v0.3.1

Added

• Package metadata for PyPI (summary, keywords, project links, classifiers) and
  CITATION.cff.

Changed

• The README introduces the whole project, links absolutely (so it also works on PyPI), shows
  the spiking sequence memory learning, and lists the limitations.

v0.3.0

Added

• Spiking temporal memory: ActiveSegments (distal dendritic segments with NMDA-like
  plateaus and a coincidence window) and MinicolumnInhibition; a layer built from them
  activates exactly the cells TemporalMemory activates, checked by
  tests/validation/test_segment_math.py and test_sequence_math.py.
• SegmentLearning: the temporal memory's learning in spike time, and sequence_memory,
  which builds a spiking sequence memory layer with derived and checked timing. It learns
  the same curves as TemporalMemory (tests/validation/test_sequence_learning.py).
• experiments/sequence_learning.py: the full learning curves, with parameters and seeds.
• SpatialPooler.state_dict()/load_state_dict() and the same for TemporalMemory
  (including its random generator), so both resume exactly.
• More of tests/validation: conv2d, local2d, lateral and pooling currents and the conv and
  local STDP rules against the dense synapse-by-synapse definition; reward-modulated STDP
  against its closed form, and the distal reward problem (Izhikevich 2007).
• benchmarks/suite.py and compare.py, run weekly by the Benchmarks workflow: throughput
  normalized by a calibration workload, compared with a baseline recorded on a GitHub
  runner (a case that halves fails).

Fixed

• STDP computed in torch's default dtype instead of the traces' dtype, so a float64
  network learned in float32 precision (and conv and local synapses got float32 updates).

v0.2.0

Added

• neurosush.htm: SDR operations with exact match probabilities, scalar, random
  distributed and category encoders, SDRClassifier, SpatialPooler, TemporalMemory,
  grid cell modules, ActiveDendrites, and voting columns for object recognition.
• neurosush.predictive_coding.PredictiveCodingNetwork: hierarchical predictive coding with
  learned weights, priors and variances.
• examples/sequence_prediction.py and examples/object_recognition.py.
• Recorder: records attributes of a network, neuron group or synapse group over time.
• neurosush.checkpoint: save and load the complete state of a network and resume it
  exactly, batched or not.
• Behavior.state_dict()/load_state_dict(), and the same for delay buffers.
• Strict mypy type checking of the package (in scripts/check.sh and CI).
• tests/validation: the spiking core checked against the closed-form solutions of its
  equations (LIF, exponential and adaptive LIF, traces, STDP, inhibitory STDP, homeostasis,
  dopamine, Poisson encoders and delays).
• Tests marked gpu that compare CUDA and CPU runs, pre-commit hooks (also run by the CI
  lint job), and Dependabot updates for pip and GitHub Actions.

Changed

• A synaptic input without SpikeGather is now an error; before, the synapse silently
  delivered no current.
• SpikeGather needs an Axon only on the source; syn.post_spike is gathered when the
  destination has one too (traces and plasticity need it and say so).
• Axon checks that the dst_delay of incoming synapses fits its history.
• A group shape in a JSON spec that is neither an int nor three ints raises ValueError.

v0.1.0

Added

• Simulation core: Network, NeuronGroup, SynapseGroup, Behavior with an explicit
  execution Order, delay buffers, and seeded randomness.
• Neurons: LIF, ELIF, AdaptiveELIF with pure dynamics, Fire, KWTA, InherentNoise,
  Axon delays, DendriteStructure and DendriteIntegration (proximal, distal, apical),
  ActivityHomeostasis, VoltageHomeostasis, and SpikeInput.
• Synapses: WeightInit (dense or sparse), DelayInit, dense, one-to-one, sparse, conv2d,
  local2d, lateral and average-pooling inputs, SpikeGather, Traces, STDP, RSTDP,
  ISTDP, WeightClip, WeightNormalization, CurrentNormalization.
• Payoff and Dopamine reward modulation.
• Poisson and latency encoders, DoG and Gabor kernels, grid masks, LocationDataset and
  spike_frames.
• Cortical structures: layers with ports, connect, CorticalColumn, and JSON specs built
  through a class registry.
• examples/two_patterns.py, which learns two input patterns with STDP.
• Fast paths: ring-buffer delays validated once, and event-driven in-place STDP for dense
  synapses (5-6x faster on CPU); benchmarks/dense_stdp.py measures it.
• Batched simulation: Network(batch_size=B) and spike_frames(..., batch_size=B) run B
  samples side by side with shared weights (about 23,000 sample-steps/s on a laptop GPU for
  the 784 -> 400 benchmark).
• CI (lint, tests on Python 3.10 to 3.13, coverage, build) and a tag-driven release workflow
  with PyPI trusted publishing.