This video is both an introduction to the recent paper Thinking Fast and Slow in AI by Francesca Rossi and her team at IBM, and to Luis Lamb's most recent paper Neurosymbolic AI: the 3rd Wave of AI.
Both these papers are drawing inspiration from human capabilities to build a future generation of artificial intelligence that would be more general and trustworthy.
Then, there are 10 important questions for the AI community to consider in their research.
Enjoy the video!
Chapters:
- 0:00 Hey! Tap the Thumbs Up button and Subscribe. You'll learn a lot of cool stuff, I promise.
- 5:33 Neuro-Symbolic AI: The Third Wave
- 11:01 Thinking Fast and Slow
- 16:06 Thinking Fast and Slow in AI - 10 Questions for the AI Research Community
References:
- Thinking Fast and Slow by Daniel Kahneman (Book): https://amzn.to/2XHehG1
- Neurosymbolic AI: The 3rd Wave: https://arxiv.org/pdf/2012.05876.pdf
- Thinking Fast And Slow in AI: https://arxiv.org/abs/2010.06002
- AI Debate 2 by Montreal AI: https://youtu.be/VOI3Bb3p4GM
Follow me for more AI content:
- Instagram: https://www.instagram.com/whats_ai/
- LinkedIn: https://www.linkedin.com/in/whats-ai/
- Twitter: https://twitter.com/Whats_AI
- Facebook: https://www.facebook.com/whats.artifi...
- Medium: https://medium.com/@whats_ai
- Youtube: https://www.youtube.com/channel/UCUzGQrN-lyyc0BWTYoJM_Sg/join
Video Transcript
(Note: this transcript is auto-generated by YouTube and may not be 100% accurate)
00:00
okay let me say let me tell you about my
00:03
experience you know in dealing with
00:05
trust and
00:06
ai ethics in general and also what i
00:09
think are some
00:11
main points so some of them are
00:12
technical and some of them are not
00:14
in really achieving this ecosystem of
00:17
trust around the ai so this is the
00:19
overall picture that i think that
00:21
many of the previous panelists put
00:23
forward you know we want the future
00:25
i would say of and with ai because of
00:28
course i don't think as ai
00:30
uh at ai as just autonomous systems but
00:33
also systems that work with us
00:34
so it's not just the future of the eye
00:36
for me but also with ai
00:38
uh and it has all these uh desirable
00:41
properties of course trustworthiness is
00:44
one but of course general collaborative
00:46
and
00:46
as already mentioned for gdp three and
00:48
language very huge launch and matter
00:50
also sustainably
00:51
uh computationally but how to focus on
00:55
this
00:55
third panel how do we build an ecosystem
00:58
of trust
00:58
and i and i talk about an ecosystem of
01:01
trust because
01:02
it has many dimensions just like
01:04
trusting other people as many dimensions
01:07
so of course we want ai system to be
01:09
accurate
01:10
and that's but beyond accuracy we really
01:13
want
01:14
a lot of desirable properties one of
01:16
them i called it and
01:18
some other people call it value
01:19
alignment which is around fairness you
01:22
know what how do we want these
01:25
machines to behave to behave according
01:27
to some values that we care about
01:28
one of them of course is fairness so we
01:31
want the bias to be
01:32
identified removed and so on but
01:35
it may have maybe other values that are
01:37
beyond fairness
01:38
robustness also you know
01:40
generalizability beyond
01:42
you know some data distribution and
01:44
explainability explainability is very
01:46
important especially in the context of
01:48
machines that work together with human
01:50
beings
01:52
now but differently from what we would
01:55
expect in building trust with another
01:56
human being
01:58
here we are not in in the presence of
02:01
another human being we are the presence
02:03
of ai systems that are created by human
02:05
beings
02:06
so just like margaret and others have
02:08
said we want
02:09
something also from those that create
02:12
that ai
02:12
system from the developers from the
02:14
deployers for those that use the ai
02:16
and one of the things that i think
02:18
margaret pointed out very clearly and
02:21
we have a very similar approach is we
02:23
want transparency we want transparency
02:25
about
02:26
the decisions that have been made during
02:28
the ai pipeline
02:30
whether their decision about the
02:31
training data or other decision and
02:34
in that very nice visual way margaret
02:37
showed that bias can be injected in many
02:39
different uh
02:40
places in the ai pipeline uh we don't
