Changelog

DRIADA follows semantic versioning and maintains a changelog of all notable changes.

The full changelog is available in the repository root:

Changelog

All notable changes to DRIADA will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

---

[1.1.0] - 2026-03-15

Recurrence analysis module and INTENSE improvements

68 commits, 136 files changed, 12,017 insertions, 1,118 deletions since v1.0.0. 2,052 tests passing on Linux, macOS, Windows (Python 3.9-3.13).

Recurrence analysis (new module)

Full nonlinear time series analysis toolkit in driada.recurrence:

• RecurrenceGraph — k-NN and epsilon-ball recurrence graphs with Theiler window exclusion, inheriting from Network for full spectral/community analysis

• VisibilityGraph — horizontal (O(N) monotone stack) and natural visibility graphs for amplitude-structure analysis

• OrdinalPartitionNetwork — ordinal rank-pattern transition networks with permutation entropy for temporal ordering analysis

• compute_rqa — recurrence quantification analysis extracting DET, LAM, ENTR, and 7 other measures from diagonal/vertical line structures

• estimate_tau / estimate_embedding_dim — automated Takens embedding parameter selection via TDMI first minimum and FNN criterion. return_fractions=True mode for diagnostic plots

• pairwise_jaccard_sparse — vectorized pairwise Jaccard similarity via sparse matmul. Accepts RecurrenceGraph objects directly; trim_to_min=True handles different-sized embeddings from per-neuron parameters

• population_recurrence_graph — combine per-neuron recurrence graphs via joint (AND threshold) or mean strategies

• TimeSeries integration — all recurrence functions available as cached methods on TimeSeries and MultiTimeSeries, with automatic tau/dim estimation and Theiler window calculation

• Optional Numba acceleration — JIT-compiled RQA scanning for large recurrence matrices

INTENSE features

• Signal ratio — quantitative selectivity strength metric for binary and linear continuous features. Measures how much stronger the neural response is during active vs baseline periods (b913762, b87d39d)

• Anti-selectivity filter — remove_anti_selective option drops neurons with signal_ratio <= 1 before disentanglement, removing suppressed responses that confuse redundancy/synergy analysis (ca68532)

• Disentanglement delay correction — optimal delays now applied during CMI computation in the disentanglement pipeline, fixing a bias where unshifted features underestimated multi-feature interactions (81cb05c)

• Exhaustive shuffle sampling — when the number of valid circular shifts is small, INTENSE now samples exhaustively instead of drawing with replacement, eliminating duplicate shuffles (20de17b)

INTENSE bug fixes

• Circular feature _2d handling — original circular features now correctly excluded from the feature bunch when their _2d (cos, sin) counterparts are present, preventing double-counting in significance testing (2ec0299)

• 2D copula axis bug — np.roll axis corrected for 2D copula-normal data in the FFT path (cef50be)

• Signal ratio delay alignment — uses np.roll instead of array truncation, preserving frame count consistency (c9ee988)

• Feature-feature pipeline return arity — compute_feat_feat_significance now returns a consistent 5-tuple across all code paths (c7f3ddb)

Dimensionality reduction

• ClassificationLoss — supervised loss component for autoencoders that trains embeddings to preserve categorical structure alongside reconstruction (46e5a75)

• FlexibleVAE latent dimension — _latent_dim now correctly passed during VAE construction (f33b31b)

• Optimizer parameter deduplication — duplicate parameter registration in autoencoder models fixed (70426b9)

• Graph preprocessing — explicit None for graph_preprocessing now respected instead of falling through to default (ecd2e45)

Performance

• FFT-accelerated TDMI — get_tdmi uses FFT cross-correlation for GCMI estimator, matching the INTENSE pipeline optimization (8a4848d)

• INTENSE pair masking — FFT cache builder skips unused neuron-feature pairs when pair_mask is provided (ca59fa8)

• Cell-cell memory — float32/int32 casting for cell-cell significance arrays reduces peak memory by 50% (bb2661c)

• ProximityGraph construction — skip directionality detection, use cKDTree for neighbor search, vectorized sparse operations (66a327a, 6cd8823)

