for users who don't need heatmapping functionality.

• Removed reexport of heatmapping functionality by updating XAIBase dependency to v3.0.0 (#162).
• Added GradCAM analyzer (#155). Try it with VisionHeatmaps.jl's new heatmap_overlay feature.

Version v0.7.0

This release moves the core interface (Explanation, heatmap, analyze) into a separate package called XAIBase.jl. Developers can make use of the XAIBase.jl interface to quickly implement or prototype new methods without having to write boilerplate code.

As announced with version v0.6.2, this is first release without LRP, which has been moved to a separate package called RelevancePropagation.jl. This separation is enabled by the new common XAIBase.jl interface.

• Move core interface into XAIBase.jl package (#154).
  • Renamed Explanation field neuron_selection to output_selection
  • Added Explanation field heatmap for heatmapping presets
• Move LRP into RelevancePropagation.jl (#157)
• Remove ImageNet preprocessing utilities (#159)
• Partially move documentation into the Julia-XAI ecosystem documentation

Version v0.6.3

• Allow Gradient analyzers on non-Flux models (#150)
• Fix typo in BATCHDIM_MISSING error (#150)

Version v0.6.2

This is first release of ExplainableAI.jl as part of the Julia-XAI organization (#149) and the last minor release that includes LRP before it is moved to its own separate package.

• Add Concept Relevance Propagation analyzer CRP (#146, #148)
• Add option to process heatmaps batch-wise using keyword argument process_batch=true (#146, #148)
• Remove FlatRule on dense layers from EpsilonPlusFlat and EpsilonAlpha2Beta1Flat composite presets (#147)

Version v0.6.1

This release brings GPU support to all analyzers.

• Support LRP on GPUs (#142, #140)
• Support gradient analyzers on GPUs (#144)
• Make Tullio optional dependency using package extensions (#141)
• Document GPU support (#145)

Version v0.6.0

This release brings a large refactor of LRP analyzers, supporting nested "dataflow layers" from Flux.jl like Chain and Parallel layers. This enables LRP on more complex model architectures like ResNets.

Due to the fact that these new features require a breaking release, we've used the occasion to clean up the API. Since the number of changes is large, this changelog has been split between changes to LRP analyzers and more general changes to the package.

Changes to LRP analyzers

Breaking changes:

• Remove all unicode characters from user-facing API (#107)
  • EpsilonRule: argument epsilon replaces ϵ
  • GammaRule: argument gamma replaces γ
  • AlphaBetaRule: arguments alpha and beta replace α, β
• Rename LRP analyzer keyword argument is_flat=false to flatten=true (#119)
• Remove check_model, replaced by non-exported check_lrp_compat (#119)
• Replace layerwise_relevances field of Explanation return type by optional named tuple extras. Access layerwise relevances via extras.layerwise_relevances. (#126)
• Remove composite LastNTypeRule (#119)
• Rename composite primitives to avoid confusion with LRP rules (#130)
  • rename *Rule to *Map
  • rename *TypeRule to *TypeMap

Breaking changes to commonly extended internal functions:

• Internal lrp! rule calls require extra argument layer (#119)
• Pre-allocate modified layers, replacing modify_param! with modify_parameters (#102)

New features and enhancements:

• Support nested Flux Chains (#119)
• Support Parallel layers (#135, #138)
• Support BatchNorm layers (#129, #134)
• Add GeneralizedGammaRule (#109)
• Support nested indexing in composite primitive LayerMap (#131)
• Pre-allocate modified layers in LRP analyzer field modified_layers (#119)
• Set LRP output relevance to one (#128)
• lrp! rule calls require extra argument layer, avoiding copies of unmodified layers (#119)
• Performance fixes for LRP rules, reducing number of generated pullback functions (#106, #108)
• Simplify LRP analyzer (#112, #119)
• Simplify LRP model checks (#110, #119)
• Improve type stability of LRP rules

Documentation:

• Update documentation, adding pages on model preparation, composites, custom LRP rules, developer documentation and a separate API reference for LRP analyzers (#137, #105)

Package maintenance:

• Refactor LRP rule tests (#103, #127)
• Fix LRP benchmarks (#104, #127)

General changes

Breaking changes:

