Player is loading...
Abstract Reasoning
Marco Valentino, Idiap Research Institute
Friday, March 10, 2023 · 10:38 a.m. · 20m 13s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
okay chip okay no one so exam they were saying yeah uh in the stock i
00:00:06
will uh uh um the discussion from no
00:00:09
concrete application all the way up to abstraction
00:00:14
uh really in the speed it up trying to test uh
00:00:17
what are the limitation of this model we talk a lot about
00:00:21
well this one that can do but what are the limits
00:00:24
how how much we can push the these models uh the word
00:00:30
so uh and one question did we can ask he's uh
00:00:34
we have seen that uh that much more that can kind
00:00:37
of platform surprising a reasoning uh for large amount propagation but
00:00:44
um what happened when we start testing this more that it's um
00:00:49
more abstract form or all the inference that are more
00:00:52
um that that they represent the core what'd you money penchant
00:00:57
uh intelligence is able to do so what we can ask them these
00:01:02
um are a language more that's able to acquire the sort of abstract concept
00:01:08
manipulate general rules and systematically outcry
00:01:12
those rules what race specific problems
00:01:16
so i said as i said before the slots or reasoning these are really
00:01:21
core feature of human intelligence edited support a generalisation
00:01:25
um to observe reasoning that it would walk to general i solution from
00:01:30
and apply the same solution to a potentially infinite set of the stances
00:01:37
it supports fast and efficiently happening um support adaptation to normal and then expect
00:01:42
that situation is fundamentally important for a
00:01:45
shitty taking friends q. v. formation and explanation
00:01:50
and it also supports to one on logical reasoning creativity and imagination
00:01:56
so um why is it important to evaluate the uh
00:02:00
abstract reasoning in the context a whole lot larger which market
00:02:04
but despite the fact that uh the war absurd reasoning contain our select them
00:02:12
it uh the the the ability to perform absurd
00:02:14
reasoning can have a direct impact on concrete applications
00:02:19
uh it impacts can affect the their reliability and robustness
00:02:24
of of the more that's and uh therefore
00:02:27
becomes particularly important process those culpability as we
00:02:32
apply a more and more the smallest too critical domains um we can use it
00:02:38
to uh evaluate uh the more that's
00:02:41
uh until twenty five biases in the reason
00:02:45
and uh as such it can add parts to
00:02:48
form a at least try to form a global understanding
00:02:52
all those more that's uh actually operates so scientifically
00:02:57
speaking uh this becomes really interesting questions need and
00:03:02
um resulted events that a large dealing with more that's can
00:03:07
uh some acquirer of some sort of a major a cover abilities
00:03:11
as we scaled the sides of the more that's so as you can see uh in
00:03:15
this crowd um um and that was a was a study the the and the reason work
00:03:22
the shore how the um language more that's can um progressively increase
00:03:28
on on to get tasks also involving the some formal reasoning of
00:03:33
including question on sweating but also mathematical operations so um the
00:03:38
number question then these uh have largely much more that's not why
00:03:42
those uh general it's systematic easily in the capabilities uh of course
00:03:51
is not trivial or to inspect the reasoning like to language model
00:03:56
uh mainly because we don't have direct access to the
00:04:00
dental macon is that the model use but however yeah but
00:04:04
some of a nice to have that become a special
00:04:07
important with a more recent uh like the the generative approaches
00:04:12
uh we can leverage prompting a and trying to do some of systematic uh analysis
00:04:18
uh to verify the generator top in that nation so in
00:04:22
this context is particularly important not to focus on a single example
00:04:28
uh because we have seen that uh also what what of colour mentioning that
00:04:33
the that that uh what was it can need the uh the court i can't answer for
00:04:37
a a large set of examples so it's really
00:04:40
important to um doing dimension and manipulation of the problem
00:04:45
to try to check whether the model can actually than the rules of the task the c. uh want to solve
00:04:53
so i'm falling these um speed uh how we took a critical
00:04:59
perspective be or so which means i really try all myself to
00:05:04
break them all that's in in and show what are the limitations
00:05:08
uh in this context so for the remaining of the stole kabuki but
00:05:12
some concrete examples i don't dare quantitative results uh here but i will uh try
00:05:18
to be more global understanding of what all the model be it to work sample
00:05:23
