Open In Colab

Bonus Tutorial: Diving Deeper into Decoding & Encoding

Week 2, Day 1: Deep Learning

By Neuromatch Academy

Content creators: Jorge A. Menendez, Carsen Stringer

Content reviewers: Roozbeh Farhoodi, Madineh Sarvestani, Kshitij Dwivedi, Spiros Chavlis, Ella Batty, Michael Waskom

Our 2021 Sponsors, including Presenting Sponsor Facebook Reality Labs


Tutorial Objectives

In this tutorial, we’ll will dive deeper into our decoding model from Tutorial 1 and we will fit a convolutional neural network directly to neural activities.


Setup

# Imports

import os
import numpy as np

import torch
from torch import nn
from torch import optim

import matplotlib as mpl
from matplotlib import pyplot as plt

Data retrieval and loading

#@title Data retrieval and loading
import hashlib
import requests

fname = "W3D4_stringer_oribinned1.npz"
url = "https://osf.io/683xc/download"
expected_md5 = "436599dfd8ebe6019f066c38aed20580"

if not os.path.isfile(fname):
  try:
    r = requests.get(url)
  except requests.ConnectionError:
    print("!!! Failed to download data !!!")
  else:
    if r.status_code != requests.codes.ok:
      print("!!! Failed to download data !!!")
    elif hashlib.md5(r.content).hexdigest() != expected_md5:
      print("!!! Data download appears corrupted !!!")
    else:
      with open(fname, "wb") as fid:
        fid.write(r.content)

Figure Settings

#@title Figure Settings
%config InlineBackend.figure_format = 'retina'
plt.style.use("https://raw.githubusercontent.com/NeuromatchAcademy/course-content/master/nma.mplstyle")

Plotting Functions

#@title Plotting Functions

def plot_data_matrix(X, ax):
  """Visualize data matrix of neural responses using a heatmap

  Args:
    X (torch.Tensor or np.ndarray): matrix of neural responses to visualize
        with a heatmap
    ax (matplotlib axes): where to plot

  """

  cax = ax.imshow(X, cmap=mpl.cm.pink, vmin=np.percentile(X, 1), vmax=np.percentile(X, 99))
  cbar = plt.colorbar(cax, ax=ax, label='normalized neural response')

  ax.set_aspect('auto')
  ax.set_xticks([])
  ax.set_yticks([])

def plot_decoded_results(train_loss, test_loss, test_labels, predicted_test_labels):
  """ Plot decoding results in the form of network training loss and test predictions

  Args:
    train_loss (list): training error over iterations
    test_labels (torch.Tensor): n_test x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data, in radians
    predicted_test_labels (torch.Tensor): n_test x 1 tensor with predicted orientations of the
      stimuli from decoding neural network

  """

  # Plot results
  fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 6))

  # Plot the training loss over iterations of GD
  ax1.plot(train_loss)
  # Plot the testing loss over iterations of GD
  ax1.plot(test_loss)
  ax1.legend(['train loss', 'test loss'])

  # Plot true stimulus orientation vs. predicted class
  ax2.plot(stimuli_test.squeeze(), predicted_test_labels, '.')

  ax1.set_xlim([0, None])
  ax1.set_ylim([0, None])
  ax1.set_xlabel('iterations of gradient descent')
  ax1.set_ylabel('negative log likelihood')
  ax2.set_xlabel('true stimulus orientation ($^o$)')
  ax2.set_ylabel('decoded orientation bin')
  ax2.set_xticks(np.linspace(0, 360, n_classes + 1))
  ax2.set_yticks(np.arange(n_classes))
  class_bins = [f'{i * 360 / n_classes: .0f}$^o$ - {(i + 1) * 360 / n_classes: .0f}$^o$' for i in range(n_classes)]
  ax2.set_yticklabels(class_bins);

  # Draw bin edges as vertical lines
  ax2.set_ylim(ax2.get_ylim())  # fix y-axis limits
  for i in range(n_classes):
    lower = i * 360 / n_classes
    upper = (i + 1) * 360 / n_classes
    ax2.plot([lower, lower], ax2.get_ylim(), '-', color="0.7", linewidth=1, zorder=-1)
    ax2.plot([upper, upper], ax2.get_ylim(), '-', color="0.7", linewidth=1, zorder=-1)

  plt.tight_layout()

def visualize_weights(W_in_sorted, W_out):
    plt.figure(figsize=(10,4))
    plt.subplot(1,2,1)
    plt.imshow(W_in_sorted, aspect='auto', cmap='bwr', vmin=-1e-2, vmax=1e-2)
    plt.colorbar()
    plt.xlabel('sorted neurons')
    plt.ylabel('hidden units')
    plt.title('$W_{in}$')

    plt.subplot(1,2,2)
    plt.imshow(W_out.T, cmap='bwr', vmin=-3, vmax=3)
    plt.xticks([])
    plt.xlabel('output')
    plt.ylabel('hidden units')
    plt.colorbar()
    plt.title('$W_{out}$')

def visualize_hidden_units(W_in_sorted, h):
    plt.figure(figsize=(10,8))
    plt.subplot(2,2,1)
    plt.imshow(W_in_sorted, aspect='auto', cmap='bwr', vmin=-1e-2, vmax=1e-2)
    plt.xlabel('sorted neurons')
    plt.ylabel('hidden units')
    plt.colorbar()
    plt.title('$W_{in}$')

    plt.subplot(2,2,2)
    plt.imshow(h, aspect='auto')
    plt.xlabel('stimulus orientation ($^\circ$)')
    plt.ylabel('hidden units')
    plt.colorbar()
    plt.title('$\mathbf{h}$')

    plt.subplot(2,2,4)
    plt.plot(h.T)
    plt.xlabel('stimulus orientation ($^\circ$)')
    plt.ylabel('hidden unit activity')
    plt.title('$\mathbf{h}$ tuning curves')

    plt.show()

def plot_weights(weights, channels=[0], colorbar=True):
  """ plot convolutional channel weights
  Args:
        weights: weights of convolutional filters (conv_channels x K x K)
        channels: which conv channels to plot
  """
  wmax = torch.abs(weights).max()
  fig, axs = plt.subplots(1,len(channels), figsize=(12,2.5))
  for i, channel in enumerate(channels):
    im = axs[i].imshow(weights[channel,0], vmin=-wmax, vmax=wmax, cmap='bwr')
    axs[i].set_title('channel %d'%channel)

  if colorbar:
    ax = fig.add_axes([1, 0.1, 0.05, 0.8])
    plt.colorbar(im, ax=ax)
    ax.axis('off')

def plot_tuning(ax, stimuli, respi_train, respi_test, neuron_index, linewidth=2):
  """Plot the tuning curve of a neuron"""

  ax.plot(stimuli, respi_train, 'y', linewidth=linewidth)  # plot its responses as a function of stimulus orientation
  ax.plot(stimuli, respi_test, 'm', linewidth=linewidth)  # plot its responses as a function of stimulus orientation
  ax.set_title('neuron %i' % neuron_index)
  ax.set_xlabel('stimulus orientation ($^o$)')
  ax.set_ylabel('neural response')
  ax.set_xticks(np.linspace(0, 360, 5))
  ax.set_ylim([-0.5, 2.4])


def plot_prediction(ax, y_pred, y_train, y_test):
  """ plot prediction of neural response + test neural response """
  ax.plot(y_train, 'y', linewidth=1)
  ax.plot(y_test,color='m')
  ax.plot(y_pred, 'g', linewidth=3)
  ax.set_xlabel('stimulus bin')
  ax.set_ylabel('response')


def plot_training_curves(train_loss, test_loss):
  """
  Args:
    train_loss (list): training error over iterations
    test_loss (list): n_test x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data, in radians
    predicted_test_labels (torch.Tensor): n_test x 1 tensor with predicted orientations of the
      stimuli from decoding neural network

  """

  f, ax = plt.subplots()
  # Plot the training loss over iterations of GD
  ax.plot(train_loss)
  # Plot the testing loss over iterations of GD
  ax.plot(test_loss, '.', markersize=10)
  ax.legend(['train loss', 'test loss'])
  ax.set(xlabel="Gradient descent iteration", ylabel="Mean squared error")

Helper Functions

# @title Helper Functions

def load_data(data_name=fname, bin_width=1):
  """Load mouse V1 data from Stringer et al. (2019)

  Data from study reported in this preprint:
  https://www.biorxiv.org/content/10.1101/679324v2.abstract

  These data comprise time-averaged responses of ~20,000 neurons
  to ~4,000 stimulus gratings of different orientations, recorded
  through Calcium imaging. The responses have been normalized by
  spontaneous levels of activity and then z-scored over stimuli, so
  expect negative numbers. They have also been binned and averaged
  to each degree of orientation.