Tutorial notebook

• Notebook 06: Recurrence analysis — Colab-ready tutorial covering delay embedding, recurrence plots, RQA, three graph representations (RG, HVG, OPN), windowed regime detection, and population module recovery via Jaccard similarity network. Full ReadTheDocs API cross-links and metric explanations

Documentation

• Recurrence API reference — 7 Sphinx pages covering all recurrence submodules (embedding, recurrence_graph, rqa, population, visibility, opn, plotting)

• Notebook validation protocol — added checks for API cross-links and user-friendly concept explanations

• Example validation — all 25 examples re-verified; output path fixes, ASCII compliance

---

[1.0.0] - 2026-03-01

First stable release

95 commits, 282 files changed, 30,481 insertions, 9,826 deletions since v0.7.2. 1,875 tests passing on Linux, macOS, Windows (Python 3.9-3.13).

INTENSE performance

• FFT cache build 2x faster — precompute per-signal FFT and std once (520 unique signals instead of 20,000 pairs for a 500-neuron × 20-feature problem), then combine via cross-correlation. Bit-exact results. (9d865c3)

• Stage 2 scan 1.4x faster — replaced scipy.stats.rankdata with count-based comparison, vectorized per-neuron cache lookup to eliminate Python inner loop. (fe9bb55)

• FFT-accelerated av metric — average-value selectivity metric now uses the FFT path, matching cc/pearson_cc performance. (7e37c5b)

INTENSE features

• Feature shuffle masks — independent circular-shift exclusion zones for calcium (CA_SHIFT_N_TOFF=5 frames) and features (MIN_FEAT_SHIFT_SEC=2s), replacing the single global minimum shift. Prevents trivial self-correlation at small shifts without over-restricting the shuffle range. (8744e52)

• feature_types parameter — override auto-detected feature subtypes (e.g., force head_direction to circular, trial_type to categorical) in load_exp_from_aligned_data. (98da360)

• Automatic circular → (cos, sin) encoding — circular features are automatically transformed to 2D representations for MI estimators that assume Euclidean distance. Controlled by create_circular_2d parameter. (aca9d69)

• mi_estimator_kwargs passthrough — thread estimator-specific parameters (e.g., KSG neighbors, LNC alpha) through the entire INTENSE pipeline. (9075eee)

• MTS delay optimization — MultiTimeSeries features (e.g., 2D position) now support optimal delay search. (2b0f9d4)

• Post-hoc pre_pval reconstruction — reconstruct Stage 1 p-values from saved shuffle data without re-running the analysis. (7a6b6e4)

INTENSE refactoring

• intense_base.py factored into 4 modules — extracted validation (intense_validation.py), FFT cache management (fft.py), multiple-comparison correction (corrections.py), and delay optimization (delay.py). The monolithic file is now the orchestrator only. (c5f5d59, b45cd1b)

• Removed deprecated parameters — joint_distr (use find_joint=True instead) and allow_mixed_dimensions (MTS delay optimization makes this unnecessary). Clean deprecation with no silent fallbacks. (afe68f2, f60e488)

Experiment & data loading

• Experiment.add_feature() — register dynamic features after initialization with full attribute binding, shuffle mask application, and hash rebuild. Fixes disentanglement invisibility when features are added via dict assignment. (520bc27)

• 3D place cell support — volumetric spatial manifold generator for 3-photon and light-sheet experiments. (8842516)

• Experiment partitioning — split experiments by time windows or trial indices for cross-validation. (3f44f77)

• Circular feature rescaling — detects and rescales [0,1]-normalized circular features to [0, 2π] before INTENSE analysis, preventing silent MI underestimation. (7edf26e)

Tutorial notebooks (new)

Six Colab-ready notebooks with auto-generation scripts (tools/create_notebook_0X.py):

0. DRIADA overview — Core data structures, quick tour of INTENSE, dimensionality reduction, and networks (0f9b806)

1. Data loading & neuron analysis — Experiment construction, spike reconstruction methods comparison, surrogate generation (37b71d4)

2. Selectivity detection with INTENSE — Two-stage testing, MultiTimeSeries features, circular features, multiple comparison correction (38fa6c4)

3. Population geometry & DR — MVData API, method comparison benchmark (PCA, Isomap, UMAP), sequential DR, autoencoder DR, circular manifold extraction, INTENSE-guided DR (df27aef)

4. Network analysis — Network construction, spectral decomposition, community detection, thermodynamic entropy, ProximityGraph from DR (e1515fc)

5. Advanced capabilities — INTENSE + DR pipeline, leave-one-out importance, RSA, RNN analysis (b4d2b15)

All notebooks verified executable via tools/verify_notebooks.py. Cross-navigation links and API documentation links. Sentence case throughout.

