Merge pull request #588 from htm-community/tm_anomaly_likelihood

TM anomaly: new modes: likelihood,...

Author: breznak

bindings/py/cpp_src/bindings/algorithms/py_TemporalMemory.cpp

@@ -47,6 +47,13 @@ Example usage:
     TODO
 )");
 
+	py::enum_<TemporalMemory::ANMode>(m, "ANMode")
+	  .value("DISABLED",   TemporalMemory::ANMode::DISABLED)
+	  .value("RAW",        TemporalMemory::ANMode::RAW)
+	  .value("LIKELIHOOD", TemporalMemory::ANMode::LIKELIHOOD)
+	  .value("LOGLIKELIHOOD", TemporalMemory::ANMode::LOGLIKELIHOOD)
+	  .export_values();
+
         py_HTM.def(py::init<>());
         py_HTM.def(py::init<std::vector<CellIdx>
                 , CellIdx
@@ -62,7 +69,8 @@ Example usage:
                 , SegmentIdx
                 , SynapseIdx
                 , bool
-                , UInt>(),
+                , UInt
+		, TemporalMemory::ANMode>(),
 R"(Initialize the temporal memory (TM) using the given parameters.
 
 Argument columnDimensions
@@ -123,6 +131,10 @@ Argument externalPredictiveInputs
     TemporalMemory.  If this is given (and greater than 0) then the active
     cells and winner cells of these external inputs must be given to methods
     TM.compute and TM.activateDendrites
+
+Argument anomalyMode (optional, default ANMode::RAW) selects mode for `TM.anomaly`.
+    Options are ANMode {DISABLED, RAW, LIKELIHOOD, LOGLIKELIHOOD}
+
 )"
                 , py::arg("columnDimensions")
                 , py::arg("cellsPerColumn") = 32
@@ -139,7 +151,8 @@ Argument externalPredictiveInputs
                 , py::arg("maxSynapsesPerSegment") = 255
                 , py::arg("checkInputs") = true
                 , py::arg("externalPredictiveInputs") = 0u
-            );
+		, py::arg("anomalyMode") = TemporalMemory::ANMode::RAW 
+		);
 
         py_HTM.def("printParameters",
             [](const HTM_t& self)

src/htm/algorithms/AnomalyLikelihood.cpp

@@ -27,6 +27,7 @@ AnomalyLikelihood::AnomalyLikelihood(UInt learningPeriod, UInt estimationSamples
         iteration_ = 0;
         NTA_CHECK(historicWindowSize >= estimationSamples); // cerr << "estimationSamples exceeds historicWindowSize";
         NTA_CHECK(aggregationWindow < reestimationPeriod && reestimationPeriod < historicWindowSize);
+	NTA_WARN << "C++ AnomalyLikelihood may still need some testing.";
     }
 
 

src/htm/algorithms/TemporalMemory.cpp

@@ -66,12 +66,14 @@ TemporalMemory::TemporalMemory(
     SegmentIdx maxSegmentsPerCell,
     SynapseIdx maxSynapsesPerSegment, 
     bool checkInputs, 
-    UInt externalPredictiveInputs) {
+    UInt externalPredictiveInputs,
+    ANMode anomalyMode) {
+
   initialize(columnDimensions, cellsPerColumn, activationThreshold,
              initialPermanence, connectedPermanence, minThreshold,
              maxNewSynapseCount, permanenceIncrement, permanenceDecrement,
              predictedSegmentDecrement, seed, maxSegmentsPerCell,
-             maxSynapsesPerSegment, checkInputs, externalPredictiveInputs);
+             maxSynapsesPerSegment, checkInputs, externalPredictiveInputs, anomalyMode);
 }
 
 TemporalMemory::~TemporalMemory() {}
@@ -91,7 +93,8 @@ void TemporalMemory::initialize(
     SegmentIdx maxSegmentsPerCell,
     SynapseIdx maxSynapsesPerSegment, 
     bool checkInputs, 
-    UInt externalPredictiveInputs) {
+    UInt externalPredictiveInputs,
+    ANMode anomalyMode) {
 
   // Validate all input parameters
   NTA_CHECK(columnDimensions.size() > 0) << "Number of column dimensions must be greater than 0";
@@ -132,6 +135,8 @@ void TemporalMemory::initialize(
   maxSegmentsPerCell_ = maxSegmentsPerCell;
   maxSynapsesPerSegment_ = maxSynapsesPerSegment;
 
+  tmAnomaly_.mode_ = anomalyMode;
+
   reset();
 }
 
@@ -483,15 +488,43 @@ void TemporalMemory::compute(const SDR &activeColumns,
   activateDendrites(learn, externalPredictiveInputsActive, externalPredictiveInputsWinners);
 
