import random as rng import matplotlib.pyplot as plt import noise import numpy as np from matplotlib.patches import Ellipse from tqdm import tqdm from brian2.units import * from scripts.spatial_maps.orientation_map import OrientationMap class Pickle: def __init__(self, x, y, a, b, phi): self.x = x self.y = y self.a = a # Semimajor axis self.b = b # Semiminor axis self.phi = phi self.c = 2.0 * a # Sum distance self.foc_len = np.sqrt(a ** 2 - b ** 2) self.area = np.pi * a * b def get_p1(self): x1 = self.x - self.foc_len * np.cos(self.phi) y1 = self.y - self.foc_len * np.sin(self.phi) return x1, y1 def get_p2(self): x2 = self.x + self.foc_len * np.cos(self.phi) y2 = self.y + self.foc_len * np.sin(self.phi) return x2, y2 def get_ellipse(self): ell = Ellipse((self.x, self.y), 2.0 * self.a, 2.0 * self.b, angle=self.phi * 180. / np.pi, linewidth=1, fill=False, zorder=2) return ell def get_area(self): return np.pi * self.a * self.b def plot_neural_sheet(ex_positions, ex_tunings, axonal_clouds=None, delta_xy=[]): X, Y = get_position_mesh(ex_positions) n_ex = int(np.sqrt(len(ex_positions))) head_dir_preference = np.array(ex_tunings).reshape((n_ex, n_ex)) fig = plt.figure() ax = fig.add_subplot(111) plt.set_cmap('twilight') c = ax.pcolor(X, Y, head_dir_preference.T, vmin=-np.pi, vmax=np.pi) fig.colorbar(c, ax=ax, label="Tuning") # plt.colorbar() if axonal_clouds is not None: for i, p in enumerate(axonal_clouds): ell = p.get_ellipse() # print(ell) ax.add_artist(ell) # ax.quiver(p.x, p.y, delta_xy[i][0], delta_xy[i][1], scale=2.0) # px = axonal_clouds[0].x # py = axonal_clouds[0].y # # print(px,py) # ax.plot(px, py, 'rx') # p1 = axonal_clouds[0].get_p1() # # print(p1) # ax.plot(p1[0], p1[1], 'bx') # p2 = axonal_clouds[0].get_p2() # # print(p2) # ax.plot(p2[0], p2[1], 'bx') def get_position_mesh(ex_positions): n_ex = int(np.sqrt(len(ex_positions))) X = np.array([x for x, _ in ex_positions]).reshape((n_ex, n_ex)) Y = np.array([y for _, y in ex_positions]).reshape((n_ex, n_ex)) return X, Y def get_entropy_of_axonal_coverage(ex_neuron_tuning, ie_connections): ex_phase_vals_per_pickle = [] phase_entropy_per_pickle = [] n_hist_bins = 10 for connected_idxs in ie_connections: ex_phase_vals_of_p = [ex_neuron_tuning[idx] for idx in connected_idxs] phase_entropy = 0.0 ex_phase_vals_per_pickle.append(ex_phase_vals_of_p) ex_phase_hist, _ = np.histogram(ex_phase_vals_of_p, n_hist_bins) n_ex_of_p = float(len(ex_phase_vals_of_p)) ex_phase_hist = 1. / n_ex_of_p * ex_phase_hist.astype(float) for n_phi in ex_phase_hist: if n_phi != 0: phase_entropy -= n_phi * np.log(n_phi) phase_entropy_per_pickle.append(phase_entropy) return phase_entropy_per_pickle def get_excitatory_neurons_in_inhibitory_axonal_clouds(ex_neuron_positions, inhibitory_axonal_clouds): ie_connections = [] for p in inhibitory_axonal_clouds: ie_conn_p = [] x_p_1, y_p_1 = p.get_p1() x_p_2, y_p_2 = p.get_p2() for ex_idx, ex_pos in enumerate(ex_neuron_positions): x_ex, y_ex = ex_pos d_1 = np.sqrt((x_ex - x_p_1) ** 2 + (y_ex - y_p_1) ** 2) d_2 = np.sqrt((x_ex - x_p_2) ** 2 + (y_ex - y_p_2) ** 2) if d_1 + d_2 < p.c: ie_conn_p.append(ex_idx) ie_connections.append(ie_conn_p) return ie_connections def get_excitatory_phases_in_inhibitory_axon(ex_neuron_positions, ex_neuron_tunings, inhibitory_neuron): ex_phase_vals = [] x_p_1, y_p_1 = inhibitory_neuron.get_p1() x_p_2, y_p_2 = inhibitory_neuron.get_p2() for ex_idx, ex_pos in enumerate(ex_neuron_positions): x_ex, y_ex = ex_pos d_1 = np.sqrt((x_ex - x_p_1) ** 2 + (y_ex - y_p_1) ** 2) d_2 = np.sqrt((x_ex - x_p_2) ** 2 + (y_ex - y_p_2) ** 2) if d_1 + d_2 < inhibitory_neuron.c: ex_phase_vals.append(ex_neuron_tunings[ex_idx]) return ex_phase_vals def get_interneuron_entropy(ex_phase_vals, entropy_bins): phase_entropy = 0. ex_phase_hist, _ = np.histogram(ex_phase_vals,entropy_bins) ex_phase_prob = ex_phase_hist/len(ex_phase_vals) for prob in ex_phase_prob: if prob > 0.0: phase_entropy += -prob * np.log(prob) return phase_entropy def create_interneuron_sheet_by_noise(ex_positions, ex_tunings, number_of_interneurons, long_axis, short_axis, max_x, max_y, random_seed=None, n_iterations = 100): ''' Start with uniform distribution ''' if random_seed is not None: rng.seed(random_seed) pickle_a = long_axis pickle_b = short_axis pickle_n = number_of_interneurons x_n = int(np.sqrt(pickle_n)) y_n = int(np.sqrt(pickle_n)) x_step = int(max_x/(x_n+1)) y_step = int(max_y/(y_n+1)) inhibitory_axonal_clouds = [] for n in range(pickle_n): inhibitory_axonal_clouds.append(Pickle(((n % x_n) + 1) * x_step, \ (int(n / y_n) + 1) * y_step, \ pickle_a, pickle_b, rng.random() * 2. * np.pi)) ''' Determine initial tunings and entropy values ''' entropy_bins = np.linspace(-np.pi, np.pi, 30) ex_tunings_of_interneurons = [] entropies_of_interneurons = [] for i_neuron in inhibitory_axonal_clouds: ex_phase_vals = get_excitatory_phases_in_inhibitory_axon(ex_positions, ex_tunings, i_neuron) ex_tunings_of_interneurons.append(ex_phase_vals) entropies_of_interneurons.append(get_interneuron_entropy(ex_phase_vals, entropy_bins)) ''' (Randomly) perturb interneuron and see if entropy increases, if so update position and axonal tunings ''' # print('Mean Entropy before rotation ', np.mean(entropies_of_interneurons)) pert_phi = np.pi / 4.0 pert_phi_decrease = pert_phi / (n_iterations + 1) for it in tqdm(range(n_iterations), desc="Optimizing axonal cloud distribution"): rotation_perturbation = pert_phi - it * pert_phi_decrease for i_neuron_id, i_neuron in enumerate(inhibitory_axonal_clouds): perturbed_interneuron = Pickle(i_neuron.x, i_neuron.y, long_axis, short_axis, i_neuron.phi + rotation_perturbation) perturbed_phase_vals = get_excitatory_phases_in_inhibitory_axon(ex_positions, ex_tunings, perturbed_interneuron) perturbed_entropy = get_interneuron_entropy(perturbed_phase_vals, entropy_bins) if perturbed_entropy > entropies_of_interneurons[i_neuron_id]: ex_tunings_of_interneurons.append(ex_phase_vals) entropies_of_interneurons[i_neuron_id] = perturbed_entropy i_neuron.phi = perturbed_interneuron.phi else: perturbed_interneuron = Pickle(i_neuron.x, i_neuron.y, long_axis, short_axis, i_neuron.phi - rotation_perturbation) perturbed_phase_vals = get_excitatory_phases_in_inhibitory_axon(ex_positions, ex_tunings, perturbed_interneuron) perturbed_entropy = get_interneuron_entropy(perturbed_phase_vals, entropy_bins) if perturbed_entropy > entropies_of_interneurons[i_neuron_id]: ex_tunings_of_interneurons.append(ex_phase_vals) entropies_of_interneurons[i_neuron_id] = perturbed_entropy i_neuron.phi = perturbed_interneuron.phi # print('Mean Entropy after rotation', np.mean(entropies_of_interneurons)) return inhibitory_axonal_clouds, np.mean(entropies_of_interneurons) def create_interneuron_sheet_entropy_max_orientation(ex_positions, ex_tunings, number_of_interneurons, long_axis, short_axis, max_x, max_y, trial_orientations = 30): ''' Start with uniform distribution ''' pickle_a = long_axis pickle_b = short_axis pickle_n = number_of_interneurons x_n = int(np.sqrt(pickle_n)) y_n = int(np.sqrt(pickle_n)) x_step = int(max_x/(x_n+1)) y_step = int(max_y/(y_n+1)) inhibitory_axonal_clouds = [] for n in range(pickle_n): inhibitory_axonal_clouds.append(Pickle(((n % x_n) + 1) * x_step, \ (int(n / y_n) + 1) * y_step, \ pickle_a, pickle_b, rng.random() * 2. * np.pi)) ''' Determine initial tunings and entropy values ''' entropy_bins = np.linspace(-np.pi, np.pi, 30) entropies_of_interneurons = [] # for i_neuron in tqdm(inhibitory_axonal_clouds, desc="Maximizing entropy by rotation"): for i_neuron in inhibitory_axonal_clouds: orientations = np.linspace(-np.pi, np.pi, trial_orientations) entropy_for_orientation = [] for orientation in orientations: perturbed_interneuron = Pickle(i_neuron.x, i_neuron.y, long_axis, short_axis, orientation) ex_phase_vals = get_excitatory_phases_in_inhibitory_axon(ex_positions, ex_tunings, perturbed_interneuron) entropy_for_orientation.append(get_interneuron_entropy(ex_phase_vals, entropy_bins)) optimal_orientation = orientations[np.argmax(entropy_for_orientation)] i_neuron.phi = optimal_orientation entropies_of_interneurons.append(np.max(entropy_for_orientation)) # print(len(ex_phase_vals),short_axis,long_axis) return inhibitory_axonal_clouds, np.mean(entropies_of_interneurons) def create_interneuron_sheet_by_repulsive_force(number_of_interneurons, long_axis, short_axis, max_x, max_y, random_seed=None, n_iterations = 100): # ''' # Start with random pickle distribution # ''' # if random_seed is not None: # rng.seed(random_seed) # pickle_a = long_axis # pickle_b = short_axis # pickle_n = number_of_interneurons # x_rng_min = pickle_a # x_rng_max = max_x - pickle_a # y_rng_min = pickle_a # y_rng_max = max_y - pickle_a # inhibitory_axonal_clouds = [] # for n in range(pickle_n): # inhibitory_axonal_clouds.append(Pickle(rng.uniform(x_rng_min, x_rng_max), \ # rng.uniform(y_rng_min, y_rng_max), \ # pickle_a, pickle_b, rng.random() * 2. * np.pi)) ''' Start with uniform distribution ''' if random_seed is not None: rng.seed(random_seed) pickle_a = long_axis pickle_b = short_axis pickle_n = number_of_interneurons x_n = int(np.sqrt(pickle_n)) y_n = int(np.sqrt(pickle_n)) x_step = int(max_x/(x_n+1)) y_step = int(max_y/(y_n+1)) inhibitory_axonal_clouds = [] for n in range(pickle_n): inhibitory_axonal_clouds.append(AxonalCloud(((n % x_n) + 1) * x_step, \ (int(n / y_n) + 1) * y_step, \ pickle_a, pickle_b, rng.random() * 2. * np.pi)) # plotting() ''' Distribute pickles by repulsive "force" ''' f_scale = 0.2 def point_repulsion(x_i, y_i, x_j, y_j): d_x = x_i - x_j d_y = y_i - y_j # Distance is torus-like as to approach more homogenous distribution if d_x > max_x / 2.: d_x = -max_x + d_x elif d_x < -max_x / 2.: d_x = max_x + d_x if d_y > max_y / 2.: d_y = -max_y + d_y elif d_y < -max_y / 2.: d_y = max_y + d_y d = np.sqrt(d_x ** 2 + d_y ** 2) f_i = f_scale * 1. / (d + 1) ** 2 f_x = f_i * d_x / d f_y = f_i * d_y / d return f_x, f_y def pickle_repulsion(p_i, p_j): phi = p_i.phi p_i_x1, p_i_y1 = p_i.get_p1() p_i_x2, p_i_y2 = p_i.get_p2() p_j_x1, p_j_y1 = p_j.get_p1() p_j_x2, p_j_y2 = p_j.get_p2() f_p1p1_x, f_p1p1_y = point_repulsion(p_i_x1, p_i_y1, p_j_x1, p_j_y1) # p1-p1 f_p1p2_x, f_p1p2_y = point_repulsion(p_i_x1, p_i_y1, p_j_x2, p_j_y2) # p1-p2 f_p2p1_x, f_p2p1_y = point_repulsion(p_i_x2, p_i_y2, p_j_x1, p_j_y1) # p2-p1 f_p2p2_x, f_p2p2_y = point_repulsion(p_i_x2, p_i_y2, p_j_x2, p_j_y2) # p2-p2 f_p1_x = f_p1p1_x + f_p1p2_x f_p1_y = f_p1p1_y + f_p1p2_y f_p2_x = f_p2p1_x + f_p2p2_x f_p2_y = f_p2p1_y + f_p2p2_y m_p1 = (f_p1_x * np.sin(phi) - f_p1_y * np.cos(phi)) m_p2 = (-f_p2_x * np.sin(phi) + f_p2_y * np.cos(phi)) f_x_sum = f_p1_x + f_p2_x f_y_sum = f_p1_y + f_p2_y m_sum = m_p1 + m_p2 return f_x_sum, f_y_sum, m_sum for i in tqdm(range(n_iterations), desc="Optimizing axonal cloud distribution"): delta_xyphi = [] for i, p_i in enumerate(inhibitory_axonal_clouds): delta_x = 0. delta_y = 0. delta_phi = 0. for p_j in inhibitory_axonal_clouds: if p_i is not p_j: del_x, del_y, del_m = pickle_repulsion(p_i, p_j) delta_x += del_x delta_y += del_y delta_phi += del_m delta_xyphi.append([delta_x, delta_y, delta_phi]) # plotting(delta_xy) for i, p_i in enumerate(inhibitory_axonal_clouds): x_new = p_i.x + delta_xyphi[i][0] y_new = p_i.y + delta_xyphi[i][1] p_i.phi = p_i.phi + delta_xyphi[i][2] # Torus-like distance so pickles can cross borders and behave like a subset of a bigger distribution if x_new < max_x and x_new > 0.0: p_i.x = x_new elif x_new < 0.: p_i.x = max_x + x_new elif x_new > max_x: p_i.x = x_new - max_x if y_new < max_y and y_new > 0.0: p_i.y = y_new elif y_new < 0.: p_i.y = max_y + y_new elif y_new > max_y: p_i.y = y_new - max_y # plotting() # plotting() return inhibitory_axonal_clouds def create_grid_of_excitatory_neurons(max_x, max_y, number_of_excitatory_neurons_per_row, tuning_map): grid_x = number_of_excitatory_neurons_per_row grid_y = number_of_excitatory_neurons_per_row x = np.linspace(0, max_x, grid_x) y = np.linspace(0, max_y, grid_y) ex_neuron_positions = [] ex_neuron_tuning = [] head_dir_preference = np.zeros((grid_x, grid_y)) for y_idx, yy in enumerate(y): for x_idx, xx in enumerate(x): p_noise = tuning_map(xx, yy) head_dir_preference[x_idx, y_idx] = p_noise ex_neuron_positions.append([xx, yy]) ex_neuron_tuning.append(p_noise) return ex_neuron_positions, ex_neuron_tuning def test_run(): ''' Neural sheet ''' # To go from the 3d data obtained in Peng 2017 to a 2d sheet is not straightforward # Assuming a depth of 100um, the density becomes a 1000 interneurons/mm2 and thus about 3000-4000 exneurons/mm2 # A long 200um short 50um interneuron axon would cover about 90 partners # On the other hand, the typical contacts per excitatory neuron would be A_total/(A_I*N_I) = 30 N_E = 400 N_I = 40 sheet_x = 300 * um sheet_y = 300 * um inhibitory_axon_long_axis = 100 * um inhibitory_axon_short_axis = 25 * um number_of_excitatory_neurons_per_row = int(np.sqrt(N_E)) ''' Tuning Maps ''' # tuning_label = "Perlin" tuning_label = "Orientation map" if tuning_label == "Perlin": tuning_map = lambda x, y: noise.pnoise2(x / 100.0, y / 100.0, octaves=2) * np.pi elif tuning_label == "Orientation map": map = OrientationMap(20, 20, 7, sheet_x / um, sheet_y / um) map.improve(10) tuning_map = lambda x, y: map.orientation_map_float(x, y) ex_positions, ex_tunings = create_grid_of_excitatory_neurons(sheet_x / um, sheet_y / um, number_of_excitatory_neurons_per_row, tuning_map) # inhibitory_axonal_clouds = create_interneuron_sheet(N_I, inhibitory_axon_long_axis / um, # inhibitory_axon_short_axis / um, sheet_x / um, # sheet_y / um, random_seed=2, n_iterations=1000) inhibitory_axonal_clouds = create_interneuron_sheet_by_noise(ex_positions, ex_tunings, N_I, inhibitory_axon_long_axis / um, inhibitory_axon_short_axis / um, sheet_x / um, sheet_y / um, random_seed=3, n_iterations=10) plot_neural_sheet(ex_positions, ex_tunings, inhibitory_axonal_clouds) if __name__ == "__main__": test_run()