02:42
have the concept of a model card but a
02:44
very similar one called the ai faction
02:46
and in fact we work together with
02:48
margaret and others also within the
02:50
partnership and ai
02:51
to compare and learn from each other
02:53
these different ways to achieve
02:55
uh transparency the second point that i
02:58
want to make is that
02:59
of course we want these developers also
03:02
to work according to some ai
03:04
ethics and guidelines and principles but
03:07
principles are just the first step
03:10
in in in a corporate place where
03:13
ai is being produced and deployed so it
03:16
really needs a lot of multi-stakeholder
03:18
consultation
03:19
education training and as margaret
03:22
already mentioned diverse teams
03:24
you know to bring all many different
03:26
backgrounds it needs a lot of technical
03:28
tools
03:29
for example to detect mitigate bias to
03:31
generate explanation and so on
03:33
it needs a lot of work in helping
03:36
developers understand
03:37
how to change the way they're doing
03:39
things how to make it as easy as
03:41
possible to adopt a
03:43
new methodology and how to build an
03:45
overall governance
03:47
in a company within you know that is a
03:50
kind of an
03:50
umbrella over what developers are doing
03:53
what the business units are doing and so
03:55
on so it's really a process
03:57
and that's why i put this picture of
03:58
trust with all the cranes because it's a
04:00
process to build trust in ai
04:04
so the last point that i want to make is
04:06
that uh
04:07
for all these properties that we want in
04:09
the ai systems in order to be able to
04:12
trust them
04:12
unfortunately current ayai is not there
04:15
yet
04:16
these are the reasons why francisca
04:18
rossi and her team at ibm
04:20
published this paper proposing a
04:22
research direction to advance ai
04:24
drawing inspiration from cognitive
04:26
theories of human decision making
04:29
where the premise is if we gain insights
04:32
into human capabilities that are still
04:34
lacking in
04:35
ai such as adaptability robustness
04:38
abstraction generalizability common
04:41
sense and causal reasoning
04:43
we may obtain similar capabilities as we
04:46
have
04:46
in an ai system nobody knows yet what
04:49
will be the future of ai
04:51
will it be neural networks or do we need
04:53
to integrate machine learning with
04:55
symbolic and
04:56
logic based ai techniques the latest
04:59
is similar to the neurosymbolic learning
05:01
systems
05:02
which integrate two fundamental
05:04
phenomena of intelligent behavior
05:06
reasoning and the ability to learn from
05:09
experience
05:10
they argue that a better comprehension
05:12
of how humans have
05:14
and have evolved to obtain these
05:16
advanced capabilities
05:18
can inspire innovative ways to imbue ai
05:21
systems with these competencies
05:23
but nobody will be better placed than
05:25
lewis lamb
05:26
himself to explain this learning system
05:28
shown in his recent paper
05:30
neurosymbolic ai the third wave
05:34
what we want to do here is exactly this
05:36
convergence because one of the key
05:38
questions is to identify
05:40
the building blocks of ai and how to
05:44
make
05:44
a.i more trustworthy ai
05:48
explainable but not only explainable
05:50
interpretable as well
05:52
so in order to make ai interpretable
05:55
sound and to use the right models
05:59
right computational models so that one
06:01
can explain what's going on in ai
06:03
we need better representations we need
06:06
models that are sound
06:07
and soundness and the results that come
06:10
from logic
06:12
the correctness results and all of that
06:14
can benefit
06:15
of course the great results you are
06:17
having on deep learning so
06:19
our work corroborates this point that uh
06:22
gary marcus made and also that danny
06:25
kanema made
06:26
at tripoli i that system one i mean the
06:29
fast system one that's associated with
06:32
concepts like deep learning
06:34
certainly knows language as daniel
06:36
kahneman said and system
06:37
2 which is more reflective certainly
06:40
does involve
06:41
certain manipulation of symbols so this
06:44
analogy of system 1 and 2
06:46
leads us to build the ideas that are
06:50
the inspiration the inspiration that
06:52
gary brought in his book the algebraic
06:54
the algebraic mind and also that we
06:56
formalized in several
