In machine learning, we are given a dataset of the form {(xi, yi)}, i ∈ [1..N] and aim at learning a function f(x) which maps unseen input feature vectors to ŷ - the predicted value. Distinguish between the 3 types of learning problems by characterizing the mathematical type of the predicted values ŷ:
Classification:
Regression:
Ranking:
Given a training dataset D = {(xi, yi)}, i ∈ [1..N], we want to identify a function fΘ() such that the predictions ŷ = fΘ(x) over the training dataset are as accurate as possible. Given a Loss Function L(y,ŷ) - write the criterion that the optimal value of Θ must satisfy:
Find θ such that:
Write the expression of the cross-entropy loss which is useful when the predicted output of the model we learn is interpreted
as a discrete distribution p(yc|x) for c ∈ [1..C] (C-way classification model).
f(x) = ŷ = (ŷ1 ... ŷC) is a distribution over the C possible classes.
L(ŷ,y) =
The deep learning approach learns a trainable non-linear mapping function φ from x to a representation φ(x) which can be used as
an input to a linearly separable classification problem. The general form of this trainable mapping we consider is:
ŷ = W φ(x) + b
φ(x) = g(W'x + b')
where g is a non-linear function.
Why do we need non-linear mappings such as g() in this formulation?
Consider the task of predicting the sentiment of a text document as either {positive, negative, neutral}.
We want to use a neural network to learn a model for this task, given a training dataset of the form { (documenti, labeli) } i in [1..N].
Each document di contains Ni words (wi,1, ..., wi, Ni), where the words wi,j belong to the
vocabulary V = { w1, ..., w|V| }.
Describe how the documents are encoded as vectors of size |V| for each of the following two methods: