ELEC 483 Final Project

import numpy as np import random as rand # load data with np.load('/datasets/hw3_dataset/ReachData.npz', allow_pickle=True) as data: r = data['r'] cfr = data['cfr'] print(data.files) data = np.load('/datasets/hw3_dataset/ReachData.npz', allow_pickle=True) # Description of data in 'r' # - r[i].unit[j].spikeTimes - array of spiketimes relative to trial start (in ms) for trial i, neuron j # - r[i].timeTouchHeld - the time the reach target appears and movement planning nominally begins for trial i # - r[i].timeGoCue - the time the animal is cued to move and planning nominally ends # - r[i].timeTargetAcquire - the time movement ends # 'pp' - delay period duration # RT - unknown but length of 1127 # NumberInClass - use this to balance across # There are a total of 1127 trials and 190 neurons

# Initialize spike rate matrices spike_rates_undiff = np.zeros((1127, 190)) # Matrix with dimensions (# of trials) x (number of neurons), each row is spike rates vector (in spikes/ms) spike_rates_plan_only = np.zeros((1127, 190)) # Matrix with dimensions (# of trials) x (number of neurons), each row is spike rates vector (in spikes/ms) spike_rates_move_only = np.zeros((1127, 190)) # Matrix with dimensions (# of trials) x (number of neurons), each row is spike rates vector (in spikes/ms) for i in range(1127): for j in range(190): spiketimes = np.array([r[i].unit[j].spikeTimes]) if isinstance(r[i].unit[j].spikeTimes, float) else r[i].unit[j].spikeTimes # spike times for current trial/neuron spike_rates_undiff[i][j] = len(spiketimes[(spiketimes > r[i].timeTouchHeld) & (spiketimes <= r[i].timeTargetAcquire)]) # count of spikes in plan + move phases spike_rates_undiff[i][j] = 1000*spike_rates_undiff[i][j] / (r[i].timeTargetAcquire - r[i].timeTouchHeld) spike_rates_plan_only[i][j] = len(spiketimes[(spiketimes > r[i].timeTouchHeld) & (spiketimes <= r[i].timeGoCue)]) # count of spikes in plan phase spike_rates_plan_only[i][j] = 1000*spike_rates_plan_only[i][j] / (r[i].timeGoCue - r[i].timeTouchHeld) spike_rates_move_only[i][j] = len(spiketimes[(spiketimes > r[i].timeGoCue) & (spiketimes <= r[i].timeTargetAcquire)]) # count of spikes in move phase spike_rates_move_only[i][j] = 1000*spike_rates_move_only[i][j] / (r[i].timeTargetAcquire - r[i].timeGoCue) spike_rates_plan_move = np.concatenate((spike_rates_plan_only, spike_rates_move_only), axis=1)

from math import pi, sqrt, exp def gauss_kernel(length,sigma=15): ''' Gaussian kernel function, centered at 0, with given length and standard deviation ''' r = range(-int(length/2),int(length/2)+1) return [1 / (sigma * sqrt(2*pi)) * exp(-float(x)**2/(2*sigma**2)) for x in r] # Help with generating spike rate over time - https://rcweb.dartmouth.edu/~mvdm/wiki/doku.php?id=analysis:nsb2016:week9