06:58
neural symbolic systems since the early
07:01
2000s
07:02
and some of them several of them
07:04
temporal reasoning model reasoning
07:06
reason about knowledge are formalized in
07:08
this book and
07:09
of course we have been evolving this
07:11
concept so that we one can deal with
07:13
combinatory explosion
07:14
and several other symbolic problems
07:17
within a neurosymbolic framework
07:20
and so the approach that we have been
07:22
defending over the years
07:24
is that we need a foundational approach
07:26
for neurosymbolic computing
07:28
neurosymbolic ai
07:29
that's based both on logical
07:31
formalization and we have francesca here
07:34
judy pearl that have been that have been
07:36
outstanding results on symbolic ai
07:39
and machine learning and we use logic
07:42
and knowledge representation to
07:43
represent the reasoning process
07:45
that is integrated with machine learning
07:47
systems
07:48
so that we can also effective
07:50
effectively perform
07:52
neural learning using deep learning
07:54
machinery so our approach has been
07:56
tested in training
07:58
assessment simulators by tno which is a
08:01
dutch subsidiary of the government it
08:03
has been applied in robotics and ai
08:06
and several other applications but what
08:08
we offer here
08:09
is a sound way including some formal
08:12
results
08:13
that our neurosymbolic systems in order
08:16
so that we can have more effective and
08:18
more trustful
08:19
ai we need to have models
08:23
interpretive models that are based on
08:25
sound logical models
08:27
and in this way we can explain what the
08:29
neural learning process is doing
08:32
at the same way that we can prove that
08:35
the results that we obtain via machine
08:38
learning
08:39
can even be related to the formal
08:41
results that one typically expects from
08:43
symbolic logic for instance here
08:46
in a system that we call connections
08:47
connectionist modal logic
08:49
which was by the way published in the
08:51
same issue of neuro computation that
08:54
jeff hinton published one of his
08:56
inflation paper
08:57
on deep belief nets we proved that model
09:01
and temporal logic programs
09:02
can be computed soundly in neural
09:06
network models so in this way what we
09:08
provide
09:09
in a way is a way of providing neural
09:12
networks
09:12
as a learning system which can also
09:15
learn to compute
09:16
in a deep way the evolution of knowledge
09:19
in time
09:20
and this is what we explained in several
09:23
of our papers
09:24
and also in recent work that we
09:26
published gary in
09:27
um in tripoli 2018
09:31
and now each guy 2020 where we present a
09:34
survey paper
09:36
so the the the final message here
09:39
is that there have been some
09:41
developments including
09:43
uh the ai debate the great ai debate
09:46
between banjo and
09:47
gary marcus last year which we saw also
09:50
at triple i 2020 that we need more
09:52
convergency
09:53
towards building more effective ai
09:56
systems
09:57
and ai systems that most people can
09:59
trust since ai
10:00
is becoming a lingua franca for science
10:03
these days
10:04
neurosymbolic is basically another type
10:06
of ai system that is trying to use a
10:09
well-funded knowledge representation
10:11
and reasoning it integrates neural
10:13
network-based learning
10:15
with symbolic knowledge representation
10:17
and logical reasoning
10:18
with the goal of creating an ai system
10:20
that is both interpretable
10:22
and trustworthy this is where francesca
10:25
rossi's work comes into play
10:27
with their purpose called thinking fast
10:29
and slow in ai
10:30
as the name suggests they focus on
10:32
daniel kahneman's theory regarding the
10:35
two systems
10:36
explained in his book thinking fast and
10:38
slow
10:39
an attempt to connect them into a
10:41
unified theory
10:43
with the aim to identify some of the
10:45
roots of the desired human capabilities
10:48
here is a quick extract taken from the
10:50
ai debate 2
10:51
organized by montreal ai where daniel
10:54
cantman himself clearly explained these
10:56
two systems
10:58
and their link with artificial
10:59
intelligence
11:01
i seem to be identified with the idea of
11:04
two systems system one and system two
11:06
although they're not my idea but i did
11:08
write a book
11:09
that described them and
11:12
as quite a few of you sure know
11:16
