CHANGELOG.md

Change Log

Unreleased

Changed - in summary

• Removed regions completely. HTM networks now consist only of layers.
• Layers are constructed in unembedded state, then embedded in a network.
• Layers receive signals - consisting of SDRs - from surrounding layers/senses.
• core namespace now largely abstract and separate from layer implementation.

Changed - in detail

• Layer topography value is that which is exposed as a signal - the cells.
• PFeedForward, PFeedBack, PFeedForwardMotor replaced with PSignalSource.
• region-network renamed to network and now has simpler args.
• regions-in-series removed.
• region-seq, region-keys => layer-seq, layer-keys.
• corresponding argument changes to these and related functions.
• in demos, n-region-model => build.
• Namespace changes:
  • .cells renamed to .layer.
  • .columns renamed to .homeostasis and init fn moved into .layer.
  • .protocols absorbed into .core.
  • .layer.tools created and holds cell-excitation-breakdowns.
  • .layer.params created for parameter defaults and specs - helps with docs.
  • .synapses hosts its own protocols.
  • .topology renamed to .topography and hosts its own protocols.
• column-state-freqs moved into .layer.
• PLayer gains a layer-decode-to-ff-bits* so that predict can be generic.
• Renames: p/topology, p/PTopology, p/PTopological (to p/PTopographic).
• Rename in LayerOfCells :state to :active-state, also :prior-active-state.
• Fixed predictions function, it was dropping votes from cells with index > 0.
• Boosting is now off by default - see parameters :boost-active-duty-ratio etc.

0.0.15 - 2016-09-05

Changed

• The parameter specification map now referred to as params instead of spec. To avoid confusion with clojure.spec where specs define function usage. The breaking part of this is renaming key :spec to :params in LayerOfCells.
• within layer, state key :matching-seg-paths is now :fully-matching-segs.
• protocols namespace functions active-columns, bursting-columns, active-cells, winner-cells, predictive-cells and prior-predictive-cells replaced with just layer-state, which returns all the above under keys of the same names. This seems more clojurey in avoiding inventing detailed APIs, while still allowing for different implementations... and is easier to spec.
• some protocols fns renamed adding a star*, to allow global specs on wrappers.
• params :ff-perm-init-lo, :ff-perm-init-hi now :ff-perm-init [lo hi].

Added

• minimal Javascript API in new js namespace.
• Javascript API demo in public/comportexjs.html
• function specification and testing using clojure.spec.

0.0.14 - 2016-07-01

Changed

• No more temporal pooling excitation. Instead, synapses from predicted cells excite their target cells over multiple time steps.
• No more union pooling; activation level is constant.
• First level layers are not treated any differently from higher layers.
  • Proximal synapses grow to any active synapses, not just stable ones.
  • Winner cells remain the same in continuing active columns unless reset; we may rely on an external timing signal to distinguish repeats.
  • Learn on winner cells only when they become active (even at first level). But might revisit this to learn auto-associatively for pattern completion.
• Inhibition radius is now full span of receptive field size, not half of it.
• Do not displace synapses when growing; instead wait for them to be culled.
• Spec :spatial-pooling :local-inhibition replaces :global-inhibition? true.

Added

• Spec parameter :stable-activation-steps
• State key :stable-cells-buffer
• Can create 2D regions with 3D inputs.
• Can specify custom spatial pooling function via spec key :spatial-pooling.
• Adjust overlap based on overlap duty cycle, to compare to NuPIC.
• Sampling Linear Encoder
• periodic? option to linear-encoder
• core/segs-proximal-bit-votes
• Spec parameter :inh-radius-scale
• Optional instantaneous lateral synapse graph, eg for learned lateral inhibition.
• Multimethods for customisation of algorithms, selected by spec keys :spatial-pooling and :temporal-pooling.
• Adjust overlap by activation frequency. disable by :float-overlap-duty-ratio 0
• Effective time steps by sufficient difference in columns :transition-similarity.

Removed

• Spec parameters :temporal-pooling-max-exc, -fall, -amp
• Spec parameters :activation-level-max, :stable-inbit-frac-threshold
• State keys :engaged?, :newly-engaged?
• temporal-pooling-cells protocol function.
• Spec parameter :global-inhibition?

Fixed

• coord-distance now uses Chebyshev distance to match neighbours function.
• Local inhibition failed when inhibition-base-distance >= inh-radius.
• Bursting defined by depolarisation; distinguishes bursting when depth=1.
• Distal punishment cells were incorrect.

0.0.13 - 2016-01-12

Changed

• Winner cells are not selected in layer-activate phase, only later in the layer-learn phase.
• LayerActiveState no longer has keys :distal-learning, :proximal-learning, etc.
• LayerOfCells now has key :learn-state which is a LayerLearnState record.
• Abandoned the bias against cells with existing inactive segments.
• :matching-[ff-]seg-paths now holds [seg-path exc] tuples instead of just seg-path.
• Removed :well-matching-seg-paths and :well-matching-ff-seg-paths.
• In LayerDistalState renamed :on-bits to :active-bits and :on-lc-bits to :learnable-bits
• 2D and 3D topologies now index across rows first, like image pixels.

Added

• LayerActiveState gains a key :col-active-cells, a map keyed by col id.
• Spec parameter :apical-bias-frac
• NoEncoder to pass in sparse distributed representations directly.

Fixed

• Segments with connected synapses below :new-synapse-count but above :stimulus-threshold were not automatically chosen, instead falling through to partial matches (including disconnected synapses).
• The wrong segment index was chosen for a partial match when some cells had no segments. (cell-segs was filtered)

0.0.12 - 2015-12-01

• Before this I didn't keep a change log.