  This function returns the relevant data (neural responses and
  stimulus orientations) in a torch.Tensor of data type torch.float32
  in order to match the default data type for nn.Parameters in
  Google Colab.

  This function will actually average responses to stimuli with orientations
  falling within bins specified by the bin_width argument. This helps
  produce individual neural "responses" with smoother and more
  interpretable tuning curves.

  Args:
    bin_width (float): size of stimulus bins over which to average neural
      responses

  Returns:
    resp (torch.Tensor): n_stimuli x n_neurons matrix of neural responses,
        each row contains the responses of each neuron to a given stimulus.
        As mentioned above, neural "response" is actually an average over
        responses to stimuli with similar angles falling within specified bins.
    stimuli: (torch.Tensor): n_stimuli x 1 column vector with orientation
        of each stimulus, in degrees. This is actually the mean orientation
        of all stimuli in each bin.

  """
  with np.load(data_name) as dobj:
    data = dict(**dobj)
  resp = data['resp']
  stimuli = data['stimuli']

  if bin_width > 1:
    # Bin neural responses and stimuli
    bins = np.digitize(stimuli, np.arange(0, 360 + bin_width, bin_width))
    stimuli_binned = np.array([stimuli[bins == i].mean() for i in np.unique(bins)])
    resp_binned = np.array([resp[bins == i, :].mean(0) for i in np.unique(bins)])
  else:
    resp_binned = resp
    stimuli_binned = stimuli

  # Return as torch.Tensor
  resp_tensor = torch.tensor(resp_binned, dtype=torch.float32)
  stimuli_tensor = torch.tensor(stimuli_binned, dtype=torch.float32).unsqueeze(1)  # add singleton dimension to make a column vector

  return resp_tensor, stimuli_tensor


def identityLine():
  """
  Plot the identity line y=x
  """
  ax = plt.gca()
  lims = np.array([ax.get_xlim(), ax.get_ylim()])
  minval = lims[:, 0].min()
  maxval = lims[:, 1].max()
  equal_lims = [minval, maxval]
  ax.set_xlim(equal_lims)
  ax.set_ylim(equal_lims)
  line = ax.plot([minval, maxval], [minval, maxval], color="0.7")
  line[0].set_zorder(-1)

def get_data(n_stim, train_data, train_labels):
  """ Return n_stim randomly drawn stimuli/resp pairs

  Args:
    n_stim (scalar): number of stimuli to draw
    resp (torch.Tensor):
    train_data (torch.Tensor): n_train x n_neurons tensor with neural
      responses to train on
    train_labels (torch.Tensor): n_train x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data, in radians

  Returns:
    (torch.Tensor, torch.Tensor): n_stim x n_neurons tensor of neural responses and n_stim x 1 of orientations respectively
  """
  n_stimuli = train_labels.shape[0]
  istim = np.random.choice(n_stimuli, n_stim)
  r = train_data[istim]  # neural responses to this stimulus
  ori = train_labels[istim]  # true stimulus orientation

  return r, ori

def stimulus_class(ori, n_classes):
  """Get stimulus class from stimulus orientation

  Args:
    ori (torch.Tensor): orientations of stimuli to return classes for
    n_classes (int): total number of classes

  Returns:
    torch.Tensor: 1D tensor with the classes for each stimulus

  """
  bins = np.linspace(0, 360, n_classes + 1)
  return torch.tensor(np.digitize(ori.squeeze(), bins)) - 1  # minus 1 to accomodate Python indexing

def grating(angle, sf=1 / 28, res=0.1, patch=False):
  """Generate oriented grating stimulus

  Args:
    angle (float): orientation of grating (angle from vertical), in degrees
    sf (float): controls spatial frequency of the grating
    res (float): resolution of image. Smaller values will make the image
      smaller in terms of pixels. res=1.0 corresponds to 640 x 480 pixels.
    patch (boolean): set to True to make the grating a localized
      patch on the left side of the image. If False, then the
      grating occupies the full image.

  Returns:
    torch.Tensor: (res * 480) x (res * 640) pixel oriented grating image

  """

  angle = np.deg2rad(angle)  # transform to radians

  wpix, hpix = 640, 480  # width and height of image in pixels for res=1.0

  xx, yy = np.meshgrid(sf * np.arange(0, wpix * res) / res, sf * np.arange(0, hpix * res) / res)

  if patch:
    gratings = np.cos(xx * np.cos(angle + .1) + yy * np.sin(angle + .1))  # phase shift to make it better fit within patch
    gratings[gratings < 0] = 0
    gratings[gratings > 0] = 1
    xcent = gratings.shape[1] * .75
    ycent = gratings.shape[0] / 2
    xxc, yyc = np.meshgrid(np.arange(0, gratings.shape[1]), np.arange(0, gratings.shape[0]))
    icirc = ((xxc - xcent) ** 2 + (yyc - ycent) ** 2) ** 0.5 < wpix / 3 / 2 * res
    gratings[~icirc] = 0.5

  else:
    gratings = np.cos(xx * np.cos(angle) + yy * np.sin(angle))
    gratings[gratings < 0] = 0
    gratings[gratings > 0] = 1

  gratings -= 0.5

  # Return torch tensor
  return torch.tensor(gratings, dtype=torch.float32)

def filters(out_channels=6, K=7):
  """ make example filters, some center-surround and gabors
  Returns:
      filters: out_channels x K x K
  """
  grid = np.linspace(-K/2, K/2, K).astype(np.float32)
  xx,yy = np.meshgrid(grid, grid, indexing='ij')

  # create center-surround filters
  sigma = 1.1
  gaussian = np.exp(-(xx**2 + yy**2)**0.5/(2*sigma**2))
  wide_gaussian = np.exp(-(xx**2 + yy**2)**0.5/(2*(sigma*2)**2))
  center_surround = gaussian - 0.5 * wide_gaussian

  # create gabor filters
  thetas = np.linspace(0, 180, out_channels-2+1)[:-1] * np.pi/180
  gabors = np.zeros((len(thetas), K, K), np.float32)
  lam = 10
  phi = np.pi/2
  gaussian = np.exp(-(xx**2 + yy**2)**0.5/(2*(sigma*0.4)**2))
  for i,theta in enumerate(thetas):
    x = xx*np.cos(theta) + yy*np.sin(theta)
    gabors[i] = gaussian * np.cos(2*np.pi*x/lam + phi)

  filters = np.concatenate((center_surround[np.newaxis,:,:],
                            -1*center_surround[np.newaxis,:,:],
                            gabors),
                           axis=0)
  filters /= np.abs(filters).max(axis=(1,2))[:,np.newaxis,np.newaxis]
  filters -= filters.mean(axis=(1,2))[:,np.newaxis,np.newaxis]
  # convert to torch
  filters = torch.from_numpy(filters)
  # add channel axis
  filters = filters.unsqueeze(1)

  return filters

def regularized_MSE_loss(output, target, weights=None, L2_penalty=0, L1_penalty=0):
  """loss function for MSE

  Args:
    output (torch.Tensor): output of network
    target (torch.Tensor): neural response network is trying to predict
    weights (torch.Tensor): fully-connected layer weights of network (net.out_layer.weight)
    L2_penalty : scaling factor of sum of squared weights
    L1_penalty : scalaing factor for sum of absolute weights

  Returns:
    (torch.Tensor) mean-squared error with L1 and L2 penalties added

  """

  loss_fn = nn.MSELoss()
  loss = loss_fn(output, target)

  if weights is not None:
    L2 = L2_penalty * torch.square(weights).sum()
    L1 = L1_penalty * torch.abs(weights).sum()
    loss += L1 + L2

  return loss

Section 1: Decoding - Investigating model and evaluating performance

In this section, we will return to our decoding model from Tutorial 1 and further investigate its performance, and then improve it in the next section. Let’s first load the data again and train our model, as we did in Tutorial 1.

