Related
When we train neural networks, we typically use gradient descent, which relies on a continuous, differentiable real-valued cost function. The final cost function might, for example, take the mean squared error. Or put another way, gradient descent implicitly assumes the end goal is regression - to minimize a real-valued error measure.
Sometimes what we want a neural network to do is perform classification - given an input, classify it into two or more discrete categories. In this case, the end goal the user cares about is classification accuracy - the percentage of cases classified correctly.
But when we are using a neural network for classification, though our goal is classification accuracy, that is not what the neural network is trying to optimize. The neural network is still trying to optimize the real-valued cost function. Sometimes these point in the same direction, but sometimes they don't. In particular, I've been running into cases where a neural network trained to correctly minimize the cost function, has a classification accuracy worse than a simple hand-coded threshold comparison.
I've boiled this down to a minimal test case using TensorFlow. It sets up a perceptron (neural network with no hidden layers), trains it on an absolutely minimal dataset (one input variable, one binary output variable) assesses the classification accuracy of the result, then compares it to the classification accuracy of a simple hand-coded threshold comparison; the results are 60% and 80% respectively. Intuitively, this is because a single outlier with a large input value, generates a correspondingly large output value, so the way to minimize the cost function is to try extra hard to accommodate that one case, in the process misclassifying two more ordinary cases. The perceptron is correctly doing what it was told to do; it's just that this does not match what we actually want of a classifier. But the classification accuracy is not a continuous differentiable function, so we can't use it as the target for gradient descent.
How can we train a neural network so that it ends up maximizing classification accuracy?
import numpy as np
import tensorflow as tf
sess = tf.InteractiveSession()
tf.set_random_seed(1)
# Parameters
epochs = 10000
learning_rate = 0.01
# Data
train_X = [
[0],
[0],
[2],
[2],
[9],
]
train_Y = [
0,
0,
1,
1,
0,
]
rows = np.shape(train_X)[0]
cols = np.shape(train_X)[1]
# Inputs and outputs
X = tf.placeholder(tf.float32)
Y = tf.placeholder(tf.float32)
# Weights
W = tf.Variable(tf.random_normal([cols]))
b = tf.Variable(tf.random_normal([]))
# Model
pred = tf.tensordot(X, W, 1) + b
cost = tf.reduce_sum((pred-Y)**2/rows)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
tf.global_variables_initializer().run()
# Train
for epoch in range(epochs):
# Print update at successive doublings of time
if epoch&(epoch-1) == 0 or epoch == epochs-1:
print('{} {} {} {}'.format(
epoch,
cost.eval({X: train_X, Y: train_Y}),
W.eval(),
b.eval(),
))
optimizer.run({X: train_X, Y: train_Y})
# Classification accuracy of perceptron
classifications = [pred.eval({X: x}) > 0.5 for x in train_X]
correct = sum([p == y for (p, y) in zip(classifications, train_Y)])
print('{}/{} = perceptron accuracy'.format(correct, rows))
# Classification accuracy of hand-coded threshold comparison
classifications = [x[0] > 1.0 for x in train_X]
correct = sum([p == y for (p, y) in zip(classifications, train_Y)])
print('{}/{} = threshold accuracy'.format(correct, rows))
How can we train a neural network so that it ends up maximizing classification accuracy?
I'm asking for a way to get a continuous proxy function that's closer to the accuracy
To start with, the loss function used today for classification tasks in (deep) neural nets was not invented with them, but it goes back several decades, and it actually comes from the early days of logistic regression. Here is the equation for the simple case of binary classification:
The idea behind it was exactly to come up with a continuous & differentiable function, so that we would be able to exploit the (vast, and still expanding) arsenal of convex optimization for classification problems.
It is safe to say that the above loss function is the best we have so far, given the desired mathematical constraints mentioned above.
Should we consider this problem (i.e. better approximating the accuracy) solved and finished? At least in principle, no. I am old enough to remember an era when the only activation functions practically available were tanh and sigmoid; then came ReLU and gave a real boost to the field. Similarly, someone may eventually come up with a better loss function, but arguably this is going to happen in a research paper, and not as an answer to a SO question...
That said, the very fact that the current loss function comes from very elementary considerations of probability and information theory (fields that, in sharp contrast with the current field of deep learning, stand upon firm theoretical foundations) creates at least some doubt as to if a better proposal for the loss may be just around the corner.
There is another subtle point on the relation between loss and accuracy, which makes the latter something qualitatively different than the former, and is frequently lost in such discussions. Let me elaborate a little...
All the classifiers related to this discussion (i.e. neural nets, logistic regression etc) are probabilistic ones; that is, they do not return hard class memberships (0/1) but class probabilities (continuous real numbers in [0, 1]).