from sklearn.decomposition import PCA from scipy.ndimage import gaussian_filter1d from sklearn.preprocessing import StandardScaler def pc_directions(bin_size, kernel_length, kernel_sigma): ''' Computes leading principal component direction for each unit Returns: dictionaries mapping neuron -> 1st principal component (for plan and move) ''' plan_pc_directions = {} # dictionary mapping neuron number -> 1st principal component move_pc_directions = {} # dictionary mapping neuron number -> 1st principal component for i in range(190): avg_spike_trains_plan = [] # to store average plan spike train across trials, length = number of targets (8) avg_spike_trains_move = [] # to store average move spike train across trials, length = number of targets (8) for target in range(1, 9): relevant_trials = np.squeeze(np.argwhere(cfr == target)) # indices of trials for neuron i which have given target spiketimes = [np.array([r[trial].unit[i].spikeTimes]) if isinstance(r[trial].unit[i].spikeTimes, float) else r[trial].unit[i].spikeTimes for trial in relevant_trials] # array containing spiketimes for each trial for this neuron plan_start = [r[trial].timeTouchHeld for trial in relevant_trials] plan_end = [r[trial].timeGoCue for trial in relevant_trials] move_end = [r[trial].timeTargetAcquire for trial in relevant_trials] all_plan_conv = [] all_move_conv = [] for spike_train, start, middle, end in zip(spiketimes, plan_start, plan_end, move_end): # bin the spikes binned_plan, bin_edges_plan = np.histogram(spike_train, bins=bin_size, range=(start, middle)) binned_move, bin_edges_move = np.histogram(spike_train, bins=bin_size, range=(middle, end)) bin_size_plan = bin_edges_plan[1] - bin_edges_plan[0] # bin width for the plan data bin_size_move = bin_edges_move[1] - bin_edges_move[0] # bin width for the move data # convolve with gaussian plan_convolved = np.convolve(binned_plan, gauss_kernel(kernel_length, kernel_sigma), mode='same') / bin_size_plan # normalize by bin size (gives rates instead of counts) move_convolved = np.convolve(binned_move, gauss_kernel(kernel_length, kernel_sigma), mode='same') / bin_size_move # normalize by bin size (gives rates instead of counts) all_plan_conv.append(plan_convolved) all_move_conv.append(move_convolved) avg_spike_trains_plan.append(np.mean(all_plan_conv, axis=0)) avg_spike_trains_move.append(np.mean(all_move_conv, axis=0)) # PCA on average_spike_trains & save first PC pca = PCA() sc = StandardScaler() avg_spike_trains_plan = sc.fit_transform(avg_spike_trains_plan) pca.fit(avg_spike_trains_plan) plan_pc_directions[i] = pca.components_[0] pca = PCA() sc = StandardScaler() avg_spike_trains_move = sc.fit_transform(avg_spike_trains_move) pca.fit(avg_spike_trains_move) move_pc_directions[i] = pca.components_[0] return move_pc_directions, plan_pc_directions

def get_pc_scores(bin_size, kernel_length, kernel_sigma, move_pc_directions, plan_pc_directions): ''' Given PC directions as found above, compute the spike rate PC scores needed for downstream modeling tasks ''' spike_rates_move_pc = np.zeros((1127, 190)) spike_rates_plan_pc = np.zeros((1127, 190)) for i in range(1127): for j in range(190): spiketimes = np.array([r[i].unit[j].spikeTimes]) if isinstance(r[i].unit[j].spikeTimes, float) else r[i].unit[j].spikeTimes # spike train for this neuron/trial start = r[i].timeTouchHeld middle = r[i].timeGoCue end = r[i].timeTargetAcquire # bin the spikes binned_plan, bin_edges_plan = np.histogram(spiketimes, bins=bin_size, range=(start, middle)) binned_move, bin_edges_move = np.histogram(spiketimes, bins=bin_size, range=(middle, end)) bin_size_plan = bin_edges_plan[1] - bin_edges_plan[0] # bin width for the plan data bin_size_move = bin_edges_move[1] - bin_edges_move[0] # bin width for the move data # convolve with gaussian plan_convolved = np.convolve(binned_plan, gauss_kernel(kernel_length, kernel_sigma), mode='same') / bin_size_plan # normalize by bin size move_convolved = np.convolve(binned_move, gauss_kernel(kernel_length, kernel_sigma), mode='same') / bin_size_move # normalize by bin size spike_rates_move_pc[i][j] = np.inner(move_convolved, move_pc_directions[j]) spike_rates_plan_pc[i][j] = np.inner(plan_convolved, plan_pc_directions[j]) return spike_rates_move_pc, spike_rates_plan_pc

move_pc_directions, plan_pc_directions = pc_directions(500, 50,10) spike_rates_move_pc, spike_rates_plan_pc = get_pc_scores(500, 50, 10, move_pc_directions, plan_pc_directions)