• Rename Explanation field attribution to val (#136)

Documentation:

• Update documentation, adding pages on heatmapping and input augmentations (#137, #105)

Package maintenance:

• Compatibility with Flux.jl v0.14 (#116)
• Drop dependency on LinearAlgebra.jl and PrettyTables.jl (#119)
• Add Aqua.jl tests (#125)

Version v0.5.7

• Fix WSquareRule dispatch on Dense layers
• Fix Vararg deprecation warnings from composites

Version v0.5.6

• Drop Flux v0.12 due to compatibility issues in preactivation (#99)

Version v0.5.5

• Ignore bias in WSquareRule
• Faster FlatRule on Dense layers (#96)
• Faster WSquareRule on Dense layers (#98)
• Update rule tests and references

Version v0.5.4

This release brings bugfixes and usability features:

• Add pretty printing of LRP analyzers, summarizing how layers and rules are matched up (#89)
• Add LRP support for ConvTranspose and CrossCor layers
• Add equations of LRP rules to docstrings

Bugfixes:

• Fix bug affecting AlphaBetaRule, ZPlusRule and ZBoxRule, where mutating the layer modified Zygote pullbacks (#92)
• Fix bug in FlatRule bias (#92)
• Fix input modification for FlatRule and WSquareRule (#93)

Version v0.5.3

Big feature release that adds LRP composites and presets:

• Add LRP Composite and composite primitives (#84)
• Add LRP composite presets (#87)
• Add LRP ZPlusRule (#88)
• Export union-types of Flux layers for easy definition of LRP composites
• Improvements to docstrings and documentation
• Add test/Project.toml with compat entries for test dependencies (#87)

Version v0.5.2

This release temporarily adds ImageNet pre-processing utilities. This enables users users to apply XAI methods on pretrained vision models from Metalhead.jl. Note that this functionality will be deprecated once matching functionality is in either Metalhead.jl or MLDatasets.jl.

• Add ImageNet preprocessing utility preprocess_imagenet (#80)
• Change default heatmap color scheme to seismic
• Updated README with the JuliaCon 2022 talk and examples on VGG16

Version v0.5.1

Small bugfix release addressing a bug in v0.5.0. Version of ExplainableAI.jl shown in the JuliaCon 2022 talk.

• Fix bug in FlatRule (#77)

Version v0.5.0

Breaking release that refactors the internals of LRP analyzers and adds several rules.

List of breaking changes:

• Introduce compatibility checks for LRP rule & layer combinations using check_compat(rule, layer) (#75)
• Applying GammaRule and ZBoxRule on a layer without weights and biases will now throw an error (#75)
• In-place updating modify_layer!(rule, layer) replaces modify_layer(rule, layer) (#73)
• In-place updating modify_param!(rule, param) replaces modify_params(rule, W, b) (#73)
• Removed named LRP constructors LRPZero, LRPEpsilon, LRPGamma (#75)

Added new LRP rules:

• Add PassRule (#76)
• Add AlphaBetaRule (#78)
• Add FlatRule (a6e2c59)
• Add WSquareRule (a6e2c59)

Bug fixes:

• Fix bug in ZBoxRule (#77)
• Fix broadcasting for Julia 1.6 (#74)
• Support MLUtils.flatten

Performance improvements:

• Replace LRP gradient computation with VJP using Zygote.pullback (#72)
• Faster GammaRule

Version v0.4.0

Changes:

• Update heatmapping normalizer, using ColorScheme's get. Breaking due to renaming normalize to ColorScheme's rangescale. (#57)
• Rename InputAugmentation to NoiseAugmentation. (#65)
• GammaRule and EpsilonRule now use default arguments instead of keyword arguments, removing the need for users to type unicode symbols. (#70)
• ZBoxRule now requires parameters low and high instead of computing them from the input. (#69)
• Add IntegratedGradients analyzer. (#65)
• Add InterpolationAugmentation wrapper. (#65)
• Allow any type of Sampleable in NoiseAugmentation. (#65)

Performance improvements:

• Remove use of mapreduce. (#58)
• Load LoopVectorization.jl in tests and benchmarks to speed up Tullio on CPU. (#66)
• Type stability fixes for GammaRule. (#70)