Execute this cell to load and visualize data

#@title

#@markdown Execute this cell to load and visualize data

# Load data
resp_all, stimuli_all = load_data()  # argument to this function specifies bin width
n_stimuli, n_neurons = resp_all.shape

print(f'{n_neurons} neurons in response to {n_stimuli} stimuli')

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(2 * 6, 5))

# Visualize data matrix
plot_data_matrix(resp_all[:100, :].T, ax1)  # plot responses of first 100 neurons
ax1.set_xlabel('stimulus')
ax1.set_ylabel('neuron')

# Plot tuning curves of three random neurons
ineurons = np.random.choice(n_neurons, 3, replace=False)  # pick three random neurons
ax2.plot(stimuli_all, resp_all[:, ineurons])
ax2.set_xlabel('stimulus orientation ($^o$)')
ax2.set_ylabel('neural response')
ax2.set_xticks(np.linspace(0, 360, 5))

plt.tight_layout()
23589 neurons in response to 360 stimuli
../../../_images/W2D1_Tutorial4_17_1.png

Execute this cell to split into training and test sets

#@title
#@markdown Execute this cell to split into training and test sets

# Set random seeds for reproducibility
np.random.seed(4)
torch.manual_seed(4)

# Split data into training set and testing set
n_train = int(0.6 * n_stimuli)  # use 60% of all data for training set
ishuffle = torch.randperm(n_stimuli)
itrain = ishuffle[:n_train]  # indices of data samples to include in training set
itest = ishuffle[n_train:]  # indices of data samples to include in testing set
stimuli_test = stimuli_all[itest]
resp_test = resp_all[itest]
stimuli_train = stimuli_all[itrain]
resp_train = resp_all[itrain]

Execute this cell to train the network

# @markdown Execute this cell to train the network

class DeepNetReLU(nn.Module):
  """ network with a single hidden layer h with a RELU """

  def __init__(self, n_inputs, n_hidden):
    super().__init__()  # needed to invoke the properties of the parent class nn.Module
    self.in_layer = nn.Linear(n_inputs, n_hidden) # neural activity --> hidden units
    self.out_layer = nn.Linear(n_hidden, 1) # hidden units --> output

  def forward(self, r):

    h = torch.relu(self.in_layer(r)) # h is size (n_inputs, n_hidden)
    y = self.out_layer(h) # y is size (n_inputs, 1)

    return y


def train(net, loss_fn, train_data, train_labels,
          n_epochs=50, learning_rate=1e-4):
  """Run gradient descent to opimize parameters of a given network

  Args:
    net (nn.Module): PyTorch network whose parameters to optimize
    loss_fn: built-in PyTorch loss function to minimize
    train_data (torch.Tensor): n_train x n_neurons tensor with neural
      responses to train on
    train_labels (torch.Tensor): n_train x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data
    n_epochs (int, optional): number of epochs of gradient descent to run
    learning_rate (float, optional): learning rate to use for gradient descent

  Returns:
    (list): training loss over iterations

  """

  # Initialize PyTorch SGD optimizer
  optimizer = optim.SGD(net.parameters(), lr=learning_rate)

  # Placeholder to save the loss at each iteration
  train_loss = []

  # Loop over epochs
  for i in range(n_epochs):

    # compute network output from inputs in train_data
    out = net(train_data)  # compute network output from inputs in train_data

    # evaluate loss function
    loss = loss_fn(out, train_labels)

    # Clear previous gradients
    optimizer.zero_grad()

    # Compute gradients
    loss.backward()

    # Update weights
    optimizer.step()

    # Store current value of loss
    train_loss.append(loss.item())  # .item() needed to transform the tensor output of loss_fn to a scalar


    # Track progress
    if (i + 1) % (n_epochs // 5) == 0:
      print(f'iteration {i + 1}/{n_epochs} | loss: {loss.item():.3f}')

  return train_loss


# Set random seeds for reproducibility
np.random.seed(1)
torch.manual_seed(1)

# Initialize network with 10 hidden units
net = DeepNetReLU(n_neurons, 10)

# Initialize built-in PyTorch MSE loss function
loss_fn = nn.MSELoss()

# Run gradient descent on data
train_loss = train(net, loss_fn, resp_train, stimuli_train)

# Plot the training loss over iterations of GD
plt.plot(train_loss)
plt.xlim([0, None])
plt.ylim([0, None])
plt.xlabel('iterations of gradient descent')
plt.ylabel('mean squared error')
plt.show()
iteration 10/50 | loss: 12219.761
iteration 20/50 | loss: 1672.732
iteration 30/50 | loss: 548.096
iteration 40/50 | loss: 235.619
iteration 50/50 | loss: 144.001
../../../_images/W2D1_Tutorial4_22_2.png

Section 1.1: Peering inside the decoding model

We have built a model to perform decoding that takes as input neural activity and outputs the estimated angle of the stimulus. We can imagine that an animal that needs to determine angles would have a brain area that acts like the hidden layer in our model. It transforms the neural activity from visual cortex and outputs a decision. Decisions about orientations of edges could include figuring out how to jump onto a branch, how to avoid obstacles, or determining the type of an object, e.g. food or predator.

What sort of connectivity would this brain area have with visual cortex? Determining this experimentally would be very difficult, perhaps we can look at the model we have and see if its structure constrains the type of connectivity we’d expect.

Below we will visualize the weights from the neurons in visual cortex to the hidden units \(\mathbf{W}_{in}\), and the weights from the hidden units to the output orientation \(\mathbf{W}_{out}\).

PyTorch Note:

An important thing to note in the code below is the .detach() method. The PyTorch nn.Module class is special in that, behind the scenes, each of the variables inside it are linked to each other in a computational graph, for the purposes of automatic differentiation (the algorithm used in .backward() to compute gradients). As a result, if you want to do anything that is not a torch operation to the parameters or outputs of an nn.Module class, you’ll need to first “detach” it from its computational graph. This is what the .detach() method does. In this code below, we need to call it on the weights of the network. We also convert the variable from a pytorch tensor to a numpy array using the .numpy() method.

W_in = net.in_layer.weight.detach().numpy() # we can run .detach() and .numpy() to get a numpy array
print('shape of W_in:')
print(W_in.shape)

W_out = net.out_layer.weight.detach().numpy() # we can run .detach() and .numpy() to get a numpy array
print('shape of W_out:')
print(W_out.shape)

plt.figure(figsize=(10,4))
plt.subplot(1,2,1)
plt.imshow(W_in, aspect='auto', cmap='bwr', vmin=-1e-2, vmax=1e-2)
plt.xlabel('neurons')
plt.ylabel('hidden units')
plt.colorbar()
plt.title('$W_{in}$')

plt.subplot(1,2,2)
plt.imshow(W_out.T, cmap='bwr', vmin=-3, vmax=3)
plt.xticks([])
plt.xlabel('output')
plt.ylabel('hidden units')
plt.colorbar()
plt.title('$W_{out}$')

plt.show()
shape of W_in:
(10, 23589)
shape of W_out:
(1, 10)
../../../_images/W2D1_Tutorial4_25_1.png

Coding Exercise 1.1: Visualizing weights

It’s difficult to see any structure in this weight matrix. How might we visualize it in a better way?