spike_rates_plan_move_pc = np.concatenate((spike_rates_plan_only, spike_rates_move_pc), axis=1) spike_rates_plan_pc_move_pc = np.concatenate((spike_rates_plan_pc, spike_rates_move_pc), axis=1) # print(np.count_nonzero(spike_rates_move_pc==0))

import matplotlib.pyplot as plt plt.plot(plan_pc_directions[0]) plt.xlabel("Time (ms)") plt.ylabel("Gaussian Smoothed Rate (spikes/ms)") plt.title("Example Principal Component for Neuron 1 - Plan") plt.show()

def learn_gaussian(training_data, labels, correction): ''' Inputs: training data, class labels for all training trials (as in cfr) Outputs: dict mapping class (1-8) to a vector of gaussian parameters (tuple of (mean, covariance)). Assumes diagonal covariance! ''' gaussian_params = {} # dict mapping target number -> (mean, covariance) for k in range(1, 9): rows = training_data[labels == k] # Get training data corresponding to given indices, for current class # Compute max-likelihood parameter estimates mean = np.mean(rows, axis=0) cov = np.diag(np.cov(rows.T, bias=True)) # Accounting for 0 variance values cov.setflags(write=1) cov[cov == 0] = correction # correcting for 0 variance. gaussian_params[k] = (mean, cov) return gaussian_params

from scipy.stats import multivariate_normal from scipy.stats import norm import math # def gaussian_llh(data, params, penalty): # ''' # NOT USED # Inputs: data, tuple containing mean and variances of Gaussian # Outputs: log-likelihood of data under the given Gaussian model # ''' # means = params[0] # variances = params[1] # llh = 0 # for i in range(len(data)): # if norm.pdf(data[i], means[i], variances[i]) != 0: # llh += np.log(norm.pdf(data[i], means[i], variances[i])) # else: # llh += penalty # return llh def predict_target(test_data, gaussian_params): ''' Inputs: test_data, and dictionary containing the gaussian params for each target Outputs: predicted targets ''' predicted_targets = [] # for row in test_data: # predicted_targets.append(max(range(1, 9), key=lambda target: gaussian_llh(row, gaussian_params[target], penalty))) # choose target with max log-likelihood # print("Prediction: " + str(predicted_targets[-1])) for row in test_data: max_target, probs = target_prob(row, gaussian_params) predicted_targets.append(max_target) # print("Prediction: " + str(predicted_targets[-1])) return predicted_targets

from functools import reduce import math def target_prob(data, gaussian_params): ''' Inputs: data from a single trial, gaussian params for each target Outputs: max likelihood target estimate, array of p(target | data) for targets 1-8 ''' probs = [] # probs of target given data log_probs = [] # log probs of data given each target for i in range(1, 9): mean = gaussian_params[i][0] variance = gaussian_params[i][1] # print(np.log(norm.pdf(data, mean, variance))) temp_logs = np.log(norm.pdf(data, mean, variance)) temp_logs[temp_logs == -np.inf] = -1000 # temp = sum(temp_logs) log_probs.append(sum(temp_logs)) # log_probs.append(temp if temp != -np.inf else -1000) # print(log_probs) max_target = 1 + np.argmax(log_probs) # target with highest log(p(data|target)) # print(max_target) for i in range(1, 9): log_numerator = log_probs[i-1] log_denominator = max_target + np.log(np.sum(np.exp(log_probs - max_target))) # print(log_denominator) probs.append(np.exp(log_numerator - log_denominator)) # print(probs) return max_target, probs def predict_target_AP(test_data, gaussian_params): ''' Uses mean of the a posteriori distribution as the target estimator Inputs: test_data, and dictionary containing the gaussian params for each target Outputs: predicted targets ''' predicted_targets = [] for row in test_data: max_target, probs = target_prob(row, gaussian_params) prediction = np.inner(np.arange(1,9), probs) predicted_targets.append(prediction if prediction != np.inf else max_target) # print("Prediction: " + str(predicted_targets[-1])) return predicted_targets

