AKA: The art of being wrong in front of 700,000 people.
Seems like neural networks are everywhere these days. Google uses them for translation services, chatbots use them to mimic human conversation, investors use them to analyse and predict markets movements, car manufacturers use them to edge ever closer to self-driving vehicles.
But what are neural networks? How do they differ from other, more conventional software structures? And, most importantly, are they intelligent?
These notes, from a talk I presented at the Melbourne International Film Festival in 2017, offer a relatively non-technical look at what's going in inside those little black boxes and how they work.
The talk was first presented in August, 2017 at MIFF Talks: Technology in the Cinema. The event was co-sponsored by Inspire9 and General Assembly.
2. P R O L O G U E
THE ART OF BEING WRONG IN PUBLIC
3. Around the start of 2017, I
published an article on Medium,
enthusiastically raving about
Google Translate’s neural
network.
You can read the full post at
http://bit.ly/gilfew170822, but the
basic gist is that Google’s neural
network developed an entirely
original meta-language as a
means of translating more
effectively.
Which, as you can see from the
excerpts on this page, blew my
mind.
I WROTE A STORY
4. LOTS OF PEOPLE READ IT
Mostly, I’m lucky to get a couple of hundred
views on my Medium posts.
This time, however, the internet did that thing
it does every now and then. It went a bit crazy.
to date, nearly 700,000 people have viewed
my post.
Which is very nice indeed.
Except for one small problem…
5. EVERYTHING I WROTE WAS WRONG
Well, not 100% wrong exactly. But wrong-headed. I was
more-or-less right about what the neural network had done,
but I had vastly overstated the significance of its
achievements.
My post was inundated with comments, explaining that
what Google Translate’s software was doing was perfectly
in line with how neural networks function.
In the end, the comments and conversations responding to
my article turned into some of the most informative and
positive interactions I’ve had in a social media forum. So
that’s a happy ending, but still…
But still… there I was. A technologist talking about
technology and being about as wrong as I could be in front
of seven hundred thousand people.
6. BACK TO SCHOOL
So what wasn’t I understanding when I so enthusiastically
paraded my ignorance to the internet back in January?
Essentially, I was fooled by a simulation. Neural networks
can offer a startlingly good simulation of an intelligance,
decision making and intent.
But it’s just a simulation.
7. WHAT DO WE MEAN BY INTELLIGENT?
I CAN’T DEFINE IT,
BUT I KNOW IT WHEN I SEE IT”
“
U.S Supreme Court Justice Potter Stewart’s famously pithy
summation of the difficulty in defining hardcore pornography,
feels equally apt when talking about intelligence.
8. MATHS WHIZZ?
We often characterise people who have a talented for the
arcane logic and processes of mathematics as having
above average intelligence.
MAYBE
But then again, a five dollar calculator can perform
calculations faster and more accurately than any human
being on the planet. Is your calculator intelligent?
9. QUIZMASTER?
Who was Marco Polo? What is an Alpaca?
Why does Donald Trump?
Knowing lots of stuff is often considered a good indicator of
intelligence. But being able to hold and retrieve data does not
equal intelligence.
The difference between having access to data, and being able to
creatively or intelligently use that tdata is the difference between
Wikipedia and a librarian.
10.
Key Jie, the world’s best human Go player, about to be lose a match against Google’s AlphaGo Software.
IF A MACHINE CAN OUTSMART
A HUMAN, IS IT INTELLIGENT?
11. Complex Rules
Simple Rules
Concealed
State
Exposed
State
GO
CHESS
POKER
GO is a tough game for computers, because it’s
almost impossible to assess the current game state
(winning/losing) at any point in time.
Every piece in Go has the same value as every
other piece, and can be placed on any unoccupied
space on the board.
Chess, perhaps counterintuitively, is
relatively easy. The restricted rules of
movement and varying value of different
pieces means game state is easy to assess,
and the range of available strategies at any
given point are significantly fewer.
12. And this where neural networks come in. Neural
networks are the secret sauce behind Google’s
champion Go-bot and language-inventing
translation software.
In fact, they are the brains of most AI tools at the
moment. And the reason for this is simple:
NETWORKS & NEURONS
NEURAL NETWORKS SIMULATE
THE HUMAN BRAIN’S ABILITY TO
L E A R N A N D A D A P T.
14. • The central processor accesses data and instruction sets stored in
an array of memory locations.
• The processor rertrieves an instruction from memory, along with
any data that the instruction requires., and executes the instruction.
• The processor stores the result in a new memory location, and
continues on to the next instruction speficied in the program.
CONVENTIONAL COMPUTING
15. Neural networks are made up of interconnected
nodes which link together to develop efficient
mathematical models for solving specific problems.
A Neural Network evolves over time.
It Learns.
NEURAL NETWORKS ARE…
DIFFERENT
16. Neural networks are designed to mimic the way brain cells
function.
Through a combination of input and feedback, they can ne trained
to recognize patterns, and make decisions in a humanlike way.
Each neural network consists of
thousands, millions, billions of nodes -
units of functionality whcih mimic
individual neurons.
Data goes in — it is inspected — and then
a result as passed to the output.