   // Update Anomaly Metric.  The anomaly is the percent of active columns that
-  // were not predicted.
-  anomaly_ = computeRawAnomalyScore(
-                activeColumns,
-                cellsToColumns( getPredictiveCells() ));
+  // were not predicted. 
+  // Must be computed here, between `activateDendrites()` and `activateCells()`.
+  switch(tmAnomaly_.mode_) {
+
+    case ANMode::DISABLED: {
+      tmAnomaly_.anomaly_ = 0.5f; 
+			   } break;
+
+    case ANMode::RAW: {
+      tmAnomaly_.anomaly_ = computeRawAnomalyScore(
+                             activeColumns,
+                             cellsToColumns( getPredictiveCells() ));
+		      } break;
+
+    case ANMode::LIKELIHOOD: {
+      const Real raw = computeRawAnomalyScore(
+                         activeColumns,
+                         cellsToColumns( getPredictiveCells() ));
+      tmAnomaly_.anomaly_ = tmAnomaly_.anomalyLikelihood_.anomalyProbability(raw);
+			     } break;
+
+    case ANMode::LOGLIKELIHOOD: {
+      const Real raw = computeRawAnomalyScore(
+                         activeColumns,
+                         cellsToColumns( getPredictiveCells() ));
+      const Real like = tmAnomaly_.anomalyLikelihood_.anomalyProbability(raw);
+      const Real log  = tmAnomaly_.anomalyLikelihood_.computeLogLikelihood(like);
+      tmAnomaly_.anomaly_ = log;
+				} break;
   // TODO: Update mean & standard deviation of anomaly here.
+  };
+  NTA_ASSERT(tmAnomaly_.anomaly_ >= 0.0f and tmAnomaly_.anomaly_ <= 1.0f) << "TM.anomaly is out-of-bounds!";
 
   activateCells(activeColumns, learn);
 }
 
+
 void TemporalMemory::compute(const SDR &activeColumns, const bool learn) {
   SDR externalPredictiveInputsActive({ externalPredictiveInputs_ });
   SDR externalPredictiveInputsWinners({ externalPredictiveInputs_ });
@@ -504,7 +537,7 @@ void TemporalMemory::reset(void) {
   activeSegments_.clear();
   matchingSegments_.clear();
   segmentsValid_ = false;
-  anomaly_ = -1.0f;
+  tmAnomaly_.anomaly_ = -1.0f; //TODO reset rather to 0.5 as default (undecided) anomaly
 }
 
 // ==============================
@@ -713,7 +746,9 @@ bool TemporalMemory::operator==(const TemporalMemory &other) const {
       winnerCells_ != other.winnerCells_ ||
       maxSegmentsPerCell_ != other.maxSegmentsPerCell_ ||
       maxSynapsesPerSegment_ != other.maxSynapsesPerSegment_ ||
-      anomaly_ != other.anomaly_ ) {
+      tmAnomaly_.anomaly_ != other.tmAnomaly_.anomaly_ ||
+      tmAnomaly_.mode_ != other.tmAnomaly_.mode_ || 
+      tmAnomaly_.anomalyLikelihood_ != other.tmAnomaly_.anomalyLikelihood_ ) {
     return false;
   }
 

src/htm/algorithms/TemporalMemory.hpp

@@ -27,6 +27,8 @@
 #include <htm/types/Sdr.hpp>
 #include <htm/types/Serializable.hpp>
 #include <htm/utils/Random.hpp>
+#include <htm/algorithms/AnomalyLikelihood.hpp>
+
 #include <vector>
 
 
@@ -35,6 +37,7 @@ namespace htm {
 using namespace std;
 using namespace htm;
 
+
 /**
  * Temporal Memory implementation in C++.
  *
@@ -53,6 +56,7 @@ using namespace htm;
 class TemporalMemory : public Serializable
 {
 public:
+  enum class ANMode { DISABLED = 0, RAW = 1, LIKELIHOOD = 2, LOGLIKELIHOOD = 3};
   TemporalMemory();
 
   /**
@@ -124,6 +128,10 @@ class TemporalMemory : public Serializable
    * TemporalMemory.  If this is given (and greater than 0) then the active
    * cells and winner cells of these external inputs must be given to methods
    * TM.compute and TM.activateDendrites
+   *
+   * @param anomalyMode (optional, default `ANMode::RAW`)from enum ANMode, how is 
+   * `TM.anomaly` computed. Options ANMode {DISABLED, RAW, LIKELIHOOD, LOGLIKELIHOOD}
+   * 
    */
   TemporalMemory(
       vector<CellIdx> columnDimensions,
@@ -140,13 +148,15 @@ class TemporalMemory : public Serializable
       SegmentIdx      maxSegmentsPerCell          = 255,
       SynapseIdx      maxSynapsesPerSegment       = 255,
       bool            checkInputs                 = true,
-      UInt            externalPredictiveInputs    = 0);
+      UInt            externalPredictiveInputs    = 0,
+      ANMode	      anomalyMode 		  = ANMode::RAW
+      );
 