def train_test_split(training_size): ''' Inputs: desired number of trials per target in the training set Outputs: train indices and test indices, balanced by each target ''' trials_by_target = {} # dict mapping each target (1 to 8) to the indices of trials for that target for target in range(1, 9): trials_by_target[target] = list(np.squeeze(np.argwhere(cfr == target))) train_indices = [] test_indices = [] for target in range(1, 9): train = rand.sample(trials_by_target[target], training_size) # train indices for this target test = np.setdiff1d(trials_by_target[target], train) # test indices for this target train_indices.extend(train) test_indices.extend(test) return train_indices, test_indices train_indices, test_indices = train_test_split(75)

def angular_error(predicted, actual): """ Takes predicted and actual values and computes the angular error Assumptions: Assumes that each target is 45 degrees apart from the others Outputs: Average angular error (as defined by Yu et al.) """ distances = abs(predicted - actual) return np.mean(45*np.minimum(distances, 8 - distances))

import itertools def tuning(bin_sizes, kernel_sigmas, corrections, penalties): """ Tuning the different parameters, saving best parameters out / print """ search_results = {} param_grid = [bin_sizes, kernel_sigmas, corrections, penalties] options = list(itertools.product(*param_grid)) for params in options: # print(params) move_pc_directions, plan_pc_directions = pc_directions(params[0], 50,params[1]) spike_rates_move_pc, spike_rates_plan_pc = get_pc_scores(params[0], 50, params[1], move_pc_directions, plan_pc_directions) gaussian_params = learn_gaussian(spike_rates_plan_pc_move_pc[train_indices], cfr[train_indices], params[2]) predicted = predict_target(spike_rates_plan_pc_move_pc[test_indices], gaussian_params, params[3]) err = angular_error(predicted, cfr[test_indices]) search_results[params] = err print("param_grid: ", params, "error: ", err) return search_results

# potential_bins = [500,750,1000,1250] # potential_sigmas = [10,15,20] # potential_corrections = [1e-2, 1e-5] # potential_penalties = [-1000,-50000,-100000] # # 500,10,0.01, -1000 - 2.73 # results = tuning(potential_bins, potential_sigmas, potential_corrections, potential_penalties)

gaussian_params = learn_gaussian(spike_rates_undiff[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_undiff[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

gaussian_params = learn_gaussian(spike_rates_move_only[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_move_only[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

gaussian_params = learn_gaussian(spike_rates_plan_only[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_only[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

gaussian_params = learn_gaussian(spike_rates_plan_move[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

gaussian_params = learn_gaussian(spike_rates_plan_move_pc[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move_pc[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

gaussian_params = learn_gaussian(spike_rates_plan_pc_move_pc[train_indices], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_pc_move_pc[test_indices], gaussian_params) print(angular_error(predicted, cfr[test_indices]))

import matplotlib.pyplot as plt

# num_trials = 10 # num_units = 40 # number of neurons to use (out of 190) # training_size = 90 # number of trials in training set # train_indices, test_indices = train_test_split(training_size) # unit_indices = np.random.choice(190, num_units, replace=False) # spike_rates_plan_move_pc[np.ix_(train_indices, unit_indices)].shape