Documentation overhaul

• 28 docstring-signature mismatches fixed — parameter names, types, and counts verified against source across all public functions. (ce1065c)

• All Sphinx cross-references fixed — broken :func:, :class:, :meth: references and non-clickable See Also entries corrected. (c5d6112, a05c655)

• Quickstart rewritten — removed references to deleted functions, fixed undefined variables in code examples. (09dcbcb)

• README rewrite — added data requirements section, pipeline examples, Colab badges, macOS CI badge, synced citation sections. (5244ec3, 33b98fd, 235874f)

• Network-DR relationship documented — ProximityGraph inherits from Network; documented in module docstrings, Sphinx RST, and notebook 03/04 cross-references. (c6f5fd8, b223d3e, 7f89c8b, c18193b)

• INTENSE math framework — added FFT-accelerated permutation testing explanation to docs. (34c25dc)

Test quality

• 36 RED findings resolved — zero-assertion tests, loose tolerances for exact results, dead code, and lowered-bar comments identified by tools/validate_tests.py and fixed. (add6f37, c57e700, cec65cd)

• Flaky CI tests fixed — RSA cache division by zero, diffusion maps threshold too tight for cross-platform variance, JIT benchmark timing resolution on Windows, warmup fixture durations too short for shuffle mask exclusion zones. (aefcc22, 4d46560, e70353c)

• Python 3.13 CI — added to test matrix. (4aa1b36)

Bug fixes

• Shift index out-of-bounds after downsampling — invalid-shift indices could exceed array length when ds > 1; now clipped. (6aeeff7)

• Ground truth _2d names — synthetic generator expected_pairs now use _2d suffixed names for circular features, matching the auto-created 2D representations. (a9ee6ef)

• t-SNE/UMAP parameter forwarding — perplexity and random_state were silently swallowed by **kwargs; now properly forwarded. (28533cd)

• IntenseResults JSON round-trip — dict keys converted from int to string during serialization; now handled correctly on load. (04b7690)

• NumPy scalar compatibility — ca_mse_error returned ndarray instead of scalar on newer NumPy; extracted with .item(). (eebcc0a)

• Network.n_cc guard — connected components count now handles uninitialized graph state. (28533cd)

• Bare except narrowed — neuron.py:2826 catch-all replaced with except (ValueError, TypeError). (4aa1b36)

• Gamma-ZI fit warning — silent p=1.0 fallback on MLE failure now emits RuntimeWarning. (4aa1b36)

• Windows encoding — emoji in doc tool output replaced with ASCII; em dash bytes in RST fixed. (f20fed4, e70353c)

Code quality & refactoring

• FFT module reorganization — info_fft.py (1,130 lines) and info_fft_utils.py (256 lines) extracted from info_base.py. rfft/irfft replacing fft/ifft for 50% speedup. 100-1000x memory reduction in MTS functions. Shared helpers eliminate ~40% code duplication. All public APIs preserved via re-exports. (info_base.py reduced from 3,246 to 2,311 lines)

• Neuron module reorganization — neuron.py (3,807 lines) factored into calcium_kinetics.py, signal_preprocessing.py, event_detection.py. Backward compatible via static method delegation.