Limiting the discussion for simplicity to the binary case, when converting a class probability to a (hard) class membership, we are implicitly involving a threshold, usually equal to 0.5, such as if p[i] > 0.5, then class[i] = "1". Now, we can find many cases whet this naive default choice of threshold will not work (heavily imbalanced datasets are the first to come to mind), and we'll have to choose a different one. But the important point for our discussion here is that this threshold selection, while being of central importance to the accuracy, is completely external to the mathematical optimization problem of minimizing the loss, and serves as a further "insulation layer" between them, compromising the simplistic view that loss is just a proxy for accuracy (it is not). As nicely put in the answer of this Cross Validated thread:
the statistical component of your exercise ends when you output a probability for each class of your new sample. Choosing a threshold beyond which you classify a new observation as 1 vs. 0 is not part of the statistics any more. It is part of the decision component.
Enlarging somewhat an already broad discussion: Can we possibly move completely away from the (very) limiting constraint of mathematical optimization of continuous & differentiable functions? In other words, can we do away with back-propagation and gradient descend?
Well, we are actually doing so already, at least in the sub-field of reinforcement learning: 2017 was the year when new research from OpenAI on something called Evolution Strategies made headlines. And as an extra bonus, here is an ultra-fresh (Dec 2017) paper by Uber on the subject, again generating much enthusiasm in the community.
I think you are forgetting to pass your output through a simgoid. Fixed below:
import numpy as np
import tensorflow as tf
sess = tf.InteractiveSession()
tf.set_random_seed(1)
# Parameters
epochs = 10000
learning_rate = 0.01
# Data
train_X = [
[0],
[0],
[2],
[2],
[9],
]
train_Y = [
0,
0,
1,
1,
0,
]
rows = np.shape(train_X)[0]
cols = np.shape(train_X)[1]
# Inputs and outputs
X = tf.placeholder(tf.float32)
Y = tf.placeholder(tf.float32)
# Weights
W = tf.Variable(tf.random_normal([cols]))
b = tf.Variable(tf.random_normal([]))
# Model
# CHANGE HERE: Remember, you need an activation function!
pred = tf.nn.sigmoid(tf.tensordot(X, W, 1) + b)
cost = tf.reduce_sum((pred-Y)**2/rows)
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
tf.global_variables_initializer().run()
# Train
for epoch in range(epochs):
# Print update at successive doublings of time
if epoch&(epoch-1) == 0 or epoch == epochs-1:
print('{} {} {} {}'.format(
epoch,
cost.eval({X: train_X, Y: train_Y}),
W.eval(),
b.eval(),
))
optimizer.run({X: train_X, Y: train_Y})
# Classification accuracy of perceptron
classifications = [pred.eval({X: x}) > 0.5 for x in train_X]
correct = sum([p == y for (p, y) in zip(classifications, train_Y)])
print('{}/{} = perceptron accuracy'.format(correct, rows))
# Classification accuracy of hand-coded threshold comparison
classifications = [x[0] > 1.0 for x in train_X]
correct = sum([p == y for (p, y) in zip(classifications, train_Y)])
print('{}/{} = threshold accuracy'.format(correct, rows))
The output:
0 0.28319069743156433 [ 0.75648874] -0.9745011329650879
1 0.28302448987960815 [ 0.75775659] -0.9742625951766968
2 0.28285878896713257 [ 0.75902224] -0.9740257859230042
4 0.28252947330474854 [ 0.76154679] -0.97355717420578
8 0.28187844157218933 [ 0.76656926] -0.9726400971412659
16 0.28060704469680786 [ 0.77650583] -0.970885694026947
32 0.27818527817726135 [ 0.79593837] -0.9676888585090637
64 0.2738055884838104 [ 0.83302218] -0.9624817967414856
128 0.26666420698165894 [ 0.90031379] -0.9562843441963196
256 0.25691407918930054 [ 1.01172411] -0.9567816257476807
512 0.2461051195859909 [ 1.17413962] -0.9872989654541016
1024 0.23519910871982574 [ 1.38549554] -1.088881492614746
2048 0.2241383194923401 [ 1.64616168] -1.298340916633606
4096 0.21433120965957642 [ 1.95981205] -1.6126530170440674
8192 0.2075471431016922 [ 2.31746769] -1.989408016204834
9999 0.20618653297424316 [ 2.42539024] -2.1028473377227783
4/5 = perceptron accuracy
4/5 = threshold accuracy
I'm a newbie to machine learning and this is one of the first real-world ML tasks challenged.
