1
00:00:00,000 --> 00:00:04,995
In this final video, I was tempted to 
make some predictions about the future of 

2
00:00:04,995 --> 00:00:09,295
research on neural networks. 
Instead, I'm going to explain to you why 

3
00:00:09,295 --> 00:00:13,785
it would be extremely foolish to try to 
make any long term predictions. 

4
00:00:13,785 --> 00:00:19,034
I'm going to try and explain why we can't 
predict the long term future by using an 

5
00:00:19,034 --> 00:00:22,560
analogy. 
Imagine you're driving a car at night, 

6
00:00:22,560 --> 00:00:27,160
and you're looking at the taillights of 
the car in front. 

7
00:00:27,160 --> 00:00:31,696
The number of photons that you receive 
from the taillights of the car in front 

8
00:00:31,696 --> 00:00:36,460
falls off at one over d squared, where d 
is the distance to the car in front. 

9
00:00:36,460 --> 00:00:42,180
That's assuming that the air is clear but 
now suppose there's fog. 

10
00:00:42,180 --> 00:00:47,603
Over short ranges, the number of photons 
you get from the tail lights in front of 

11
00:00:47,603 --> 00:00:52,224
you, still falls off as one over d^2. 
Because over a short range, the fog 

12
00:00:52,224 --> 00:00:56,576
hardly absorbs any light. 
But for large distances, it falls off as 

13
00:00:56,576 --> 00:01:00,125
E to the -D. 
And that's because fog has an exponential 

14
00:01:00,125 --> 00:01:03,138
effect. 
Fog absorbs a certain fraction of the 

15
00:01:03,138 --> 00:01:07,780
photons per unit distance. 
So for small distances, fog looks very 

16
00:01:07,780 --> 00:01:12,120
transparent, but for large distances, it 
looks very opaque. 

17
00:01:12,120 --> 00:01:16,639
So, the car in front of us becomes 
completely invisible at a distance at 

18
00:01:16,639 --> 00:01:21,597
which our short range model, the one of a 
discreet model, predicts it will be very 

19
00:01:21,597 --> 00:01:25,276
visible. 
That causes people to drive into the back 

20
00:01:25,276 --> 00:01:27,400
of cars in fog. 
It kills people. 

21
00:01:27,400 --> 00:01:31,510
The development of technology is also 
typically exponential. 

22
00:01:31,510 --> 00:01:35,103
So over the short term, things appear to 
change fairly slowly. 

23
00:01:35,103 --> 00:01:39,463
And it's easy to predict progress. 
All of us, for example, can probably make 

24
00:01:39,463 --> 00:01:42,762
quite good guesses about what will be in 
the iPhone six. 

25
00:01:42,762 --> 00:01:47,358
But in the longer run, our perception of 
the future hits a wall, just like with 

26
00:01:47,358 --> 00:01:49,656
fog. 
So the long term future of machine 

27
00:01:49,656 --> 00:01:52,601
learning in neural nets is really a total 
mystery. 

28
00:01:52,601 --> 00:01:55,960
We have no idea what's going to happen in 
30 years' time. 

29
00:01:55,960 --> 00:01:59,452
There's just no way to predict it from 
what we know now. 

30
00:01:59,452 --> 00:02:02,259
Because we're going to get exponential 
progress. 

31
00:02:02,259 --> 00:02:07,311
In the short run however, in a period of 
say three to ten years, we can predict it 

32
00:02:07,311 --> 00:02:10,552
fairly well. 
And it seems obvious to me that over the 

33
00:02:10,552 --> 00:02:14,509
next five years or so. 
Big deep neural networks, you are going 

34
00:02:14,509 --> 00:02:18,233
to do amazing things. 
I'd like to congratulate all of you who 

35
00:02:18,233 --> 00:02:23,254
stuck it out long enough to get this far. 
I hope you've enjoyed the course and good 

36
00:02:23,254 --> 00:02:24,822
luck with the final test.