• Black + isort applied across entire codebase (177 files)

• Removed MDS and t-SNE from notebook 03 — too slow for educational use; replaced with Laplacian Eigenmaps in visualization. (fa43796)

• visualize_circular_manifold exported from public API. (28533cd)

Tooling

• tools/verify_notebooks.py — executes all notebook code cells against current API, catching broken examples before release. (fa43796)

• INTENSE validation benchmark — SNR × skip_prob parameter sweep across 315 configurations comparing GCMI vs Pearson methods. (ef00bc2, 844905f)

• tools/validate_tests.py — automated test quality checker flagging zero-assertion tests, loose tolerances, and dead code. (add6f37)

Project governance (new)

• Academic paper (paper.md) with literature review covering NIT, MINT, FRITES, CaImAn, Suite2p, CEBRA

• Bibliography (paper.bib) with 38 references

• Publications list (PUBLICATIONS.md) — 8 research papers using DRIADA

• CONTRIBUTING.md, CODE_OF_CONDUCT.md, CITATION.cff, AUTHORS.yaml

• PR template (.github/pull_request_template.md)

---

[0.7.2] - 2026-01-21

🔧 Batch processing CLI and disentanglement analysis enhancements

Performance

• Discrete-Discrete FFT Vectorization - 50-100x speedup for mi_dd_fft computation via vectorized operations

• FFT Memory Optimization - Up to 200x memory reduction in FFT modules through improved buffer management

• MI Value Caching - Pre-computed MI values cached in FFT cache for 4-5x speedup in repeated lookups

• Disentanglement Parallelization - MI lookup tables and parallel FFT cache; parallelized per-neuron processing

• IPC Overhead Reduction - Split fft_cache per worker to reduce inter-process communication in parallel jobs

• Feature Copula Pre-computation - Pre-compute copula cache before parallel loop to avoid redundant computation

• Calcium Signal Vectorization - Vectorized convolution for spike-to-calcium signal reconstruction

Features

• Batch Processing CLI

  • --dir option for processing multiple experiment files in a directory

  • --skip-computed flag to resume interrupted batch runs, loading existing stats from CSV

  • --parallel-backend option with global PARALLEL_BACKEND config (loky, threading, multiprocessing)

  • Backend-specific parallel config with pre_dispatch tuning for memory control

  • Timeout wrapper for autonomous batch processing using subprocess.Popen polling

  • Incremental batch summary saving after each file (fault tolerance)

  • Automatic cache clearing for memory management in long-running jobs

  • Upgraded to joblib 1.5.0 with idle_worker_timeout for worker cleanup

• Disentanglement Analysis

  • Composable pre-filters for disentanglement analysis (spatial, temporal, threshold-based)

  • Filter examples added to run_intense_analysis script

  • Disentangled stats and significance output tables for publication-ready results

• I/O & Data Loading

  • New INTENSE I/O module using NPZ format (removes HDF5 dependency)

  • Tool for loading synchronized experiment .npz files

  • Experiment configs for selectivity_dynamics package

• Profiling

  • Internal timing profiling for INTENSE pipelines

  • Timing instrumentation for FFT cache building

Fixed

• Windows Process Handling - Changed joblib backend to loky to prevent CMD hanging; use Popen polling to prevent timeout wrapper freeze

• Memory Leaks - Added gc.collect() after each experiment to prevent memory accumulation in batch processing

• Zero Column Errors - Use aggregate_multiple_ts in exp_build to add noise and avoid degenerate feature matrices

• Disentanglement Robustness - Improved numerical stability with spatial filter fallbacks

• Google API Warning - Suppressed spurious authentication warning

• Test Stability - Relaxed flaky test tolerances for CI reliability

Changed

• Consolidated IABS filename parsing into driada.utils.naming module

• Removed HDF5 dependency from INTENSE results (replaced with NPZ format)

---

[0.7.1] - 2026-01-17

🚀 Enhanced performance and data loading capabilities

Performance

• MTS-MTS FFT Acceleration - 3-200x speedup for mutual information computation between MultiTimeSeries pairs (d1+d2 ≤ 6). Uses FFT-based cross-covariance computation with block determinant formula via Schur complement. Enables efficient analysis of multivariate neural dynamics (e.g., 2D position vs 2D neural trajectory). Complexity: O(d₁ × d₂ × n log n + nsh × d³).

• Memory Optimization for INTENSE - New store_random_shifts parameter (default: False) reduces memory usage by ~400MB for typical datasets (N=500 neurons, M=20 features, nsh=10,000). Random shift indices are not persisted unless explicitly requested, while preserving all MI values for analysis. Added memory_usage() method to IntenseResults for diagnostics.