Some experimental data contains 512 independent boolean features and a boolean result.
There are about 1e6 real experiment records in the provided data set.
In a classic XOR example all 4 out of 4 possible states are required to train NN. In my case its only 2^(10-512) = 2^-505 which is close to zero.
I have no more information about the data nature, just these (512 + 1) * 1e6 bits.
Tried NN with 1 hidden layer on available data. Output of the trained NN on the samples even from the training set are always close to 0, not a single close to "1". Played with weights initialization, gradient descent learning rate.
My code utilizing TensorFlow 1.3, Python 3. Model excerpt:
with tf.name_scope("Layer1"):
#W1 = tf.Variable(tf.random_uniform([512, innerN], minval=-2/512, maxval=2/512), name="Weights_1")
W1 = tf.Variable(tf.zeros([512, innerN]), name="Weights_1")
b1 = tf.Variable(tf.zeros([1]), name="Bias_1")
Out1 = tf.sigmoid( tf.matmul(x, W1) + b1)
with tf.name_scope("Layer2"):
W2 = tf.Variable(tf.random_uniform([innerN, 1], minval=-2/512, maxval=2/512), name="Weights_2")
#W2 = tf.Variable(tf.zeros([innerN, 1]), name="Weights_2")
b2 = tf.Variable(tf.zeros([1]), name="Bias_2")
y = tf.nn.sigmoid( tf.matmul(Out1, W2) + b2)
with tf.name_scope("Training"):
y_ = tf.placeholder(tf.float32, [None,1])
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
labels = y_, logits = y)
)
train_step = tf.train.GradientDescentOptimizer(0.005).minimize(cross_entropy)
with tf.name_scope("Testing"):
# Test trained model
correct_prediction = tf.equal( tf.round(y), tf.round(y_))
# ...
# Train
for step in range(500):
batch_xs, batch_ys = Datasets.train.next_batch(300, shuffle=False)
_, my_y, summary = sess.run([train_step, y, merged_summaries],
feed_dict={x: batch_xs, y_: batch_ys})
I suspect two cases:
my fault – bad NN implementation, wrong architecture;
bad data. Compared to XOR example, incomplete training data would result in a failing NN. However, the training examples fed to the trained NN are supposed to give right predictions, aren't they?
How to evaluate if it is possible at all to train a neural network (a 2-layer perceptron) on the provided data to forecast the result? A case of aceptable set would be the XOR example. Opposed to some random noise.
There are only ad hoc ways to know if it is possible to learn a function with a differentiable network from a dataset. That said, these ad hoc ways do usually work. For example, the network should be able to overfit the training set without any regularisation.
A common technique to gauge this is to only fit the network on a subset of the full dataset. Check that the network can overfit to that, then increase the size of the subset, and increase the size of the network as well. Unfortunately, deciding whether to add extra layers or add more units in a hidden layer is an arbitrary decision you'll have to make.
However, looking at your code, there are a few things that could be going wrong here:
Are your outputs balanced? By that I mean, do you have the same number of 1s as 0s in the dataset targets?
Your initialisation in the first layer is all zeros, the gradient to this will be zero, so it can't learn anything (although, you have a real initialisation above it commented out).
Sigmoid nonlinearities are more difficult to optimise than simpler nonlinearities, such as ReLUs.
I'd recommend using the built-in definitions for layers in Tensorflow to not worry about initialisation, and switching to ReLUs in any hidden layers (you need sigmoid at the output for your boolean target).
Finally, deep learning isn't actually very good at most "bag of features" machine learning problems because they lack structure. For example, the order of the features doesn't matter. Other methods often work better, but if you really want to use deep learning then you could look at this recent paper, showing improved performance by just using a very specific nonlinearity and weight initialisation (change 4 lines in your code above).
Given a set of training examples for training a neural network, we want to give more or less weight to various examples in training. We apply a weight between 0.0 and 1.0 to each example based on some criteria for the "value" (e.g. validity or confidence) of the example. How can this be implemented in Tensorflow, in particular when using tf.nn.sparse_softmax_cross_entropy_with_logits()?
In the most common case where you call tf.nn.sparse_softmax_cross_entropy_with_logits with logits of shape [batch_size, num_classes] and labels of shape [batch_size], the function returns a tensor of shape batch_size. You can multiply this tensor with a weight tensor before reducing them to a single loss value:
weights = tf.placeholder(name="loss_weights", shape=[None], dtype=tf.float32)
loss_per_example = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels)
loss = tf.reduce_mean(weights * loss_per_example)
I am messing around with LSTMs and have a conceptual question. I created a matrix of bogus data on the following rules:
For each 1-D list in the matrix:
If previous element is less than 10, then this next element is the previous one plus 1.