Perhaps we can sort the neurons by their preferred orientation. We will use the resp_all matrix which is 360 stimuli (360\(^\circ\) of angles) by number of neurons. How do we find the preferred orientation?

Let’s visualize one column of this resp_all matrix first as we did at the beginning of the notebook. Can you see how we might want to first process this tuning curve before choosing the preferred orientation?

idx = 235
plt.plot(resp_all[:,idx])
plt.ylabel('neural response')
plt.xlabel('stimulus orientation ($^\circ$)')
plt.title(f'neuron {idx}')
plt.show()
../../../_images/W2D1_Tutorial4_27_0.png

Looking at this tuning curve, there is a bit of noise across orientations, so let’s smooth with a gaussian filter and then find the position of the maximum for each neuron. After getting the maximum position aka the “preferred orientation” for each neuron, we will re-sort the \(\mathbf{W}_{in}\) matrix. The maximum position in a matrix can be computed using the .argmax(axis=_) function in python – make sure you specify the right axis though! Next, to get the indices of a matrix sorted we will need to use the .argsort() function.

from scipy.ndimage import gaussian_filter1d

# first let's smooth the tuning curves resp_all to make sure we get
# an accurate peak that isn't just noise
# resp_all is size (n_stimuli, n_neurons)
resp_smoothed = gaussian_filter1d(resp_all, 5, axis=0)
# resp_smoothed is size (n_stimuli, n_neurons)

############################################################################
## TO DO for students
# Fill out function and remove
raise NotImplementedError("Student exercise: find preferred orientation")
############################################################################

## find position of max response for each neuron
## aka preferred orientation for each neuron
preferred_orientation = ...

## Resort W_in matrix by preferred orientation
isort = preferred_orientation.argsort()
W_in_sorted = W_in[:,isort]

# plot resorted W_in matrix
visualize_weights(W_in_sorted, W_out)

Click for solution

Example output:

Solution hint

We can plot the activity of hidden units across various stimuli to better understand the hidden units. Recall the formula for the hidden units

(153)\[\begin{equation} \mathbf{h}^{(n)} = \phi(\mathbf{W}^{in} \mathbf{r}^{(n)} + \mathbf{b}^{in}) \end{equation}\]

We can compute the activity \(\mathbf{h}^{(n)}\) directly using \(\mathbf{W}^{in}\) and \(\mathbf{b}^{in}\) or we can modify our network above to return h in the .forward() method. In this case, we’ll compute using the equation, but in practice the second method is recommended.

W_in = net.in_layer.weight.detach() # size (10, n_neurons)
b_in = net.in_layer.bias.detach().unsqueeze(1) # size (10, 1)

# Compute hidden unit activity h
h = torch.relu(W_in @ resp_all.T + b_in)
h = h.detach().numpy() # we can run .detach() and .numpy() to get a numpy array

# Visualize
visualize_hidden_units(W_in_sorted, h)

Think! 1.1: Interpreting weights

We have just visualized how the model transforms neural activity to hidden layer activity. How should we interpret these matrices? Here are some guiding questions to explore:

  • Why are some of the \(\mathbf{W}_{in}\) weights close to zero for some of the hidden units? Do these correspond to close to zero weights in \(\mathbf{W}_{out}\)?

  • Note how each hidden unit seems to have strongest weights to two groups of neurons in \(\mathbf{W}_{in}\), corresponding to two different sets of preferred orientations. Why do you think that is? What does might this tell us about the structure of the tuning curves of the neurons?

  • It appears that there is at least one hidden unit active at each orientation, which is necessary to decode across all orientations. What would happen if some orientations did not activate any hidden units?

Click for solution

Section 1.2: Generalization performance with test data

Note that gradient descent is essentially an algorithm for fitting the network’s parameters to a given set of training data. Selecting this training data is thus crucial for ensuring that the optimized parameters generalize to unseen data they weren’t trained on. In our case, for example, we want to make sure that our trained network is good at decoding stimulus orientations from neural responses to any orientation, not just those in our data set.

To ensure this, we have split up the full data set into a training set and a testing set. In Coding Exercise 3.2, we trained a deep network by optimizing the parameters on a training set. We will now evaluate how good the optimized parameters are by using the trained network to decode stimulus orientations from neural responses in the testing set. Good decoding performance on this testing set should then be indicative of good decoding performance on the neurons’ responses to any other stimulus orientation. This procedure is commonly used in machine learning (not just in deep learning)and is typically referred to as cross-validation.

We will compute the MSE on the test data and plot the decoded stimulus orientations as a function of the true stimulus.

Execute this cell to evaluate and plot test error

#@title
#@markdown Execute this cell to evaluate and plot test error

out = net(resp_test)  # decode stimulus orientation for neural responses in testing set
ori = stimuli_test  # true stimulus orientations
test_loss = loss_fn(out, ori)  # MSE on testing set (Hint: use loss_fn initialized in previous exercise)

plt.plot(ori, out.detach(), '.')  # N.B. need to use .detach() to pass network output into plt.plot()
identityLine()  # draw the identity line y=x; deviations from this indicate bad decoding!
plt.title('MSE on testing set: %.2f' % test_loss.item())  # N.B. need to use .item() to turn test_loss into a scalar
plt.xlabel('true stimulus orientation ($^o$)')
plt.ylabel('decoded stimulus orientation ($^o$)')
axticks = np.linspace(0, 360, 5)
plt.xticks(axticks)
plt.yticks(axticks)
plt.show()
../../../_images/W2D1_Tutorial4_38_0.png

If interested, please see the next section to think more about model criticism, improve the loss function accordingly, and add regularization.

Section 2: Decoding - Evaluating & improving models


Section 2.1: Model criticism

Let’s now take a step back and think about how our model is succeeding/failing and how to improve it.

Execute this cell to plot decoding error

#@title
#@markdown Execute this cell to plot decoding error

out = net(resp_test)  # decode stimulus orientation for neural responses in testing set
ori = stimuli_test  # true stimulus orientations
error = out - ori  # decoding error


plt.plot(ori, error.detach(), '.')   # plot decoding error as a function of true orientation (make sure all arguments to plt.plot() have been detached from PyTorch network!)

# Plotting
plt.xlabel('true stimulus orientation ($^o$)')
plt.ylabel('decoding error ($^o$)')
plt.xticks(np.linspace(0, 360, 5))
plt.yticks(np.linspace(-360, 360, 9))
plt.show()
../../../_images/W2D1_Tutorial4_44_0.png

Think! 2.1: Delving into error problems

In the cell below, we will plot the decoding error for each neural response in the testing set. The decoding error is defined as the decoded stimulus orientation minus true stimulus orientation

(154)\[\begin{equation} \text{decoding error} = y^{(n)} - \tilde{y}^{(n)} \end{equation}\]

In particular, we plot decoding error as a function of the true stimulus orientation.

  • Are some stimulus orientations harder to decode than others?

  • If so, in what sense? Are the decoded orientations for these stimuli more variable and/or are they biased?

  • Can you explain this variability/bias? What makes these stimulus orientations different from the others?

  • (Will be addressed in next exercise) Can you think of a way to modify the deep network in order to avoid this?

Click for solution

Section 2.2: Improving the loss function

As illustrated in the previous exercise, the squared error is not a good loss function for circular quantities like angles, since two angles that are very close (e.g. \(1^o\) and \(359^o\)) might actually have a very large squared error.

Here, we’ll avoid this problem by changing our loss function to treat our decoding problem as a classification problem. Rather than estimating the exact angle of the stimulus, we’ll now aim to construct a decoder that classifies the stimulus into one of \(C\) classes, corresponding to different bins of angles of width \(b = \frac{360}{C}\). The true class \(\tilde{y}^{(n)}\) of stimulus \(i\) is now given by

(155)\[\begin{equation} \tilde{y}^{(n)} = \begin{cases} 1 &\text{if angle of stimulus $n$ is in the range } [0, b] \\ 2 &\text{if angle of stimulus $n$ is in the range } [b, 2b] \\ 3 &\text{if angle of stimulus $n$ is in the range } [2b, 3b] \\ \vdots \\ C &\text{if angle of stimulus $n$ is in the range } [(C-1)b, 360] \end{cases} \end{equation}\]

We have a helper function stimulus_class that will extract n_classes stimulus classes for us from the stimulus orientations.

To decode the stimulus class from neural responses, we’ll use a deep network that outputs a \(C\)-dimensional vector of probabilities \(\mathbf{p} = \begin{bmatrix} p_1, p_2, \ldots, p_C \end{bmatrix}^T\), corresponding to the estimated probabilities of the stimulus belonging to each class \(1, 2, \ldots, C\).

To ensure the network’s outputs are indeed probabilities (i.e. they are positive numbers between 0 and 1, and sum to 1), we’ll use a softmax function to transform the real-valued outputs from the hidden layer into probabilities. Letting \(\sigma(\cdot)\) denote this softmax function, the equations describing our network are

(156)\[\begin{align} \mathbf{h}^{(n)} &= \phi(\mathbf{W}^{in} \mathbf{r}^{(n)} + \mathbf{b}^{in}), && [\mathbf{W}^{in}: M \times N], \\ \mathbf{p}^{(n)} &= \sigma(\mathbf{W}^{out} \mathbf{h}^{(n)} + \mathbf{b}^{out}), && [\mathbf{W}^{out}: C \times M], \end{align}\]

The decoded stimulus class is then given by that assigned the highest probability by the network:

(157)\[\begin{equation} y^{(n)} = \underset{i}{\arg\max} \,\, p_i \end{equation}\]

The softmax function can be implemented in PyTorch simply using torch.softmax().

Often log probabilities are easier to work with than actual probabilities, because probabilities tend to be very small numbers that computers have trouble representing. We’ll therefore actually use the logarithm of the softmax as the output of our network,

(158)\[\begin{equation} \mathbf{l}^{(n)} = \log \left( \mathbf{p}^{(n)} \right) \end{equation}\]

which can implemented in PyTorch together with the softmax via an nn.LogSoftmax layer. The nice thing about the logarithmic function is that it’s monotonic, so if one probability is larger/smaller than another, then its logarithm is also larger/smaller than the other’s. We therefore have that

(159)\[\begin{equation} y^{(n)} = \underset{i}{\arg\max} \,\, p_i^{(n)} = \underset{i}{\arg\max} \, \log p_i^{(n)} = \underset{i}{\arg\max} \,\, l_i^{(n)} \end{equation}\]

See the next cell for code for constructing a deep network with one hidden layer that of ReLU’s that outputs a vector of log probabilities.

# Deep network for classification
class DeepNetSoftmax(nn.Module):
  """Deep Network with one hidden layer, for classification

  Args:
    n_inputs (int): number of input units
    n_hidden (int): number of units in hidden layer
    n_classes (int): number of outputs, i.e. number of classes to output
      probabilities for

  Attributes:
    in_layer (nn.Linear): weights and biases of input layer
    out_layer (nn.Linear): weights and biases of output layer

  """

  def __init__(self, n_inputs, n_hidden, n_classes):
    super().__init__()  # needed to invoke the properties of the parent class nn.Module
    self.in_layer = nn.Linear(n_inputs, n_hidden)  # neural activity --> hidden units
    self.out_layer = nn.Linear(n_hidden, n_classes)  # hidden units --> outputs
    self.logprob = nn.LogSoftmax(dim=1)  # probabilities across columns should sum to 1 (each output row corresponds to a different input)

  def forward(self, r):
    """Predict stimulus orientation bin from neural responses

    Args:
      r (torch.Tensor): n_stimuli x n_inputs tensor with neural responses to n_stimuli

    Returns:
      torch.Tensor: n_stimuli x n_classes tensor with predicted class probabilities

    """
    h = torch.relu(self.in_layer(r))
    logp = self.logprob(self.out_layer(h))
    return logp

What should our loss function now be? Ideally, we want the probabilities outputted by our network to be such that the probability of the true stimulus class is high. One way to formalize this is to say that we want to maximize the log probability of the true stimulus class \(\tilde{y}^{(n)}\) under the class probabilities predicted by the network,

(160)\[\begin{equation} \log \left( \text{predicted probability of stimulus } n \text{ being of class } \tilde{y}^{(n)} \right) = \log p^{(n)}_{\tilde{y}^{(n)}} = l^{(n)}_{\tilde{y}^{(n)}} \end{equation}\]

To turn this into a loss function to be minimized, we can then simply multiply it by -1: maximizing the log probability is the same as minimizing the negative log probability. Summing over a batch of \(P\) inputs, our loss function is then given by

(161)\[\begin{equation} L = -\sum_{n=1}^P \log p^{(n)}_{\tilde{y}^{(n)}} = -\sum_{n=1}^P l^{(n)}_{\tilde{y}^{(n)}} \end{equation}\]

In the deep learning community, this loss function is typically referred to as the cross-entropy, or negative log likelihood. The corresponding built-in loss function in PyTorch is nn.NLLLoss() (documentation here).

Coding Exercise 2.2: A new loss function

In the next cell, we’ve provided most of the code to train and test a network to decode stimulus orientations via classification, by minimizing the negative log likelihood. Fill in the missing pieces.

Once you’ve done this, have a look at the plotted results. Does changing the loss function from mean squared error to a classification loss solve our problems? Note that errors may still occur – but are these errors as bad as the ones that our network above was making?

Run this cell to create train function that uses test_data and L1 and L2 terms for next exercise

#@markdown Run this cell to create train function that uses test_data and L1 and L2 terms for next exercise
def train(net, loss_fn, train_data, train_labels,
          n_iter=50, learning_rate=1e-4,
          test_data=None, test_labels=None,
          L2_penalty=0, L1_penalty=0):
  """Run gradient descent to opimize parameters of a given network

  Args:
    net (nn.Module): PyTorch network whose parameters to optimize
    loss_fn: built-in PyTorch loss function to minimize
    train_data (torch.Tensor): n_train x n_neurons tensor with neural
      responses to train on
    train_labels (torch.Tensor): n_train x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data
    n_iter (int, optional): number of iterations of gradient descent to run
    learning_rate (float, optional): learning rate to use for gradient descent
    test_data (torch.Tensor, optional): n_test x n_neurons tensor with neural
      responses to test on
    test_labels (torch.Tensor, optional): n_test x 1 tensor with orientations of
      the stimuli corresponding to each row of test_data
    L2_penalty (float, optional): l2 penalty regularizer coefficient
    L1_penalty (float, optional): l1 penalty regularizer coefficient

  Returns:
    (list): training loss over iterations

  """

  # Initialize PyTorch SGD optimizer
  optimizer = optim.SGD(net.parameters(), lr=learning_rate)

  # Placeholder to save the loss at each iteration
  train_loss = []
  test_loss = []

  # Loop over epochs
  for i in range(n_iter):

    # compute network output from inputs in train_data
    out = net(train_data)  # compute network output from inputs in train_data

    # evaluate loss function
    if L2_penalty==0 and L1_penalty==0:
      # normal loss function
      loss = loss_fn(out, train_labels)
    else:
      # custom loss function from bonus exercise 3.3
      loss = loss_fn(out, train_labels, net.in_layer.weight,
                     L2_penalty, L1_penalty)

    # Clear previous gradients
    optimizer.zero_grad()
    # Compute gradients
    loss.backward()

    # Update weights
    optimizer.step()

    # Store current value of loss
    train_loss.append(loss.item())  # .item() needed to transform the tensor output of loss_fn to a scalar