Features

• Experiment Building Enhancements

  • aggregate_features parameter in load_exp_from_aligned_data() enables combining multiple 1D features into MultiTimeSeries before building Experiment objects (e.g., {('x', 'y'): 'position'})

  • Reserved key filtering prevents neural data keys (calcium, spikes, sp, asp, reconstructions) and metadata (_metadata, _sync_info) from being treated as behavioral features

  • Improves workflow for analyzing relationships between position/velocity and neural activity

Fixed

• Critical FFT Conjugation Bug - Fixed incorrect FFT cross-correlation direction in compute_mi_mts_mts_fft that caused wrong MI values at non-zero shifts. Discovered during comprehensive verification testing. Changed from conj(fft_x1) * fft_x2 to fft_x1 * conj(fft_x2) to correctly match np.roll semantics.

• Documentation - Removed invalid cross-references for missing functions and user guide sections

• Test Stability - Fixed flaky CI tests by adding proper imports, fixed random seeds, and increased connectivity parameters in manifold tests

Added

• Comprehensive test suite for MTS-MTS FFT (27 unit tests + 3 integration tests) verifying correctness at all shifts with rtol=1e-7

• Benchmark script (tools/benchmark_mts_mts_fft.py) demonstrating speedup characteristics across different parameter regimes (n=1,000 to n=60,000, nsh=100 to nsh=10,000)

---

[0.7.0] - 2026-01-13

🚀 Major performance and stability release

Performance

• FFT Acceleration for INTENSE - 10-100x speedup through FFT-based mutual information computation across shuffle loops, delay calculation, and discrete-continuous MI operations. Unified FFT pipeline with intelligent caching.

Critical Fixes

• Cache Collision Bug - Fixed critical bug where all neurons returned identical MI values due to generic TimeSeries naming causing FFT cache collisions. Implemented comprehensive naming strategy ensuring unique cache keys.

Added

• Zero-Inflated Gamma (ZIG) distribution for improved MI null hypothesis testing

• Tuned Selectivity Generator with disentanglement analysis metrics for INTENSE pipeline

• Configurable Random Seeds for graph construction ensuring reproducible manifold learning experiments

• Publication Framework Enhancements - Extended StylePreset with sizing parameters and accessor methods

Changed

• Synthetic Data Module - Consolidated all generators into unified generators.py with standardized signatures, explicit kwargs, and improved reproducibility

• Examples Updated - Validated and improved 10+ examples including complete RSA rewrite using tuned selectivity generator, enhanced Network Analysis with spectral properties, and updates across INTENSE, dimensionality reduction, and manifold examples

• Code Quality - Applied Black formatter (line-length 100) across codebase

Statistics

• 215 files changed, 20,762 insertions, 14,960 deletions

• 41 new tests added (FFT cache, multicomp correction, MTS FFT)

• Test coverage: intense_base 81.85%, info_base 88.48%

• All 1,681 tests passing

---

[0.6.6] - 2025-12-01

Added

• Dependency Management

  • Lazy torch imports with conditional GPU support for wavelet detection

  • Reduces import overhead and enables CPU-only environments without requiring PyTorch installation

  • GPU toggle allows users to control whether torch-based wavelet detection uses CUDA acceleration

Changed

• Distribution Optimization

  • Excluded images and data files from PyPI distribution package

  • Reduced tarball size to <1 MB (was >10 MB with example data)

Fixed

• Test Stability

  • Increased shuffle iterations to 1000 for 2D manifold tests (previously failed intermittently with default 100 shuffles)

  • Ensures consistent statistical power in null hypothesis testing

---

[0.6.5] - 2025-12-01

Added

• Wavelet Optimization

  • Batch optimization mode for wavelet-based event detection

  • Vectorized processing of multiple neurons simultaneously (10-50× speedup for large datasets)

  • New reconstruction workflow balancing accuracy and computational efficiency

Fixed

• Dimensionality Reduction

  • Relaxed diffusion maps variance test to check total variance preservation trend

  • Previously failed due to strict per-eigenvalue thresholds; now validates overall manifold structure

