16:43 2013-11-15
start CalTech machine learning, lec 1
the learning problem
16:46 2013-11-15
mathematical + conceptual + technical
16:47 2013-11-15
What is learning? can we learn?
how to do it? how to do it well?
16:50 2013-11-15
machine learning is a very broad subject
16:51 2013-11-15
mathematical formalization of a learning problem
16:52 2013-11-15
predicting how a viewer rating a movie?
16:54 2013-11-15
in our wide spread of machine learning
16:55 2013-11-15
What is "the essence of machine learning"?
1. a pattern exists
2. we can not pin it down mathematically
3. we have data on it
16:57 2013-11-15
if you have data, we're in business
if you don't, we're out of luck!
16:59 2013-11-15
match, mismatch
17:01 2013-11-15
to produce this thing, you have to watch
the movie, analyze the content, you have to
interview the viewer, and ask about their
taste, and then after that you combine them
and try to get a prediction for the rating
17:02 2013-11-15
the idea of machine learning is that you don't
have to do any of that, you jus sit down and
sip on your tea, while the machine is doing something
to come up with this figure on its own!
17:04 2013-11-15
the learning approach
17:04 2013-11-15
viewer & movie are both vectors!
17:08 2013-11-15
the metaphor we're using is "Credit approval"!
1. Applicant information
17:09 2013-11-15
What are the "components of learning"?
17:11 2013-11-15
first, let's formaliza this
17:11 2013-11-15
Formalization
17:13 2013-11-15
What is a formalization?
1. input: x // customer application
2. output: y // good/bad customer
3. Target function: f: X -> Y // ideal credit approval formula
4. Data: (x1, y1), (x2, y2), ... (xn, yn)
5. Hypothesis: g: X -> Y
17:14 2013-11-15
the target function is always unknown to us!
17:15 2013-11-15
there are plenty of historical records
17:16 2013-11-15
hypothesis lives in the same world as "target function"!
17:17 2013-11-15
unknown target function, // f
training examples,
final hypothesis // g ≈ f
17:19 2013-11-15
What is a learning algorithm?
a learning algorith takes training examples
and produce the final hypothesis!
17:19 2013-11-15
What is a hypothesis set?
a hypothesis set is a set of candicate formulas!
17:20 2013-11-15
hypothesis set
17:22 2013-11-15
there is no downside of including a hypothesis set,
the upside is that the hypothesis plays a pivotal
role in the field of machine learning!
17:26 2013-11-15
What is a Hypothesis Set?
a Hypothesis Set == set of candidate formulas
17:28 2013-11-15
it's a simple model to pinpoint we're talking
about
17:30 2013-11-15
w == weigh
17:32 2013-11-15
What you defines h is your choice of w
& threshold!
17:33 2013-11-15
linearly separable data
17:36 2013-11-15
perceptron
17:37 2013-11-15
we're introduce an artifical coordinates x0 == 1
to simplify our formula!
17:38 2013-11-15
perceptron hypothesis
17:40 2013-11-15
a simple learning algorithm: PLA
PLA == Perceptron Learning Algorithm
17:41 2013-11-15
training set
17:42 2013-11-15
misclassified point
17:43 2013-11-15
weight vector
17:46 2013-11-15
pick a "misclassified point" and update
the weight vector!
w <--- w + yn * xn
17:49 2013-11-15
PLA == Perceptron Learning Algorithm
17:49 2013-11-15
iteration of PLA
17:50 2013-11-15
linearly separable
17:52 2013-11-15
now we have one learning model
17:52 2013-11-15
all you do is trying to match a historical
records
17:53 2013-11-15
PLA == Perceptron Learning Algorithm
17:55 2013-11-15
the basic premise of learning?
using a set of observations to uncover
an underlying process!
17:58 2013-11-15
supervised learning
unsupervised learning
reinforcement learning
18:00 2013-11-15
What is a supervised learning?
supervised learning has (input, correct output)
18:04 2013-11-15
for example: coin recognition
18:04 2013-11-15
What is a unsupervised learning?
in unsupervised learning, instead of
(input, correct output), we get (input, ? )
18:07 2013-11-15
using the coin recgnition example, I can
classify those coin without knowing what they are?
18:16 2013-11-15
What is a reinforcement learning?
instead of (input, correct output), you get
(input, some output, grade for this output)
18:21 2013-11-15
neutral network
18:22 2013-11-15
this is a pretty good deal, this might be
too good to be true!
18:25 2013-11-15
a learning puzzle
18:33 2013-11-15
the target function will still be unknown!
18:34 2013-11-15
Q & A == Question & Answer
18:35 2013-11-15
linear separability
18:35 2013-11-15
PLA == Perceptron Learning Algorithm
18:39 2013-11-15
linear regression
18:47 2013-11-15
statistics can also be viewed as a kind
of machine learning,
find the PDF(Probability Distribution Function)
18:49 2013-11-15
in general, machine learning tries to make
the least assumptions, to accomodate the most
general territory! // compare with statistical inference
18:50 2013-11-15
machine learning, statistical learning
18:52 2013-11-15
optimization is a tool for machine learning
18:54 2013-11-15
hypotheis set
18:55 2013-11-15
even the hypothesis set is huge & complicated!
18:56 2013-11-15
sampling bias
19:01 2013-11-15
you were swimming in data
19:03 2013-11-15
learning is about be able to predict!
19:04 2013-11-15
figure out what the pattern is
19:05 2013-11-15
one of them is validation, after you learn,
you validate your solution
19:06 2013-11-15
perceptron problem
19:07 2013-11-15
PLA == Perceptron Learning Algorithm
19:10 2013-11-15
in generally, the learning algorithm has
the effect of minimizing error
19:11 2013-11-15
perceptron: minimize the classification error!
19:15 2013-11-15
What's the essence of learning?
1. A pattern exists
2. We cannot pin it down mathematically
3. We have data on it
----------------------------------------------------