    # Get loss for test_data, if given (we will use this in the bonus exercise 3.2 and 3.3)
    if test_data is not None:
      out_test = net(test_data)
      # evaluate loss function
      if L2_penalty==0 and L1_penalty==0:
        # normal loss function
        loss_test = loss_fn(out_test, test_labels)
      else:
        # (BONUS code) custom loss function from Bonus exercise 3.3
        loss_test = loss_fn(out_test, test_labels, net.in_layer.weight,
                            L2_penalty, L1_penalty)
      test_loss.append(loss_test.item())  # .item() needed to transform the tensor output of loss_fn to a scalar

    # Track progress
    if (i + 1) % (n_iter // 5) == 0:
      if test_data is None:
        print(f'iteration {i + 1}/{n_iter} | loss: {loss.item():.3f}')
      else:
        print(f'iteration {i + 1}/{n_iter} | loss: {loss.item():.3f} | test_loss: {loss_test.item():.3f}')

  if test_data is None:
    return train_loss
  else:
    return train_loss, test_loss
def decode_orientation(net, n_classes, loss_fn,
                       train_data, train_labels, test_data, test_labels,
                       n_iter=1000, L2_penalty=0, L1_penalty=0):
  """ Initialize, train, and test deep network to decode binned orientation from neural responses

  Args:
    net (nn.Module): deep network to run
    n_classes (scalar): number of classes in which to bin orientation
    loss_fn (function): loss function to run
    train_data (torch.Tensor): n_train x n_neurons tensor with neural
      responses to train on
    train_labels (torch.Tensor): n_train x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data, in radians
    test_data (torch.Tensor): n_test x n_neurons tensor with neural
      responses to train on
    test_labels (torch.Tensor): n_test x 1 tensor with orientations of the
      stimuli corresponding to each row of train_data, in radians
    n_iter (int, optional): number of iterations to run optimization
    L2_penalty (float, optional): l2 penalty regularizer coefficient
    L1_penalty (float, optional): l1 penalty regularizer coefficient

  Returns:
    (list, torch.Tensor): training loss over iterations, n_test x 1 tensor with predicted orientations of the
      stimuli from decoding neural network
  """

  # Bin stimulus orientations in training set
  train_binned_labels = stimulus_class(train_labels, n_classes)
  test_binned_labels = stimulus_class(test_labels, n_classes)


  # Run GD on training set data, using learning rate of 0.1
  # (add optional arguments test_data and test_binned_labels!)
  train_loss, test_loss = train(net, loss_fn, train_data, train_binned_labels,
                                learning_rate=0.1, test_data=test_data,
                                test_labels=test_binned_labels, n_iter=n_iter,
                                L2_penalty=L2_penalty, L1_penalty=L1_penalty)

  # Decode neural responses in testing set data
  out = net(test_data)
  out_labels = np.argmax(out.detach(), axis=1)  # predicted classes

  frac_correct = (out_labels==test_binned_labels).sum() / len(test_binned_labels)
  print(f'>>> fraction correct = {frac_correct:.3f}')

  return train_loss, test_loss, out_labels

# Set random seeds for reproducibility
np.random.seed(1)
torch.manual_seed(1)

n_classes = 20

############################################################################
## TO DO for students
# Fill out function and remove
raise NotImplementedError("Student exercise: make network and loss")
############################################################################

# Initialize network
net = ... # use M=20 hidden units

# Initialize built-in PyTorch negative log likelihood loss function
loss_fn = ...

# Train network and run it on test images
# this function uses the train function you wrote before
train_loss, test_loss, predicted_test_labels = decode_orientation(net, n_classes, loss_fn,
                                                                  resp_train, stimuli_train, resp_test, stimuli_test)

# Plot results
plot_decoded_results(train_loss, test_loss, stimuli_test, predicted_test_labels)

Click for solution

Example output:

Solution hint

How do the weights \(W_{in}\) from the neurons to the hidden layer look now?

W_in = net.in_layer.weight.detach().numpy() # we can run detach and numpy to get a numpy array
print('shape of W_in:')
print(W_in.shape)

W_out = net.out_layer.weight.detach().numpy()

# plot resorted W_in matrix
visualize_weights(W_in[:,isort], W_out)

Section 2.3: Regularization

Video 1: Regularization

As discussed in the lecture, it is often important to incorporate regularization terms into the loss function to avoid overfitting. In particular, in this case, we will use these terms to enforce sparsity in the linear layer from neurons to hidden units.

Here we’ll consider the classic L2 regularization penalty \(\mathcal{R}_{L2}\), which is the sum of squares of each weight in the network \(\sum_{ij} {\mathbf{W}^{out}_{ij}}^2\) times a constant that we call L2_penalty.

We will also add an L1 regularization penalty \(\mathcal{R}_{L1}\) to enforce sparsity of the weights, which is the sum of the absolute values of the weights \(\sum_{ij} |{\mathbf{W}^{out}_{ij}}|\) times a constant that we call L1_penalty.

We will add both of these to the loss function:

(162)\[\begin{equation} L = (y - \tilde{y})^2 + \mathcal{R}_{L2} + \mathcal{R}_{L1} \end{equation}\]

The parameters L2_penalty and L1_penalty are inputs to the train function.

Coding Exercise 2.3: Add regularization to training

We will create a new loss function that adds L1 and L2 regularization. In particular, you will:

  • add L2 loss penalty to the weights

  • add L1 loss penalty to the weights

We will then train the network using this loss function. Full training will take a few minutes: if you want to train for just a few steps to speed up the code while iterating on your code, you can decrease the n_iter input from 500.

Hint: since we are using torch instead of np, we will use torch.abs instead of np.absolute. You can use torch.sum or .sum() to sum over a tensor.

def regularized_loss(output, target, weights, L2_penalty=0, L1_penalty=0):
  """loss function with L2 and L1 regularization

  Args:
    output (torch.Tensor): output of network
    target (torch.Tensor): neural response network is trying to predict
    weights (torch.Tensor): linear layer weights from neurons to hidden units (net.in_layer.weight)
    L2_penalty : scaling factor of sum of squared weights
    L1_penalty : scalaing factor for sum of absolute weights

  Returns:
    (torch.Tensor) mean-squared error with L1 and L2 penalties added

  """

  ##############################################################################
  # TO DO: add L1 and L2 regularization to the loss function
  raise NotImplementedError("Student exercise: complete regularized_loss")
  ##############################################################################

  loss_fn = nn.NLLLoss()
  loss = loss_fn(output, target)

  L2 = L2_penalty * ...
  L1 = L1_penalty * ...
  loss += L1 + L2

  return loss

# Set random seeds for reproducibility
np.random.seed(1)
torch.manual_seed(1)

n_classes = 20

# Initialize network
net = DeepNetSoftmax(n_neurons, 20, n_classes)  # use M=20 hidden units

# Here you can play with L2_penalty > 0, L1_penalty > 0
train_loss, test_loss, predicted_test_labels = decode_orientation(net, n_classes,
                                                                  regularized_loss,
                                                                  resp_train, stimuli_train,
                                                                  resp_test, stimuli_test,
                                                                  n_iter=1000,
                                                                  L2_penalty=1e-2,
                                                                  L1_penalty=5e-4)

# Plot results
plot_decoded_results(train_loss, test_loss, stimuli_test, predicted_test_labels)

Click for solution

Example output:

Solution hint

It seems we were overfitting a little because we increased the accuracy a small amount by adding an L1 and L2 regularization penalty. What errors are still being made by the model?

Let’s see how the weights look after adding L1_penalty > 0.

W_in = net.in_layer.weight.detach().numpy() # we can run detach and numpy to get a numpy array
print('shape of W_in:')
print(W_in.shape)

W_out = net.out_layer.weight.detach().numpy()

visualize_weights(W_in[:,isort], W_out)

The weights appear to be sparser than before.