from scipy.stats import sem num_trials = 1000 num_units = 140 # number of neurons to use (out of 190) training_size = 65 # number of trials in training set move_error = [] plan_error = [] undiff_error = [] plan_move_error = [] plan_move_pc_error = [] planpc_movepc_error = [] for i in range(num_trials): train_indices, test_indices = train_test_split(training_size) unit_indices = np.random.choice(190, num_units, replace=False) # unit_indices2 = np.random.choice(380, 190, replace=False) unit_indices2 = np.concatenate((unit_indices, unit_indices + 190)) # Move Only params_move_only = learn_gaussian(spike_rates_move_only[np.ix_(train_indices, unit_indices)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_move_only[np.ix_(test_indices, unit_indices)], params_move_only) move_error.append(angular_error(predicted, cfr[test_indices])) # Plan Only params_plan_only = learn_gaussian(spike_rates_plan_only[np.ix_(train_indices, unit_indices)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_only[np.ix_(test_indices, unit_indices)], params_plan_only) plan_error.append(angular_error(predicted, cfr[test_indices])) # Undifferentiated params_undiff = learn_gaussian(spike_rates_undiff[np.ix_(train_indices, unit_indices)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_undiff[np.ix_(test_indices, unit_indices)], params_undiff) undiff_error.append(angular_error(predicted, cfr[test_indices])) # Plan and Move params_plan_move = learn_gaussian(spike_rates_plan_move[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move[np.ix_(test_indices, unit_indices2)], params_plan_move) plan_move_error.append(angular_error(predicted, cfr[test_indices])) # Plan + Move PC gaussian_params = learn_gaussian(spike_rates_plan_move_pc[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move_pc[np.ix_(test_indices, unit_indices2)], gaussian_params) plan_move_pc_error.append(angular_error(predicted, cfr[test_indices])) # Plan PC Move PC params_planpc_movepc = learn_gaussian(spike_rates_plan_pc_move_pc[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_pc_move_pc[np.ix_(test_indices, unit_indices2)], params_planpc_movepc) planpc_movepc_error.append(angular_error(predicted, cfr[test_indices])) # print(angular_error(predicted, cfr[test_indices])) # gaussian_params = learn_gaussian(spike_rates_plan_pc_move_pc[train_indices], cfr[train_indices], 1e-5) # predicted = predict_target(spike_rates_plan_pc_move_pc[test_indices], gaussian_params) # print(angular_error(predicted, cfr[test_indices])) move_error_mean = np.mean(move_error) plan_error_mean = np.mean(plan_error) undiff_error_mean = np.mean(undiff_error) plan_move_error_mean = np.mean(plan_move_error) plan_move_pc_error_mean = np.mean(plan_move_pc_error) planpc_movepc_error_mean = np.mean(planpc_movepc_error) move_error_std = np.std(move_error) plan_error_std = np.std(plan_error) undiff_error_std = np.std(undiff_error) plan_move_error_std = np.std(plan_move_error) plan_move_pc_error_std = np.std(plan_move_pc_error) planpc_movepc_error_std = np.std(planpc_movepc_error) categories = ['Move Rate', 'Plan Rate', 'Undiff. Rate', 'Plan Rate/Move Rate', 'Plan Rate/Move PC', 'Plan PC/Move PC'] x_pos = np.arange(len(categories)) errors = [move_error_mean, plan_error_mean, undiff_error_mean, plan_move_error_mean, plan_move_pc_error_mean, planpc_movepc_error_mean] stds = [move_error_std, plan_error_std, undiff_error_std, plan_move_error_std, plan_move_pc_error_std, planpc_movepc_error_std] # Build the plot fig, ax = plt.subplots() ax.bar(x_pos, errors, yerr=stds, align='center', alpha=0.6, ecolor='black', capsize=10) ax.set_ylabel('Angular Error (in degrees)') ax.set_xticks(x_pos) ax.set_xticklabels(categories) ax.set_title('Comparison of cross-validated decode performance') ax.yaxis.grid(True) plt.setp(ax.get_xticklabels(), rotation=30, horizontalalignment='right') # Save the figure and show plt.tight_layout() plt.savefig('Fig4_bar_plot.png') plt.show() print(move_error_mean, plan_error_mean, undiff_error_mean, plan_move_error_mean, plan_move_pc_error_mean, planpc_movepc_error_mean) print(move_error_std, plan_error_std, undiff_error_std, plan_move_error_std, plan_move_pc_error_std, planpc_movepc_error_std)