Input data is also assigned weight. This
weighting is an indication of the relative
importance each specific piece of data is.
17. X = 2
Y = 3
X > Y : EMIT “1”
OTHERWISE
EMIT “0”
RESULT = 0
DATA
This node takes two inputs and emits either a 1 or a 0.
If input X has a higher value than input Y, it emits a “1”
Otherwise, it emits a “0”.
A NODE DOING ITS THING
18. APPLYING WEIGHT
Here, we’ve taken the same node and fed it the same data
but, because weight has been applied to its inputs, it has
emitted a different value.
WEIGHT = 3
WEIGHT = 1
X = 2
Y = 3
X > Y : EMIT
“1”
OTHERWISE
EMIT “0”
RESULT = 1
X = 6
Y = 3
DATA REFINEMENT
weighted
value
By comparing the accuracy of
previous results, the network can
attempt the quality of its output by
a d j u s t i n g t h e w e i g h t , o r
importance, ascribed to each unit
of input data that flows through
the system.
The amount and direction of
weight to be applied is not the
product of conscious decision-
making. Instead, it is the result of
some relatively standard calculus
equations.
19. APPLYING WEIGHT
WEIGHT = 3
WEIGHT = 1
X = 2
Y = 3
X > Y : EMIT
“1”
OTHERWISE
EMIT “0”
RESULT = 1
X = 6
Y = 3
DATA REFINEMENT
weighted
value
Note that the data itself hasn’t
change. X still has a value of 2, and
Y a value of 3. All that has changed
is the relative importance, or
weight that the model has
assigned eto specific inputs.
X = 2
Y = 3
X > Y : EMIT
“1”
OTHERWISE
EMIT “0”
RESULT = 0
DATA
20. ALL TOGETHER NOW!
A neural network is
just a lot of invidual
nodes, all networked
together, passing
around little subsets
of large data
collections,
evaluating inputs
and emitting outputs.
21. TRAINING THE MODEL
So what do we mean when we say
that neural networks can learn?
This is accomplished by a process
known as training.
To train a neural network, we give it
a large, known dataset to process
and then compare it’s output to the
correct or expected results.
It’s not dissimilar to the way a student
learns maths from school textbook.
Students are given practical exercises to
work on. They write down their answers,
and then compare what they wrote to the
correct answers at the back of the book.
By looking at which ones they got right,
and which they got wrong, students can
refine their approach to solving a
particular type of problem and acheive a
higher success rate next time around.
SOUND FAMILIAR?
22. TRAINING THE MODEL
1. Feed the model lots and lots of
examples of the data containing the
patterns we want it to be able to
recognise, predict or simulate.
23. TRAINING THE MODEL
2. The model processes the
data through layers of
neurons, each making tiny
calculations or assessment
on tiny pieces of the data
3. Because the training data is
a known quantity, the model
can compare it’s own output
against the “expected” or
correct results and determine
its level of accuracy.
24. TRAINING THE MODEL
4. CALCULUS HAPPENS! The model uses some
equations to minutely adjust the weighting of each
input to each node throughout the entire model.
25. TRAINING THE MODEL
5. Start again. The
networks runs through
the same set of
training data, this time
with the newly applied
weightings distributed
throughout the model.
26. TRAINING THE MODEL
6. And ever onwards.
At the end of each iteration, the new
results are compared to the expected
results and the weighting is further
refined, incrementally improving the
accuracy with each new cycle.
27. It’s tempting to think of them as having
autonomous intelligence because:
NEURAL NETWORKS ARE
BLACK BOXES
a) when viewed in the right light from the right
angle, they appear to be making active
decisions.
b) human beings love to anthropomorphise
28. USING A NEURAL NETWORK
TO PITCH AN ORIGINAL,
MIFF-WORTHY FILM.
31. AUGMENT THE DATA
I needed a bigger dataset for the
example, so I grabbed 20,000 lines
of speeches and tweets from Donald
Trump.
In hindsight, I don’t think this helped
to male my network particularly
intelligible.
33. NEURAL NETWORKS
A COUPLE OF SIGNIFICANT LIMITATIONS
Narrow Focus
The Cat Recogniser 2000™ has been expertly trained to correctly identify
photographs containing one or more felines of any age and pedigree, but
it won’t be able to drive your car.
Rote
The same process is repeated over and over again, tweaking variables
according to specific mathematical criteria. The Cat Recogniser 2000™
will never stop after an hour and say “Maybe we’re going about this the
wrong way. What if, instead of trying to recognise cats, we try to
recognise not-cats?”
Incapable intuition or creativity
It’s not just unable to drive your car. The Cat Recogniser 2000™ doesn’t
even have the capacity to want to drive your car. In fact, it doesn’t even
want to pat all those adorable kittens its so skillfully recognised.
34. INTELLIGENT DESIGN EVOLUTIONARY DESIGN
Intentional, planned.
Begins with a plan of what is to be
created, and then figures out how to
make something which resembles
that vision.
Iterative, unopinionated.
Applies a formula or process to
raw data, over and over again, with
a mechanism for arbitrary
refinement at each cycle.
We can’t predict, or dictate, what
the end result will look like.