we talk about the contrast between one
11:18
system that works fast
11:20
another that works slow uh
11:23
but the main difference between system
11:25
one and system two as i described them
11:27
was that system one is something that
11:30
happens to you
11:32
you are not active the thought that
11:35
the words come to you the ideas come to
11:38
you the emotions come to you they happen
11:40
to you
11:40
you do not do them and the essential
11:43
distinction that i was drawing
11:45
between the two systems was really that
11:47
one
11:48
something that happens to you and
11:50
something that you do
11:52
high level skills in my description of
11:55
of things
11:56
were absolutely in system one anything
11:58
that they can do automatically
12:00
anything that happens associatively is
12:03
in system one
12:06
another distinction between system one
12:08
and system two
12:09
as psychologists see them in that
12:12
operation the system one tend to be
12:14
parallel
12:14
operations or system two tend to be
12:16
serial
12:18
so it's true that anything
12:21
any activity that we would describe as
12:23
non-symbolic
12:25
i think does belong to system one
12:28
that system one i think cannot be
12:30
described
12:31
as a non-symbolic system for one thing
12:35
uh it's it's much too complicated and
12:39
rich for that
12:40
it knows language for one thing
12:42
intuitive thoughts
12:44
are in language uh the most interesting
12:47
component
12:48
of system one the basic component as i
12:51
conceive of that
12:52
notion is that it holds a representation
12:54
of the world
12:56
and and the representation that actually
12:59
allows something that resembles the
13:01
simulation of the world
13:03
as i describe it we we live
13:06
with that representation of the world
13:08
and most of the time
13:10
we are in what i call the valley of the
13:12
normal
13:13
there are events that we positively
13:15
expect
13:16
there are events that surprise us but
13:19
most of what happens to us
13:21
neither surprises us nor is it expected
13:24
what i'm going to do
13:25
to say next will not surprise you but
13:28
you didn't actually expect it
13:30
so there is that model that compares
13:33
that
13:34
accepts many many
13:38
events as normal continuations of what
13:41
happens
13:41
but it rejects some and it distinguishes
13:45
what is surprising from what is normal
13:51
that's very difficult to describe in
13:54
terms of symbolic or non-symbolic
13:56
certainly what happens is a lot of
13:58
counter factual thinking
14:00
is in fact system one thinking because
14:04
surprise is something that happens
14:07
automatically
14:08
you're surprised when something that
14:10
happens
14:11
is not normal is not expected
14:14
and that forces common sense
14:17
and causality to be in system one and
14:20
not in system two
14:22
in short can man explains that humans
14:25
decisions
14:26
are guided by two main kinds of
14:28
capabilities
14:29
or systems which he referred to as
14:32
system 1 and system 2.
14:34
the first provides us tools for
14:36
intuitive fast and unconscious
14:38
decisions which could be viewed as
14:41
thinking
14:41
fast while the second system handles
14:44
complex situations
14:46
where we need rational thinking and
14:48
logic to make a decision
14:50
here viewed as thinking slow
14:53
if we come back to the thinking fast and
14:55
slow in ai paper
14:56
scratch frachanska russi and her team
14:58
argue that we can make a very loose
15:00
comparison
15:01
between these two systems one and two
15:04
and the two main lines of work in ai
15:06
which are machine learning and symbolic
15:09
logic reasoning
15:10
or rather data-driven versus knowledge
15:13
driven ai
15:14
systems where the comparison between
15:17
kanman's
15:18
system one and machine learning is that
15:20
both seem to be able to build
15:22
models from sensory data such as
15:25
seeing and reading where both system 1
15:28
and machine learning produce
15:29
possibly imprecise and biased results
15:32
indeed what we call deep learning is
15:35
actually not
15:36
deep enough to be explainable similarly
15:38
to system one
15:39
however the main difference is that
15:41
current machine learning algorithms
15:43
lack basic notions of causality and
15:46
common sense
15:47
reasoning compared to our system one
15:50
we can also see a comparison between the
15:52
system 2 and ai techniques