Else, this element is sin(previous element)
This way, it is a sequence that is pretty simply based on the previous information. I set up an LSTM to learn the recurrence and ran it to train on the lists one at a time. I have an LSTM layer followed by a fully connected feed-forward layer. It learns the +1 step very easily, but has trouble with the sin step. It will seemingly pick a random number between -1 and 1 when making the next element when the previous one was greater than 10. My question is this: is the training only modifying the variables in my fully connected feed forward layer? Is that why it can't learn the non-linear sin function?
Here's the code snippet in question:
lstm = rnn_cell.LSTMCell(lstmSize)
y_ = tf.placeholder(tf.float32, [None, OS])
outputs, state = rnn.rnn(lstm, x, dtype=tf.float32)
outputs = tf.transpose(outputs, [1, 0, 2])
last = tf.gather(outputs, int(outputs.get_shape()[0]) - 1)
weights = tf.Variable(tf.truncated_normal([lstmSize, OS]))
bias = tf.Variable(tf.constant(0.1, shape=[OS]))
y = tf.nn.elu(tf.matmul(last, weights) + bias)
error = tf.reduce_mean(tf.square(tf.sub(y_, y)))
train_step = tf.train.AdamOptimizer(learning_rate=1e-3).minimize(error)
The error and shape organization seems to be correct, at least in the sense that it does learn the +1 step quickly without crashing. Shouldn't the LSTM be able to handle the non-linear sin function? It seems almost trivially easy, so my guess is that I set something up wrong and the LSTM isn't learning anything.
I am trying to solve the regression task. I found out that 3 models are working nicely for different subsets of data: LassoLARS, SVR and Gradient Tree Boosting. I noticed that when I make predictions using all these 3 models and then make a table of 'true output' and outputs of my 3 models I see that each time at least one of the models is really close to the true output, though 2 others could be relatively far away.
When I compute minimal possible error (if I take prediction from 'best' predictor for each test example) I get a error which is much smaller than error of any model alone. So I thought about trying to combine predictions from these 3 diffent models into some kind of ensemble. Question is, how to do this properly? All my 3 models are build and tuned using scikit-learn, does it provide some kind of a method which could be used to pack models into ensemble? The problem here is that I don't want to just average predictions from all three models, I want to do this with weighting, where weighting should be determined based on properties of specific example.
Even if scikit-learn not provides such functionality, it would be nice if someone knows how to property address this task - of figuring out the weighting of each model for each example in data. I think that it might be done by a separate regressor built on top of all these 3 models, which will try output optimal weights for each of 3 models, but I am not sure if this is the best way of doing this.
This is a known interesting (and often painful!) problem with hierarchical predictions. A problem with training a number of predictors over the train data, then training a higher predictor over them, again using the train data - has to do with the bias-variance decomposition.
Suppose you have two predictors, one essentially an overfitting version of the other, then the former will appear over the train set to be better than latter. The combining predictor will favor the former for no true reason, just because it cannot distinguish overfitting from true high-quality prediction.
The known way of dealing with this is to prepare, for each row in the train data, for each of the predictors, a prediction for the row, based on a model not fit for this row. For the overfitting version, e.g., this won't produce a good result for the row, on average. The combining predictor will then be able to better assess a fair model for combining the lower-level predictors.
Shahar Azulay & I wrote a transformer stage for dealing with this:
class Stacker(object):
"""
A transformer applying fitting a predictor `pred` to data in a way
that will allow a higher-up predictor to build a model utilizing both this
and other predictors correctly.
The fit_transform(self, x, y) of this class will create a column matrix, whose
each row contains the prediction of `pred` fitted on other rows than this one.
This allows a higher-level predictor to correctly fit a model on this, and other
column matrices obtained from other lower-level predictors.
The fit(self, x, y) and transform(self, x_) methods, will fit `pred` on all
of `x`, and transform the output of `x_` (which is either `x` or not) using the fitted
`pred`.
Arguments:
pred: A lower-level predictor to stack.
cv_fn: Function taking `x`, and returning a cross-validation object. In `fit_transform`
th train and test indices of the object will be iterated over. For each iteration, `pred` will
be fitted to the `x` and `y` with rows corresponding to the
train indices, and the test indices of the output will be obtained
by predicting on the corresponding indices of `x`.
"""
def __init__(self, pred, cv_fn=lambda x: sklearn.cross_validation.LeaveOneOut(x.shape[0])):
self._pred, self._cv_fn = pred, cv_fn
def fit_transform(self, x, y):
x_trans = self._train_transform(x, y)
self.fit(x, y)
return x_trans
def fit(self, x, y):
"""
Same signature as any sklearn transformer.
