Modified the function to get w to a substraction of logarithmis instead

connectivity_functions.py

@@ -4,6 +4,11 @@
 epsilon = 1e-10
 
 
+def log_epsilon(x):
+
+    return np.log(np.maximum(x, epsilon))
+
+
 def calculate_probability(patterns):
     """
     Returns the probability from a list of patterns to be learned
@@ -36,17 +41,13 @@ def calculate_coactivations(patterns):
 
 
 def get_w(P, p, diagonal_zero=True):
-    outer = np.outer(p, p)
-    P_copy = np.copy(P)
-
-    outer[outer < epsilon**2] = epsilon**2
-    P_copy[P < epsilon] = epsilon**2
 
-    w = np.log(P_copy / outer)
+    outer = np.outer(p, p)
 
-    #IPython.embed()
+    w = log_epsilon(P) - log_epsilon(outer)
     if diagonal_zero:
         w[np.diag_indices_from(w)] = 0
+
     return w
 
 
@@ -65,20 +66,6 @@ def get_w_pre_post(P, p_pre, p_post, diagonal_zero=True):
     return w
 
 
-def get_w_protocol1(P, p):
-    p_copy = np.copy(p)
-    P_copy = np.copy(P)
-
-    p_copy[p < epsilon] = epsilon
-    P_copy[P < epsilon] = epsilon * epsilon
-
-    aux = np.outer(p_copy, p_copy)
-    w = np.log(P_copy / aux)
-    # IPython.embed()
-
-    return w
-
-
 def get_beta(p):
 
     probability = np.copy(p)
@@ -122,4 +109,36 @@ def softmax(x, t=1.0, minicolumns=2):
     dist = e / np.sum(e, axis=1)[:, np.newaxis]
 
     dist = np.reshape(dist, x_size)
-    return dist
+    return dist
+
+################
+# Old functions
+#################
+
+def get_w_old(P, p, diagonal_zero=True):
+    outer = np.outer(p, p)
+    P_copy = np.copy(P)
+
+    outer[outer < epsilon**2] = epsilon**2
+    P_copy[P < epsilon] = epsilon**2
+
+    w = np.log(P_copy / outer)
+
+    #IPython.embed()
+    if diagonal_zero:
+        w[np.diag_indices_from(w)] = 0
+    return w
+
+
+def get_w_protocol1(P, p):
+    p_copy = np.copy(p)
+    P_copy = np.copy(P)
+
+    p_copy[p < epsilon] = epsilon
+    P_copy[P < epsilon] = epsilon * epsilon
+
+    aux = np.outer(p_copy, p_copy)
+    w = np.log(P_copy / aux)
+    # IPython.embed()
+
+    return w

connectivity_visualization.py

@@ -0,0 +1,50 @@
+from __future__ import print_function
+
+import numpy as np
+import matplotlib.pyplot as plt
+import matplotlib.gridspec as gridspec
+from mpl_toolkits.axes_grid1 import make_axes_locatable
+
+from connectivity_functions import get_beta, get_w, get_w_old, softmax
+from connectivity_functions import calculate_probability, calculate_coactivations
+from data_transformer import build_ortogonal_patterns
+from network import BCPNN
+
+np.set_printoptions(suppress=True)
+np.set_printoptions(precision=2)
+
+hypercolumns = 3
+minicolumns = 3
+
+patterns_dic = build_ortogonal_patterns(hypercolumns, minicolumns)
+patterns = list(patterns_dic.values())[0:4]
+
+
+# The total probability
+p = calculate_probability(patterns)
+P = calculate_coactivations(patterns)
+
+w = get_w(P, p)
+x = np.outer(p, p)
+w_aux = get_w_old(P, p)
+
+# Plot it
+cmap = 'coolwarm'
+aux_max = np.max(np.abs(w))
+
+fig = plt.figure(figsize=(16, 12))
+ax1 = fig.add_subplot(121)
+# plt.imshow(w, cmap=cmap, interpolation='None', vmin=-aux_max, vmax=aux_max)
+im1 = ax1.imshow(w, cmap=cmap, interpolation='None')
+divider = make_axes_locatable(ax1)
+cax1 = divider.append_axes('right', size='5%', pad=0.05)
+fig.colorbar(im1, ax=ax1, cax=cax1)
+
+
+ax2 = fig.add_subplot(122)
+im2 = ax2.imshow(w_aux, cmap=cmap, interpolation='None')
+divider = make_axes_locatable(ax2)
+cax2 = divider.append_axes('right', size='5%', pad=0.05)
+fig.colorbar(im2, ax=ax2, cax=cax2)
+
+plt.show()

notebooks/2016-09-05(vanilla BCPNN and digit recall).ipynb

@@ -445,7 +445,7 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python [default]",
+   "display_name": "Python 2",
    "language": "python",
    "name": "python2"
   },

notebooks/2016-10-01(Convergence Properties).ipynb

@@ -316,7 +316,7 @@
  "metadata": {
   "anaconda-cloud": {},
   "kernelspec": {
-   "display_name": "Python [default]",
+   "display_name": "Python 2",
    "language": "python",
    "name": "python2"
   },

play.py

@@ -25,7 +25,7 @@
 beta = get_beta(p)
 
 dt = 0.01
-T_simulation = 1.0
+T_simulation = 10.0
 T_training = 1.0
 simulation_time = np.arange(0, T_simulation + dt, dt)
 training_time = np.arange(0, T_simulation + dt, dt)
@@ -38,11 +38,8 @@
 g_beta_set = np.arange(0, 22, 2)
 g_w_set = np.arange(0, 12, 2)
 
-nn = BCPNN(hypercolumns, minicolumns, beta, w, g_a=0, g_beta=5.0, g_w=3.0, g_I=10.0, prng=prng)
+nn = BCPNN(hypercolumns, minicolumns, beta, w, g_a=1, g_beta=1.0, g_w=1.0, g_I=10.0, prng=prng)
 nn.randomize_pattern()
-history = nn.run_network_simulation(training_time, save=True)
-o = history['o']
-s = history['s']
 
 if False:
     final_states = []
@@ -54,7 +51,7 @@
         final_states.append(nn.o)
 
 
-if True:
+if False:
     I = patterns[0]
 
     for i in range(20):
@@ -69,7 +66,7 @@
 
 # x, y = calculate_convergence_ratios(nn, 10, training_time, patterns)
 
-if False:
+if True:
     for pattern in reversed(patterns):
         I = pattern
         nn.randomize_pattern()
@@ -91,3 +88,10 @@
 
 
 
+history = nn.run_network_simulation(simulation_time, save=True)
+o = history['o']
+s = history['s']
+
+plt.imshow(o, aspect='auto', interpolation='None')
+plt.colorbar()
+plt.show()

test_connectivity.py

@@ -119,7 +119,7 @@ def test_unit_probabilities(self):
 
     def test_coactivations(self):
         test_pattern1 = np.array((1, 0, 1, 0))
-        test_pattern2 = np.array((0, 1, 0, 1))
+        test_pattern2 = np.array((0, 1,  0, 1))
 
         patterns = [test_pattern1, test_pattern2]