Changed

• Resolved merge conflicts from parallel feature development branches

---

[0.6.4] - 2025-11-01

Added

• Publication Tools

  • Multi-panel figure framework for generating journal-quality visualizations

  • Precise physical sizing control (inches/cm) for manuscript submissions

  • Customizable subplot layouts and spacing

• Testing

  • Comprehensive test suites for quantum metrics, manifold analysis, wavelet detection, and neuron quality assessment

  • Improved coverage for edge cases in signal processing

Changed

• Spike Reconstruction

  • Major refactor introducing iterative threshold detection algorithm

  • Alternates between threshold estimation and spike extraction until convergence

  • Balances sensitivity (detecting weak events) and specificity (avoiding false positives)

Fixed

• Statistical Testing

  • Eliminated spurious correlations in INTENSE null hypothesis test

  • Proper surrogate generation now preserves temporal autocorrelation structure

  • Ensures accurate Type I error control during Holm-Bonferroni correction

• Event Detection

  • Unified SNR calculation between wavelet and threshold methods (previously reported inconsistent quality metrics)

  • Metrics (amplitude, correlation, SNR) now compatible with both detection approaches

  • Threshold computation uses original signal instead of residuals (prevents threshold drift during iteration)

• Kinetics Optimization

  • Fixed silent failures for signals with long decay time constants (t_off > 200 frames)

  • Added safety checks capping kernel length at 2000 frames

  • Warns users about suspicious t_off values incompatible with typical calcium indicators

Improved

• Test tolerance relaxation for numerical edge cases

• .gitignore specificity for development artifacts

---

[0.6.1] - 2025-10-01

Added

• Frame Rate Adaptivity

  • Complete FPS (frames per second) adaptivity implementation

  • Automatic time constant conversion for varying imaging frequencies

  • Handles datasets from 5 Hz (slow volumetric) to 100+ Hz (fast planar) imaging

• Development Environment

  • Conda environment specification (environment.yml) for reproducible setup

  • Pins exact versions of scientific Python stack (numpy, scipy, scikit-learn, etc.)

Fixed

• Kinetics Optimization

  • Event regions now preserved during rise/decay time constant fitting

  • Previously could lose detected calcium transients when re-optimizing parameters

  • Added regression tests to prevent future breakage

Changed

• Added Neuron class usage example demonstrating end-to-end calcium imaging workflow

• Cleaned up accidentally committed debug and analysis artifacts

---

[0.6.0] - 2025-09-01

Added

• Kinetics Analysis System

  • Automated optimization of calcium indicator rise/decay time constants (τ_rise, τ_off)

  • Fits double-exponential kernel to observed calcium dynamics

  • Uses RMSE minimization between observed and reconstructed calcium traces

• Event Quality Metrics

  • Amplitude extraction from wavelet coefficients and threshold crossings

  • SNR (Signal-to-Noise Ratio) calculation for transient quality assessment

  • Template correlation metrics for validating event shape consistency

  • Deconvolution support for spike inference from calcium traces

Fixed

• Critical Reconstruction Bugs

  • Amplitude extraction from wavelet coefficients (was systematically underestimating event strength)

  • SNR interface unification across wavelet and threshold detection methods

  • Kinetics parameter propagation through reconstruction pipeline

Changed

• Completed major code quality fixes preparing for public pre-release

• Unified SNR calculation interface

---

[0.5.1] - 2025-08-01

Added

• Cross-Validation Framework

  • Leave-one-out (LOO) dimensionality reduction analysis

  • Validates manifold structure stability when systematically excluding neurons

  • Integration with INTENSE for selectivity-guided validation

• Visualization Enhancements

  • make_beautiful() function for publication-quality plot styling

  • Customizable legend positioning (upper/lower × left/right, outside options)

  • Improved default aesthetics (font sizes, line widths, color schemes)

• Data Management

  • Google Drive share link support for collaborative data sharing

  • router_source parameter in experiment loading for flexible data backends

  • Support for multiple simultaneous data sources in IABS router

• Analysis Features

  • Downsampling support for LOO analysis on large datasets (reduces computation 10-100×)