Section 3: Encoding - Convolutional Networks for Encoding

Video 2: Convolutional Encoding Model

In neuroscience, we often want to understand how the brain represents external stimuli. One approach to discovering these representations is to build an encoding model that takes as input the external stimuli (in this case grating stimuli) and outputs the neural responses.

Because visual cortex is often thought to be a convolutional network where the same filters are combined across the visual field, we will use a model with a convolutional layer. We learned how to build a convolutional layer in the previous section. We will add to this convolutional layer a fully connected layer from the output of the convolutions to the neurons. We will then visualize the weights of this fully connected layer.

Execute this cell to load data

# @markdown Execute this cell to load data
import hashlib
import requests

fname = "W3D4_stringer_oribinned6_split.npz"
url = "https://osf.io/p3aeb/download"
expected_md5 = "b3f7245c6221234a676b71a1f43c3bb5"

if not os.path.isfile(fname):
  try:
    r = requests.get(url)
  except requests.ConnectionError:
    print("!!! Failed to download data !!!")
  else:
    if r.status_code != requests.codes.ok:
      print("!!! Failed to download data !!!")
    elif hashlib.md5(r.content).hexdigest() != expected_md5:
      print("!!! Data download appears corrupted !!!")
    else:
      with open(fname, "wb") as fid:
        fid.write(r.content)

def load_data_split(data_name=fname):
  """Load mouse V1 data from Stringer et al. (2019)

  Data from study reported in this preprint:
  https://www.biorxiv.org/content/10.1101/679324v2.abstract

  These data comprise time-averaged responses of ~20,000 neurons
  to ~4,000 stimulus gratings of different orientations, recorded
  through Calcium imaginge. The responses have been normalized by
  spontaneous levels of activity and then z-scored over stimuli, so
  expect negative numbers. The repsonses were split into train and
  test and then each set were averaged in bins of 6 degrees.

  This function returns the relevant data (neural responses and
  stimulus orientations) in a torch.Tensor of data type torch.float32
  in order to match the default data type for nn.Parameters in
  Google Colab.

  It will hold out some of the trials when averaging to allow us to have test
  tuning curves.

  Args:
    data_name (str): filename to load

  Returns:
    resp_train (torch.Tensor): n_stimuli x n_neurons matrix of neural responses,
        each row contains the responses of each neuron to a given stimulus.
        As mentioned above, neural "response" is actually an average over
        responses to stimuli with similar angles falling within specified bins.
    resp_test (torch.Tensor): n_stimuli x n_neurons matrix of neural responses,
        each row contains the responses of each neuron to a given stimulus.
        As mentioned above, neural "response" is actually an average over
        responses to stimuli with similar angles falling within specified bins
    stimuli: (torch.Tensor): n_stimuli x 1 column vector with orientation
        of each stimulus, in degrees. This is actually the mean orientation
        of all stimuli in each bin.

  """
  with np.load(data_name) as dobj:
    data = dict(**dobj)
  resp_train = data['resp_train']
  resp_test = data['resp_test']
  stimuli = data['stimuli']

  # Return as torch.Tensor
  resp_train_tensor = torch.tensor(resp_train, dtype=torch.float32)
  resp_test_tensor = torch.tensor(resp_test, dtype=torch.float32)
  stimuli_tensor = torch.tensor(stimuli, dtype=torch.float32)

  return resp_train_tensor, resp_test_tensor, stimuli_tensor

Section 3.1: Neural tuning curves

In the next cell, we plot the turning curves of a random subset of neurons. We have binned the stimuli orientations more than in Tutorial 1. We create the gratings images as above for the 60 orientations below, and save them to a variable grating_stimuli.

Rerun the cell to look at different example neurons and observe the diversity of tuning curves in the population. How can we fit these neural responses with an encoding model?

Execute this cell to load data, create stimuli, and plot neural tuning curves

# @markdown Execute this cell to load data, create stimuli, and plot neural tuning curves

### Load data and bin at 8 degrees
# responses are split into test and train
resp_train, resp_test, stimuli = load_data_split()
n_stimuli, n_neurons = resp_train.shape
print('resp_train contains averaged responses of %i neurons to %i binned stimuli' % (n_neurons, n_stimuli))
#print(resp_train.shape)

# also make stimuli into images
orientations = np.linspace(0, 360, 61)[:-1] - 180
grating_stimuli = np.zeros((60, 1, 12, 16), np.float32)
for i, ori in enumerate(orientations):
  grating_stimuli[i,0] = grating(ori, res=0.025)#[18:30, 24:40]

grating_stimuli = torch.from_numpy(grating_stimuli)
print('grating_stimuli contains 60 stimuli of size 12 x 16')

# Visualize tuning curves
fig, axs = plt.subplots(3, 5, figsize=(15,7))
for k, ax in enumerate(axs.flatten()):
  neuron_index = np.random.choice(n_neurons)  # pick random neuron
  plot_tuning(ax, stimuli, resp_train[:, neuron_index], resp_test[:, neuron_index], neuron_index, linewidth=2)
  if k==0:
    ax.text(1.0, 0.9, 'train', color='y', transform=ax.transAxes)
    ax.text(1.0, 0.65, 'test', color='m', transform=ax.transAxes)
fig.tight_layout()
plt.show()
resp_train contains averaged responses of 23589 neurons to 60 binned stimuli
grating_stimuli contains 60 stimuli of size 12 x 16
../../../_images/W2D1_Tutorial4_76_1.png

Section 3.2: Adding a fully-connected layer to create encoding model

We will build a torch model like above with a convolutional layer. Additionally, we will add a fully connected linear layer from the convolutional units to the neurons. We will use 6 convolutional channels (\(C^{out}\)) and a kernel size (\(K\)) of 7 with a stride of 1 and padding of \(K/2\) (same as above). The stimulus is size (12, 16). Then the convolutional unit activations will go through a linear layer to be transformed into neural responses.

Think! 3.2: Number of unis and weights

  • How many units will the convolutional layer have?

  • How many weights will the fully connected linear layer from convolutional units to neurons have?

Click for solution

Coding Exercise 3.2: Add linear layer

Remember in Tutorial 1 we used linear layers. Use your knowledge from Tutorial 1 to add a linear layer to the model we created above.

Execute to get train function for our neural encoding model

# @markdown Execute to get `train` function for our neural encoding model

def train(net, custom_loss, train_data, train_labels,
          test_data=None, test_labels=None,
          learning_rate=10, n_iter=500, L2_penalty=0., L1_penalty=0.):
  """Run gradient descent for network without batches

  Args:
    net (nn.Module): deep network whose parameters to optimize with SGD
    custom_loss: loss function for network
    train_data: training data (n_train x input features)
    train_labels: training labels (n_train x output features)
    test_data: test data (n_train x input features)
    test_labels: test labels (n_train x output features)
    learning_rate (float): learning rate for gradient descent
    n_epochs (int): number of epochs to run gradient descent
    L2_penalty (float): magnitude of L2 penalty
    L1_penalty (float): magnitude of L1 penalty

  Returns:
    train_loss: training loss across iterations
    test_loss: testing loss across iterations

  """
  optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9) # Initialize PyTorch SGD optimizer
  train_loss = np.nan * np.zeros(n_iter)  # Placeholder for train loss
  test_loss = np.nan * np.zeros(n_iter)  # Placeholder for test loss

  # Loop over epochs
  for i in range(n_iter):
    if n_iter < 10:
      for param_group in self.optimizer.param_groups:
        param_group['lr'] = np.linspace(0, learning_rate, 10)[n_iter]
    y_pred = net(train_data) # Forward pass: compute predicted y by passing train_data to the model.

    if L2_penalty>0 or L1_penalty>0:
      weights = net.out_layer.weight
      loss = custom_loss(y_pred, train_labels, weights, L2_penalty, L1_penalty)
    else:
      loss = custom_loss(y_pred, train_labels)