   virtual void
   initialize(
-    vector<CellIdx>  columnDimensions            = {2048},
-    CellIdx          cellsPerColumn              = 32,
-    SynapseIdx       activationThreshold         = 13,
+    vector<CellIdx>  columnDimensions         = {2048},
+    CellIdx          cellsPerColumn           = 32,
+    SynapseIdx       activationThreshold      = 13,
     Permanence    initialPermanence           = 0.21,
     Permanence    connectedPermanence         = 0.50,
     SynapseIdx    minThreshold                = 10,
@@ -158,7 +168,9 @@ class TemporalMemory : public Serializable
     SegmentIdx    maxSegmentsPerCell          = 255,
     SynapseIdx    maxSynapsesPerSegment       = 255,
     bool          checkInputs                 = true,
-    UInt          externalPredictiveInputs    = 0);
+    UInt          externalPredictiveInputs    = 0,
+    ANMode        anomalyMode                 = ANMode::RAW
+    );
 
   virtual ~TemporalMemory();
 
@@ -231,6 +243,10 @@ class TemporalMemory : public Serializable
    * the TemporalMemory via its compute method ensures that you'll always
    * be able to call getActiveCells at the end of the time step.
    *
+   * Additionaly, this method computes anomaly for `TM.anomaly&`, if you
+   * use other learning methods (activateCells(), activateDendrites()) your
+   * anomaly scores will be off. 
+   *
    * @param activeColumns
    * Sorted SDR of active columns.
    *
@@ -467,7 +483,9 @@ class TemporalMemory : public Serializable
        CEREAL_NVP(activeCells_),
        CEREAL_NVP(winnerCells_),
        CEREAL_NVP(segmentsValid_),
-       CEREAL_NVP(anomaly_),
+       CEREAL_NVP(tmAnomaly_.anomaly_),
+       CEREAL_NVP(tmAnomaly_.mode_),
+       CEREAL_NVP(tmAnomaly_.anomalyLikelihood_),
        CEREAL_NVP(connections));
 
     cereal::size_type numActiveSegments = activeSegments_.size();
@@ -522,7 +540,9 @@ class TemporalMemory : public Serializable
        CEREAL_NVP(activeCells_),
        CEREAL_NVP(winnerCells_),
        CEREAL_NVP(segmentsValid_),
-       CEREAL_NVP(anomaly_),
+       CEREAL_NVP(tmAnomaly_.anomaly_),
+       CEREAL_NVP(tmAnomaly_.mode_),
+       CEREAL_NVP(tmAnomaly_.anomalyLikelihood_),
        CEREAL_NVP(connections));
 
     numActiveConnectedSynapsesForSegment_.assign(connections.segmentFlatListLength(), 0);
@@ -646,21 +666,34 @@ class TemporalMemory : public Serializable
   vector<SynapseIdx> numActiveConnectedSynapsesForSegment_;
   vector<SynapseIdx> numActivePotentialSynapsesForSegment_;
 
-  Real anomaly_;
-
   Random rng_;
 
+  /**
+   * holds logic and data for TM's anomaly
+   */
+  struct anomaly_tm {
+    protected:
+      friend class TemporalMemory;
+      Real anomaly_ = 0.5f; //default value
+      ANMode mode_ = ANMode::RAW;
+      AnomalyLikelihood anomalyLikelihood_; //TODO provide default/customizable params here
+  };
+
 public:
   Connections connections;
   const UInt &externalPredictiveInputs = externalPredictiveInputs_;
+
+  anomaly_tm tmAnomaly_;
   /*
    *  anomaly score computed for the current inputs
    *  (auto-updates after each call to TM::compute())
    *
    *  @return a float value from computeRawAnomalyScore()
    *  from Anomaly.hpp
    */
-  const Real &anomaly = anomaly_;
+const Real &anomaly = tmAnomaly_.anomaly_; //this is position dependant, the struct anomaly_tm must be defined before this use,
+// otherwise this surprisingly compiles, but a call to `tmAnomaly_.anomaly` segfaults!
+
 };
 
 } // namespace htm

src/test/unit/algorithms/ConnectionsPerformanceTest.cpp

@@ -70,10 +70,11 @@ float runTemporalMemoryTest(UInt numColumns, UInt w,   int numSequences, //TODO
 
   // learn
   for (int i = 0; i < 5; i++) {
-    for (auto sequence : sequences) {
-      for (auto sdr : sequence) {
+    for (const auto& sequence : sequences) {
+      for (const auto& sdr : sequence) {
         tm.compute(sdr, true);
-	avgAnomAfter = anom10.compute(tm.anomaly); //average anomaly score
+	const Real an = tm.anomaly;
+	avgAnomAfter = anom10.compute(an); //average anomaly score
       }
       tm.reset();
     }