  • Embedding visualization with automatic dimension detection

Fixed

• Quality metrics validation for calcium imaging pipelines

• Test suite stability improvements

---

[0.4.0] - 2025-07-01

Added

• CI/CD Infrastructure

  • GitHub Actions workflows for automated testing

  • Codecov integration for coverage tracking

  • Test coverage requirements: 80%+ for merging to main branch

• Comprehensive Testing

  • Coverage sprint achieving 80%+ across all major modules:

    • information/: 88.7% (INTENSE, GCMI, entropy functions)

    • dim_reduction/: 87.3% (PCA, Isomap, UMAP, Diffusion Maps)

    • network/: 96.2% (graph construction, topology metrics)

    • experiment/: extensive exp_base.py and neuron.py tests

  • Import tests ensuring all modules load without errors

  • Edge case handling for numerical stability

• Dimensionality Reduction

  • Diffusion Maps: complete eigen-decomposition implementation with heat kernel affinities

  • Intrinsic dimension estimation: correlation dimension, k-NN fractal dimension

  • Graph preprocessing: automatic defaults (e.g., UMAP requires connected graph, Isomap handles disconnected)

  • API modernization: dr_sequence() replacing legacy dr_series with better error messages

• Information Theory

  • KSG estimator: Kraskov-Stögbauer-Grassberger k-NN mutual information as alternative to GCMI

  • Estimator selection: mi_estimator parameter in INTENSE pipelines (‘gcmi’ or ‘ksg’)

  • Automatic LNC: local non-uniformity correction alpha selection for KSG

  • JIT optimization: Numba compilation for GCMI functions (10-100× speedup for large datasets)

• Network Analysis

  • Gromov hyperbolicity: curvature calculation for network geometry

  • Partial directions: construct graphs using subset of features

  • Topology metrics: clustering coefficient, path lengths, community structure

• Neural Network Models

  • Flexible autoencoders: modular loss composition system

  • Loss functions: reconstruction MSE, orthogonality penalties, correlation losses

  • Architecture options: skip connections, batch normalization, dropout

  • Time series support: advanced type detection for temporal vs. static data

Changed

• API Improvements

  • Dimensionality reduction functions accept verbose parameter (suppresses progress print statements)

  • Matrix utilities automatically branch between dense numpy and sparse scipy implementations

  • Simplified parameter passing with better validation

• RSA Module

  • Major refactoring improving code clarity

  • Better integration with dimensionality reduction workflows

  • Simplified correlation distance calculations

Fixed

• Critical Bugs

  • Manifold metric numerical stability (eliminated NaN propagation)

  • ProximityGraph edge weight doubling in k-NN construction (was creating 2× heavier graphs)

  • Lost nodes calculation in graph giant component extraction

  • Zero columns in discrete attribute handling (now raises informative error)

• Precision Issues

  • Floating-point formatting in axis labels and tick marks

  • Matrix condition number checks before inversion

Removed

• Deprecated code

  • signals module (functionality migrated to utils)

  • Legacy dr_series function (use dr_sequence instead)

---

[0.2.0] - 2025-03-01

Added

• Synthetic Data Generators

  • Mixed selectivity populations: neurons responding to multiple task variables with ground truth

  • 3D spatial manifolds: volumetric place cell populations for navigation studies

  • 2D spatial manifolds: planar place fields with adjustable coverage and overlap

  • Circular manifolds: head direction cells with von Mises tuning curves

  • Validation framework: compare dimensionality reduction against known ground truth geometry

• Dimensionality Estimation

  • Linear methods:

    • PCA-based effective dimension (proportion of variance threshold)

    • Effective rank (entropy of eigenvalue distribution)

  • Nonlinear methods:

    • k-NN fractal dimension (scales with neighbor distance)

    • Correlation dimension (box-counting on embedding)

  • Circular extraction: identify ring-like manifolds from population activity

• Documentation

  • Interactive Jupyter notebooks demonstrating end-to-end workflows

  • README overhaul explaining DRIADA’s vision: bridging single-neuron selectivity and population dynamics

  • Example gallery with 15+ analysis pipelines

Changed

• Project Scope

  • Unified framework integrating INTENSE (single-neuron selectivity) with dimensionality reduction (population structure)