move_error_mean

num_trials = 100 units = range(10, 210, 20) # number of neurons to use (out of 190) training_size = 65 # number of trials in training set undiff_means = [] undiff_stds = [] plan_move_means = [] plan_move_stds = [] plan_move_pc_means = [] plan_move_pc_stds = [] for num_units in units: undiff_error = [] plan_move_error = [] plan_move_pc_error = [] for i in range(num_trials): train_indices, test_indices = train_test_split(training_size) unit_indices = np.random.choice(190, num_units, replace=False) unit_indices2 = np.concatenate((unit_indices, unit_indices + 190)) # Undifferentiated params_undiff = learn_gaussian(spike_rates_undiff[np.ix_(train_indices, unit_indices)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_undiff[np.ix_(test_indices, unit_indices)], params_undiff) undiff_error.append(angular_error(predicted, cfr[test_indices])) # Plan and Move params_plan_move = learn_gaussian(spike_rates_plan_move[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move[np.ix_(test_indices, unit_indices2)], params_plan_move) plan_move_error.append(angular_error(predicted, cfr[test_indices])) # Plan + Move PC gaussian_params = learn_gaussian(spike_rates_plan_move_pc[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move_pc[np.ix_(test_indices, unit_indices2)], gaussian_params) plan_move_pc_error.append(angular_error(predicted, cfr[test_indices])) undiff_means.append(np.mean(undiff_error)) plan_move_means.append(np.mean(plan_move_error)) plan_move_pc_means.append(np.mean(plan_move_pc_error)) undiff_stds.append(sem(undiff_error)) plan_move_stds.append(sem(plan_move_error)) plan_move_pc_stds.append(sem(plan_move_pc_error)) # print(plan_move_means) plt.errorbar(units, undiff_means, yerr=undiff_stds, color='blue') plt.errorbar(units, plan_move_means, yerr=plan_move_stds, color='red') plt.errorbar(units, plan_move_pc_means, yerr=plan_move_pc_stds, color='green') plt.legend(["Undiff. Rate", "Plan Rate/Move Rate", "Plan Rate/Move PC"]) plt.xlabel("Number of Units") plt.ylabel("Angular Error (degrees)") plt.title("Angular Error vs. Unit Count") plt.savefig('Fig5a_err_vs_units.png') plt.show()

num_trials = 100 training_sizes = [5,10, 20, 30, 40, 50, 60, 70] # number of neurons to use (out of 190) num_units = 140 # number of trials in training set undiff_means = [] undiff_stds = [] plan_move_means = [] plan_move_stds = [] plan_move_pc_means = [] plan_move_pc_stds = [] for training_size in training_sizes: undiff_error = [] plan_move_error = [] plan_move_pc_error = [] for i in range(num_trials): train_indices, test_indices = train_test_split(training_size) unit_indices = np.random.choice(190, num_units, replace=False) unit_indices2 = np.concatenate((unit_indices, unit_indices + 190)) # Undifferentiated params_undiff = learn_gaussian(spike_rates_undiff[np.ix_(train_indices, unit_indices)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_undiff[np.ix_(test_indices, unit_indices)], params_undiff) undiff_error.append(angular_error(predicted, cfr[test_indices])) # Plan and Move params_plan_move = learn_gaussian(spike_rates_plan_move[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move[np.ix_(test_indices, unit_indices2)], params_plan_move) plan_move_error.append(angular_error(predicted, cfr[test_indices])) # Plan + Move PC gaussian_params = learn_gaussian(spike_rates_plan_move_pc[np.ix_(train_indices, unit_indices2)], cfr[train_indices], 1e-5) predicted = predict_target(spike_rates_plan_move_pc[np.ix_(test_indices, unit_indices2)], gaussian_params) plan_move_pc_error.append(angular_error(predicted, cfr[test_indices])) undiff_means.append(np.mean(undiff_error)) plan_move_means.append(np.mean(plan_move_error)) plan_move_pc_means.append(np.mean(plan_move_pc_error)) undiff_stds.append(sem(undiff_error)) plan_move_stds.append(sem(plan_move_error)) plan_move_pc_stds.append(sem(plan_move_pc_error)) # print(plan_move_means) plt.errorbar(training_sizes, undiff_means, yerr=undiff_stds, color='blue') plt.errorbar(training_sizes, plan_move_means, yerr=plan_move_stds, color='red') plt.errorbar(training_sizes, plan_move_pc_means, yerr=plan_move_pc_stds, color='green') plt.legend(["Undiff. Rate", "Plan Rate/Move Rate", "Plan Rate/Move PC"]) plt.xlabel("Training Size") plt.ylabel("Angular Error (degrees)") plt.ylim(0, 50) plt.title("Angular Error vs. Training Size") plt.savefig('Fig5b_err_vs_train.png') plt.show()