15:54
based on logic search optimization and
15:57
planning
15:58
techniques that are not using deep
16:00
learning rather employing explicit
16:02
knowledge
16:03
symbols and high level concepts to make
16:05
decisions
16:06
this is the similarity highlighted
16:08
between the humans decision making
16:10
system
16:11
and current artificial intelligence
16:13
systems i want to remind you that as
16:15
they state
16:16
the goal of this paper is mainly to
16:18
stimulate the ai research community to
16:20
define
16:21
try and evaluate new methodologies
16:24
frameworks and evaluation metrics in the
16:27
spirit of achieving a better
16:29
understanding
16:30
of both human and machine intelligence
16:33
they intend to do that by asking the ai
16:36
community to study
16:37
10 important questions and try to find
16:40
appropriate answers or at least
16:42
think about these questions here i will
16:44
only quickly list
16:45
these 10 important questions to be
16:47
considered in future research
16:50
but feel free to read their paper for
16:52
much more information regarding these
16:54
questions
16:54
and discuss it in the comments below so
16:57
here it goes
16:58
should we clearly identify the ai system
17:01
1 and system 2 capabilities
17:04
is the sequentiality of system 2 a bug
17:06
or a feature
17:07
should we carry it over to machines or
17:09
should we exploit parallel threads
17:12
performing system 2 reasoning will this
17:14
together with the greater computing
17:16
power of machines compared to humans
17:19
compensate the lack of other
17:20
capabilities in ai
17:23
what are the metrics to evaluate the
17:25
quality of a hybrid system 1 system 2
17:28
ai system should these matrix be
17:31
different
17:32
for different tasks and combination
17:34
approaches
17:35
how do we define ai's introspection in
17:38
terms of eye consciousness and
17:40
m consciousness how do we model the
17:43
governance of system 1 and
17:45
system 2 in an ai when do we switch or
17:48
combine them
17:49
which factors trigger the switch how can
17:53
we leverage a model based on system 1
17:55
and system 2 in ai to understand and
17:58
reason
17:58
in complex environments when we have
18:01
competing priorities
18:03
which capabilities are needed to perform
18:05
various forms of moral judging and
18:08
decision making
18:09
how do we model and deploy possibly
18:12
conflicting normative ethical theories
18:14
in ai
18:15
are the various ethics theories tied to
18:18
either system 1 or system 2
18:21
how do we know what to forget from the
18:23
input data during the abstraction step
18:26
should we keep knowledge at various
18:28
levels of abstraction
18:30
or just raw data and fully explicit
18:32
high-level knowledge
18:34
in a multi-agent view of several ai
18:37
systems communicating and learning from
18:39
each other
18:40
how to exploit adapt current results on
18:43
epistemic reasoning
18:44
and planning to build learn models of
18:47
the world and of others
18:49
and finally what architectural choices
18:52
best support the above vision of the
18:55
future of ai
18:58
feel free to discuss any of these
18:59
questions in the comments below
19:01
i would love to have your take on these
19:03
questions and debate over them
19:05
i definitely invite you to read the
19:07
thinking fast and slow in ai paper
19:10
as well as daniel canman's book thinking
19:13
fast and slow if you'd like to have more
19:15
information about this
19:17
theory if this subject interests you i
19:20
would also strongly recommend you to
19:22
follow the research of yeshua benjo
19:24
addressing consciousness priors and a
19:27
huge thanks to montreal ai for
19:29
organizing
19:29
this ai debate 2 providing a lot of
19:32
valuable information
19:33
for the ai community all the documents
19:36
discussed
19:37
in this video are linked in the
19:38
description below please
19:40
leave a like if you went this far in the
19:42
video and since there are over 90
19:44
of you guys watching that are not
19:46
subscribed yet consider subscribing to
19:48
the channel to not miss any further news
19:50
clearly explained
19:52
thank you for watching
Learn AI:
Join Our Discord channel, Learn AI Together:
►https://discord.gg/learnaitogether
Song credit: https://soundcloud.com/mattis-rodrigu...
►https://discord.gg/learnaitogether
Song credit: https://soundcloud.com/mattis-rodrigu...