Errata

In chapter 3 the book (sensibly) gives instructions on how to install Tensorflow for Python2.x and Python 3.x. Yet, as given the examples are incompatible with Python3.

In [5]: a = tf.constant(2)
In [6]: a = tf.constant(2)
In [7]: multiply = tf.mul(a, b)
In [7]: session.run(multiply)
Out[7]: 6

Must be:
In [5]: a = tf.constant(2)
In [6]: b = tf.constant(3) <-----------------
In [7]: multiply = tf.mul(a, b)
In [7]: session.run(multiply)
Out[7]: 6

tf.mul has been replaced by tf.multiply

the definition of the logit has a sum from i=0 to n, but the sum should be from i=1 to n (since the input/features and weights as given earlier in the paragraph are indexed from 1).

The output of the neuron is given as y = f(x.w+b) where x and w are vectors (in bold). One of either x or w should be transposed before multiplication. e.g. if the vectors are column vectors, then we should have: f(x'.w + b) where x' is the transpose of x.

"ReLU" is being described as "restricted linear units" where "Re" actually stands for "rectified".

The weight for both Soda and Fries is shown as w_2.

The last term on the third line is the partial derivative of y with respect to the weight w_k. The y is indexed with a subscript "i", but the index should instead appear as a superscript in parentheses (like the other y^(i) in the same equation). In other words, in LaTeX notation: replace y_i with y^{(i)}

In [6]: a = tf.constant(2)

should be

ln[6]: b = tf.constant(3)

and

ln[7]: multiply = tf.mul(a, b)

should be

ln[7]: multiply = tf.multiply(a, b)

`from read_data import get_minibatch()`

Module does not exist. Googling for that module or the function get minibatch finds nothing.

In the "inference()" function there are two hidden layers and one output layer being defined using the "layer()" function above.

For the output layer, the relu function should be omitted. Could be done by adding a flag to the layer() function (relu=True for example) and just conditionally adding the relu after the linear combination of inputs.
Then create hidden layers with relu, output layer without relu.

the second line of code should look like:
step_choices = range(...)

1. variable "momentum" is not defined, one line could be added:
momentum = random.random()
2. momentum_rand_walk.append(), function append() needs an argument. I guess the code should be:
momentum_rand_walk.append(new_step)

in the paragraph, it says that:
The figure demon‐
strates that to achieve a cost of 0.1 without momentum (right) requires nearly 18,000
steps (minibatches), whereas with momentum (left), we require just over 2,000.

the "left" and "right" are not consistent with what shown in the Figure 4-9.

Missing subscript on the weight for the second moment of the gradient:

Should be $\beta_2^{i-k}$ .

- Incorrect indentations, e.g. for "loss" function
- In the "inference" function the first line should read: init = tf.constant_initializer(value=0)
- The scalar_summary method used in the "training" function module does not exist ("'tensorflow' has no attribute 'scalar_summary'"), should be tf.summary.scalar
- in main, module 'tensorflow' has no attribute 'merge_all_summaries', should be tf.summary.merge_all
- in main, module 'tensorflow.python.training.training' has no attribute 'SummaryWriter'
- in main, variable mnist undefined
etc.

Note that the changes to tensorflow significantly predate the book's publication date

Errors in the mathematical description of Principal Component Analysis:

- "... we can view this operation as a project..." should be "... we can view this operation as a projection..."

- "T=X" should be "T=XW" .

". while little in a you See" should be ". while little a in you See".