  • Renamed internal documentation from _intense_backlog.md to _backlog.md reflecting broader mission

Fixed

• 3D spatial manifold generation stability (optimized firing rate parameters)

• Test numerical tolerance for nonlinear dimension estimators

---

[0.1.x] - 2024-2025

Initial development releases establishing core functionality

INTENSE Module (Information-Theoretic Evaluation of Neuronal Selectivity)

Statistical Framework

• Gaussian Copula Mutual Information (GCMI): continuous estimator handling arbitrary distributions

• Two-stage testing:

  • Stage 1: rank-based non-parametric test (fast, conservative)

  • Stage 2: parametric gamma distribution fitting (precise p-values for significant relationships)

• Multiple comparison correction: Holm-Bonferroni controlling family-wise error rate

• Optimal shift search: detect delayed relationships (e.g., sensory → motor lag)

• Mixed selectivity disentanglement: separate unique vs. shared information using interaction information

Analysis Workflows

• Cell-cell interaction information (synergy, redundancy)

• Feature-feature joint encoding

• Single-cell selectivity for behavioral variables

• Parallel processing for large-scale multi-neuron recordings

Dimensionality Reduction Framework

Classical Methods

• PCA: principal component analysis with variance-based selection

• Factor Analysis: probabilistic latent variable model with noise modeling

Manifold Learning

• Isomap: geodesic distance preservation via graph shortest paths

• UMAP: uniform manifold approximation with fuzzy topology

• t-SNE: t-distributed stochastic neighbor embedding for visualization

• Diffusion Maps: spectral embedding based on random walk diffusion

Graph Construction

• Heat kernel affinities with adaptive bandwidth

• k-NN and ε-ball proximity graphs

• Giant component extraction for disconnected graphs

• Configurable distance metrics (Euclidean, correlation, etc.)

Calcium Imaging Pipeline

Event Detection

• Wavelet-based: multi-scale decomposition with adaptive thresholding

• Threshold-based: iterative baseline-crossing detection

• Quality metrics: amplitude, SNR, shape correlation

Signal Reconstruction

• Double-exponential kernel convolution (rise + decay dynamics)

• Spike-to-calcium forward model

• Parameter optimization for indicator kinetics

Experiment Framework

Data Management

• Unified container: neural activity + behavioral variables + metadata

• Multi-cell support: analyze 10-1000+ neurons simultaneously

• HDF5 serialization: efficient storage for large datasets (>1 GB recordings)

• Flexible loading: support for various acquisition systems (Inscopix, Miniscope, 2-photon)

Integration

• Seamless INTENSE → DR pipeline: map selectivity onto population manifolds

• Synthetic data validation: test pipelines on ground truth before real data

Infrastructure

Development

• Python package structure (setuptools/pip installable)

• Modular architecture: independent modules for information theory, DR, experiment handling

• Public API via __init__.py exports

Testing

• pytest framework with fixtures

• Unit tests for mathematical functions

• Integration tests for full workflows

• Doctests in function docstrings

Documentation

• Sphinx-based HTML documentation

• NumPy-style docstrings

• Examples embedded in docstrings

• Reference API documentation

---

Alternatively, you can view the changelog on GitHub.

Version History

• 1.1.0 (2026-03-15) - Recurrence analysis module, INTENSE signal ratio, supervised DR

• 1.0.0 (2026-03-01) - First stable release

• 0.6.6 (2025-12-XX) - Lazy torch imports, distribution optimization

• 0.6.5 (2025-12-XX) - Wavelet batch optimization

• 0.6.4 (2025-11-XX) - Publication figure framework, spike reconstruction refactor

• 0.6.1 (2025-10-XX) - FPS adaptivity, conda environment

• 0.6.0 (2025-09-XX) - Kinetics optimization system

• 0.5.1 (2025-08-XX) - LOO analysis, visualization enhancements

• 0.4.0 (2025-07-XX) - CI/CD, comprehensive testing, diffusion maps

• 0.2.0 (2025-03-XX) - Synthetic data generators, dimensionality estimation

• 0.1.x (2024-2025) - Initial development releases