"""
self._pred.fit(x, y)
return self
def transform(self, x):
"""
Same signature as any sklearn transformer.
"""
return self._test_transform(x)
def _train_transform(self, x, y):
x_trans = np.nan * np.ones((x.shape[0], 1))
all_te = set()
for tr, te in self._cv_fn(x):
all_te = all_te | set(te)
x_trans[te, 0] = self._pred.fit(x[tr, :], y[tr]).predict(x[te, :])
if all_te != set(range(x.shape[0])):
warnings.warn('Not all indices covered by Stacker', sklearn.exceptions.FitFailedWarning)
return x_trans
def _test_transform(self, x):
return self._pred.predict(x)
Here is an example of the improvement for the setting described in #MaximHaytovich's answer.
First, some setup:
from sklearn import linear_model
from sklearn import cross_validation
from sklearn import ensemble
from sklearn import metrics
y = np.random.randn(100)
x0 = (y + 0.1 * np.random.randn(100)).reshape((100, 1))
x1 = (y + 0.1 * np.random.randn(100)).reshape((100, 1))
x = np.zeros((100, 2))
Note that x0 and x1 are just noisy versions of y. We'll use the first 80 rows for train, and the last 20 for test.
These are the two predictors: a higher-variance gradient booster, and a linear predictor:
g = ensemble.GradientBoostingRegressor()
l = linear_model.LinearRegression()
Here is the methodology suggested in the answer:
g.fit(x0[: 80, :], y[: 80])
l.fit(x1[: 80, :], y[: 80])
x[:, 0] = g.predict(x0)
x[:, 1] = l.predict(x1)
>>> metrics.r2_score(
y[80: ],
linear_model.LinearRegression().fit(x[: 80, :], y[: 80]).predict(x[80: , :]))
0.940017788444
Now, using stacking:
x[: 80, 0] = Stacker(g).fit_transform(x0[: 80, :], y[: 80])[:, 0]
x[: 80, 1] = Stacker(l).fit_transform(x1[: 80, :], y[: 80])[:, 0]
u = linear_model.LinearRegression().fit(x[: 80, :], y[: 80])
x[80: , 0] = Stacker(g).fit(x0[: 80, :], y[: 80]).transform(x0[80:, :])
x[80: , 1] = Stacker(l).fit(x1[: 80, :], y[: 80]).transform(x1[80:, :])
>>> metrics.r2_score(
y[80: ],
u.predict(x[80:, :]))
0.992196564279
The stacking prediction does better. It realizes that the gradient booster is not that great.
Ok, after spending some time on googling 'stacking' (as mentioned by #andreas earlier) I found out how I could do the weighting in python even with scikit-learn. Consider the below:
I train a set of my regression models (as mentioned SVR, LassoLars and GradientBoostingRegressor). Then I run all of them on training data (same data which was used for training of each of these 3 regressors). I get predictions for examples with each of my algorithms and save these 3 results into pandas dataframe with columns 'predictedSVR', 'predictedLASSO' and 'predictedGBR'. And I add the final column into this datafrane which I call 'predicted' which is a real prediction value.
Then I just train a linear regression on this new dataframe:
#df - dataframe with results of 3 regressors and true output
from sklearn linear_model
stacker= linear_model.LinearRegression()
stacker.fit(df[['predictedSVR', 'predictedLASSO', 'predictedGBR']], df['predicted'])
So when I want to make a prediction for new example I just run each of my 3 regressors separately and then I do:
stacker.predict()
on outputs of my 3 regressors. And get a result.
The problem here is that I am finding optimal weights for regressors 'on average, the weights will be same for each example on which I will try to make prediction.
What you describe is called "stacking" which is not implemented in scikit-learn yet, but I think contributions would be welcome. An ensemble that just averages will be in pretty soon: https://github.com/scikit-learn/scikit-learn/pull/4161
Late response, but I wanted to add one practical point for this sort of stacked regression approach (which I use this frequently in my work).
You may want to choose an algorithm for the stacker which allows positive=True (for example, ElasticNet). I have found that, when you have one relatively stronger model, the unconstrained LinearRegression() model will often fit a larger positive coefficient to the stronger and a negative coefficient to the weaker model.
Unless you actually believe that your weaker model has negative predictive power, this is not a helpful outcome. Very similar to having high multi-colinearity between features of a regular regression model. Causes all sorts of edge effects.
This comment applies most significantly to noisy data situations. If you're aiming to get RSQ of 0.9-0.95-0.99, you'd probably want to throw out the model which was getting a negative weighting.