    ### Update parameters
    optimizer.zero_grad() # zero out gradients
    loss.backward() # Backward pass: compute gradient of the loss with respect to model parameters
    optimizer.step() # step parameters in gradient direction
    train_loss[i] = loss.item()  # .item() transforms the tensor to a scalar and does .detach() for us

    # Track progress
    if (i+1) % (n_iter // 10) == 0 or i==0:
      if test_data is not None and test_labels is not None:
        y_pred = net(test_data)
        if L2_penalty>0 or L1_penalty>0:
          loss = custom_loss(y_pred, test_labels, weights, L2_penalty, L1_penalty)
        else:
          loss = custom_loss(y_pred, test_labels)
        test_loss[i] = loss.item()
        print(f'iteration {i+1}/{n_iter} | train loss: {train_loss[i]:.4f} | test loss: {test_loss[i]:.4f}')
      else:
        print(f'iteration {i+1}/{n_iter} | train loss: {train_loss[i]:.4f}')

  return train_loss, test_loss
class ConvFC(nn.Module):
  """Deep network with one convolutional layer + one fully connected layer

  Attributes:
    conv (nn.Conv1d): convolutional layer
    dims (tuple): shape of convolutional layer output
    out_layer (nn.Linear): linear layer

  """

  def __init__(self, n_neurons, c_in=1, c_out=6, K=7, b=12*16,
               filters=None):
    """ initialize layer
    Args:
        c_in: number of input stimulus channels
        c_out: number of convolutional channels
        K: size of each convolutional filter
        h: number of stimulus bins, n_bins
    """
    super().__init__()
    self.conv = nn.Conv2d(c_in, c_out, kernel_size=K,
                          padding=K//2, stride=1)
    self.dims = (c_out, b)  # dimensions of conv layer output
    M = np.prod(self.dims) # number of hidden units

    ################################################################################
    ## TO DO for students: add fully connected layer to network (self.out_layer)
    # Fill out function and remove
    raise NotImplementedError("Student exercise: add fully connected layer to initialize network")
    ################################################################################
    self.out_layer = nn.Linear(M, ...)

    # initialize weights
    if filters is not None:
      self.conv.weight = nn.Parameter(filters)
      self.conv.bias = nn.Parameter(torch.zeros((c_out,), dtype=torch.float32))

    nn.init.normal_(self.out_layer.weight, std=0.01) # initialize weights to be small

  def forward(self, s):
    """ Predict neural responses to stimuli s

    Args:
        s (torch.Tensor): n_stimuli x c_in x h x w tensor with stimuli

    Returns:
        y (torch.Tensor): n_stimuli x n_neurons tensor with predicted neural responses

    """
    a = self.conv(s)  # output of convolutional layer
    a = a.view(-1, np.prod(self.dims))  # flatten each convolutional layer output into a vector

    ################################################################################
    ## TO DO for students: add fully connected layer to forward pass of network (self.out_layer)
    # Fill out function and remove
    raise NotImplementedError("Student exercise: add fully connected layer to network")
    ################################################################################
    y = ...

    return y


device = torch.device('cpu')

# (Optional) To speed up processing, go to "Runtime" menu and "Change runtime"
# and select GPU processing, then uncomment line below, otherwise runtime will
# be ~ 2 minutes
# device = torch.device('cuda')

# Initialize network
n_neurons = resp_train.shape[1]
## Initialize with filters from Tutorial 2
example_filters = filters(out_channels=6, K=7)

net = ConvFC(n_neurons, filters = example_filters)
net = net.to(device)

# Run GD on training set data
# ** this time we are also providing the test data to estimate the test loss
train_loss, test_loss = train(net, regularized_MSE_loss,
                              train_data=grating_stimuli.to(device), train_labels=resp_train.to(device),
                              test_data=grating_stimuli.to(device), test_labels=resp_test.to(device),
                              n_iter=200, learning_rate=2,
                              L2_penalty=5e-4, L1_penalty=1e-6)

# Visualize
plot_training_curves(train_loss, test_loss)

Click for solution

Example output:

Solution hint

How well can we fit single neuron tuning curves with this model? What aspects of the tuning curves are we capturing?

Execute this cell to examine predictions for random subsets of neurons

# @markdown Execute this cell to examine predictions for random subsets of neurons

y_pred = net(grating_stimuli.to(device))
# Visualize tuning curves & plot neural predictions
fig, axs = plt.subplots(2, 5, figsize=(15,6))
for k, ax in enumerate(axs.flatten()):
  ineur = np.random.choice(n_neurons)
  plot_prediction(ax, y_pred[:,ineur].detach().cpu(),
                  resp_train[:,ineur],
                  resp_test[:,ineur])
  if k==0:
    ax.text(.1, 1., 'train', color='y', transform=ax.transAxes)
    ax.text(.1, 0.9, 'test', color='m', transform=ax.transAxes)
    ax.text(.1, 0.8, 'prediction', color='g', transform=ax.transAxes)

fig.tight_layout()
plt.show()
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_3806/3835599226.py in <module>
      1 # @markdown Execute this cell to examine predictions for random subsets of neurons
      2 
----> 3 y_pred = net(grating_stimuli.to(device))
      4 # Visualize tuning curves & plot neural predictions
      5 fig, axs = plt.subplots(2, 5, figsize=(15,6))

/opt/hostedtoolcache/Python/3.7.11/x64/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1049         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1050                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051             return forward_call(*input, **kwargs)
   1052         # Do not call functions when jit is used
   1053         full_backward_hooks, non_full_backward_hooks = [], []

/tmp/ipykernel_3806/2166509181.py in forward(self, r)
     31 
     32     """
---> 33     h = torch.relu(self.in_layer(r))
     34     logp = self.logprob(self.out_layer(h))
     35     return logp

/opt/hostedtoolcache/Python/3.7.11/x64/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1049         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1050                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051             return forward_call(*input, **kwargs)
   1052         # Do not call functions when jit is used
   1053         full_backward_hooks, non_full_backward_hooks = [], []

/opt/hostedtoolcache/Python/3.7.11/x64/lib/python3.7/site-packages/torch/nn/modules/linear.py in forward(self, input)
     94 
     95     def forward(self, input: Tensor) -> Tensor:
---> 96         return F.linear(input, self.weight, self.bias)
     97 
     98     def extra_repr(self) -> str:

/opt/hostedtoolcache/Python/3.7.11/x64/lib/python3.7/site-packages/torch/nn/functional.py in linear(input, weight, bias)
   1845     if has_torch_function_variadic(input, weight):
   1846         return handle_torch_function(linear, (input, weight), input, weight, bias=bias)
-> 1847     return torch._C._nn.linear(input, weight, bias)
   1848 
   1849 

RuntimeError: mat1 and mat2 shapes cannot be multiplied (720x16 and 23589x20)

We can see if the convolutional channels changed at all from their initialization as center-surround and Gabor filters. If they don’t then it means that they were a sufficient basis set to explain the responses of the neurons to orientations to the accuracy seen above.

# get weights of conv layer in convLayer
out_channels = 6 # how many convolutional channels to have in our layer
weights = net.conv.weight.detach().cpu()
print(weights.shape) # can you identify what each of the dimensions are?

plot_weights(weights, channels = np.arange(0, out_channels))
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_3806/356809345.py in <module>
      1 # get weights of conv layer in convLayer
      2 out_channels = 6 # how many convolutional channels to have in our layer
----> 3 weights = net.conv.weight.detach().cpu()
      4 print(weights.shape) # can you identify what each of the dimensions are?
      5 

/opt/hostedtoolcache/Python/3.7.11/x64/lib/python3.7/site-packages/torch/nn/modules/module.py in __getattr__(self, name)
   1129                 return modules[name]
   1130         raise AttributeError("'{}' object has no attribute '{}'".format(
-> 1131             type(self).__name__, name))
   1132 
   1133     def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None:

AttributeError: 'DeepNetSoftmax' object has no attribute 'conv'