1
00:00:00,000 --> 00:00:05,072
Let's return, for a moment, to our hidden
agenda, of trying to understand something

2
00:00:05,072 --> 00:00:09,080
about intelligence from all the stuff that
we've learned so far.

3
00:00:09,660 --> 00:00:13,880
We figured out that classes can be learned
from experience.

4
00:00:14,360 --> 00:00:18,035
Features can also be learned from
experience.

5
00:00:18,035 --> 00:00:22,528
For example genres,
That is classes as, as well as roles which

6
00:00:22,528 --> 00:00:27,920
can maybe a features merely from the
experience of people buying books.

7
00:00:30,620 --> 00:00:36,323
What is the minimum capability needed to
learn features and classes directly from

8
00:00:36,323 --> 00:00:39,945
data?
This is a, rather carefully thought out

9
00:00:39,945 --> 00:00:43,180
question,
So let me think about it for a minute.

10
00:00:44,860 --> 00:00:49,464
The first stage one needs some low level
of perception,

11
00:00:49,464 --> 00:00:56,245
One needs to be able to perceive in the
case of humans, pixels and frequencies or

12
00:00:56,245 --> 00:01:03,026
in the case of our systems, one needs to
know, be able to identify the person by a

13
00:01:03,026 --> 00:01:07,380
person ID, the book by a book ID and
that's about it.

14
00:01:08,180 --> 00:01:14,423
Second, one needs the ability to subitize
which is another way of saying counting or

15
00:01:14,423 --> 00:01:20,742
distinguishing between one and two things.
So it turns out that very young babies are

16
00:01:20,742 --> 00:01:25,350
actually able to distinguish between one
person or two people,

17
00:01:25,573 --> 00:01:31,074
One object or two objects and they get
surprised when suddenly one object

18
00:01:31,074 --> 00:01:35,980
disappears from the scene,
So this is essentially something innate.

19
00:01:36,340 --> 00:01:42,469
Similarly, the ability to break up
temporal experiences that one experiences

20
00:01:42,469 --> 00:01:47,873
over time into episodes that they
experienced something in the past five

21
00:01:47,873 --> 00:01:53,427
minutes and then the next ten minutes
another experience because suddenly,the

22
00:01:53,437 --> 00:01:57,873
scene has changed.
To break up this episodes is another

23
00:01:57,873 --> 00:02:03,600
subitizing feature in time, which babies
learn at a slightly later age.

24
00:02:04,740 --> 00:02:11,036
Given these two things, and our hidden
latent model techniques, one can

25
00:02:11,036 --> 00:02:18,412
essentially in principle learn classes and
features together simply from the fact

26
00:02:18,412 --> 00:02:22,280
that they co-occur together in
experiences.

27
00:02:22,760 --> 00:02:28,777
Theoretically it works,
But in practice, lots of research is

28
00:02:28,777 --> 00:02:35,370
currently underway to enable machines to
learn in an unsupervised manner, both the

29
00:02:35,370 --> 00:02:40,764
classes as well as the features.
So you're clustering the classes, you're

30
00:02:40,764 --> 00:02:46,608
clustering the features side-by-side,
using the fact that classes and features

31
00:02:46,608 --> 00:02:52,211
co-occur in different experiences or
objects to learn both together.

32
00:02:52,211 --> 00:02:59,001
So these is really at the frontier of
research today in both from web

33
00:02:59,001 --> 00:03:06,470
intelligence as well as understanding
human intelligence, to a certain extent.

34
00:03:06,470 --> 00:03:13,025
So when you come across articles which
talk about bottom up learning or grounded

35
00:03:13,025 --> 00:03:19,661
techniques, essentially they're talking
about things like this where one is trying

36
00:03:19,661 --> 00:03:24,760
to learn a hidden or latent model directly
without supervision.

37
00:03:24,760 --> 00:03:31,485
Of course one might really ask, to what
extent have we actually learned anything

38
00:03:31,485 --> 00:03:37,766
in the true or pure sense of the word.
In fact, in a rather celebrated 1980

39
00:03:37,766 --> 00:03:43,939
article the Philosopher, Philosopher John
Searle refuted any suggestion that

40
00:03:43,939 --> 00:03:50,187
mechanical reasoning using all manner of
learned facts or, Or, or rules could

41
00:03:50,187 --> 00:03:57,021
actually be considered intelligent and
this is the argument he used reminiscent

42
00:03:57,021 --> 00:04:03,000
of the Turing test, in fact.
Sir imagined a room where a person,

43
00:04:03,000 --> 00:04:08,724
In this case, it's a bird,
But he imagined a person armed with rules

44
00:04:08,724 --> 00:04:15,386
and facts and reasoning techniques that
allowed that person to translate from

45
00:04:15,386 --> 00:04:23,741
Chinese to English using mechanical
calculations and facts and rules about how

46
00:04:23,741 --> 00:04:28,701
to translate.
The question, so the last was does the

47
00:04:28,701 --> 00:04:35,540
translator know Chinese in the sense that
a native speaker of Chinese knows Chinese.

48
00:04:36,360 --> 00:04:44,237
He argued vehemently that this person
could not in any sense be construed to

49
00:04:44,237 --> 00:04:52,438
know Chinese, and therefore the prospect
of machine intelligence divorced from any

50
00:04:52,438 --> 00:04:59,468
direct perception or abilities in,
Say the language, Chinese was actually a

51
00:04:59,468 --> 00:05:03,630
fallacy.
Interestingly, exactly such techniques

52
00:05:03,630 --> 00:05:11,139
such as we have discussed here as well as
new ones that we'll talk about next time

53
00:05:11,139 --> 00:05:18,376
like hidden Markov models, are actually
used to parse and translate Chinese into

54
00:05:18,376 --> 00:05:22,538
English fairly well today in Google
Translate.

55
00:05:22,538 --> 00:05:29,890
The question posed by Sir has now become
popularly and known as the Chinese room

56
00:05:29,890 --> 00:05:36,104
debate and is certainly worth pondering
about, when one talks about web

57
00:05:36,104 --> 00:05:41,800
intelligence as having taught us anything
about our own abilities.