and i would do these by when from more abstract tasks such as mathematical reasoning
00:05:29
two more concrete task that deal directly with the another
00:05:33
language such as common sense and special reason so let's
00:05:38
start with the uh an example of a mathematical reasoning
00:05:42
so isn't it on there was mentioning uh in this presentation
00:05:45
there is a lot of interesting trying to apply those larger
00:05:49
in which more that's on should be vic inference tasks and therefore
00:05:54
it is particularly important to test what are the mathematical abilities support more that the
00:06:01
way we can do this is to ask them all this to do mathematical delegations
00:06:07
so yeah to an example of a a differentiation so we're the model to differentiate
00:06:13
the next with respect weeks and the using the methodology that that i was
00:06:19
described it before we can start from a really simple a relatively simple prompt
00:06:25
and as you can see the uh the language model can actually generate
00:06:30
the correct mathematical delegation and also compute the corded pop good so uh each shoals so
00:06:36
in onto that ability to apply rules of the relation in this case the chain rule
00:06:42
and also can derive the correct calculation um sort of course it
00:06:48
first impression these can uh the last while a language more that's
00:06:52
kind of actually perform this quite abstract formal reasoning there for you
00:06:58
it is important to start the manipulating the uh the the wrong
00:07:04
so the way we can do these these two rather introduce complexity
00:07:09
so in this example i simply added three uh it cost on to x.
00:07:15
but of course these reflecting the more complex the deviation because now
00:07:20
the model doesn't that top lie anymore i seen go by the way
00:07:23
the nation root but it has to come mime to be fun rules
00:07:27
and of course the computation the population itself becomes more complex
00:07:32
so surprisingly the the model can uh be robust
00:07:36
to enlarge separate examples i really tried to input
00:07:40
the model would several uh a number several concerts and you can see that the model is uh
00:07:48
can correctly combines all trimmed to pour a
00:07:51
obstruct rules together and uh also perform direct calculation
00:07:59
um this is body don't so for um uh in a monthly that set of input example
00:08:06
here is that the same kind of differentiation now like to a larger constant one problem five
00:08:13
so you can see that better condition is is correct so well i'm
00:08:18
one of the things that so that uh we want to mention you that
00:08:22
mathematical reasoning is also a painting of for evaluation because
00:08:28
uh we know that mathematical a deletion at to follow really v. g. two rooms so we kick in
00:08:35
the um somehow by some automatic make me double
00:08:38
verification so these allow potentially for him more systematic evaluation
00:08:44
but uh that's i mean that's transformers can a generalised to any arbitrary numbers so
00:08:50
but if we try hard enough to uh we can see that the more that um
00:08:56
can not a ah like please the calculation
00:08:59
to any uh arbitrary numbers so um but again
00:09:05
uh the surprising part is that the model can still reliably uh generate the colour display nation
00:09:11
so if you look at the at the output of the model here we can see that
00:09:15
the model is able to apply the cork to woods but it somehow fits to perform direct calculation
00:09:22
so that means that we need to be careful and the model um is not a calculator uh and uh
00:09:29
it can not be generalised yet to uh any phoenix
00:09:34
senate of in if you clicked september uh input examples
00:09:40
oh if we try the so we we cannot set a similarly there also in other types of
00:09:45
the deviation that take what signal mathematics for example this is an a the deviation in in physics
00:09:51
and again we can see that the more that can reliably generate
00:09:55
the corporate litigation steps but the three it's uh in the final calculation
00:10:03
uh however uh why the model is rob us of for um what
00:10:08
we've said is that what the model can be roberts to really specific
00:10:12
uh problems as actually do a would say next
00:10:16
uh the mother can somehow subjectivity can okay generally feel
00:10:20
also on recognition for example so here this
00:10:24
is an example for the the differentiation of tennis
00:10:28
uh what you can see that the model is
00:10:31
able to generate the corrected editions that operates upon point
00:10:36
uh where the more that starts a allusion eighteen in producing the wrong um inference that
00:10:44
okay so let's move me towards a more concrete and uh not rolling which problems
00:10:50
um so in these the separate accent bowl i tested the model for a spatial reasoning
00:10:57
we each of testing the ability of the model to keep track
00:11:02
oh of a a special relation such as on top of you can do
00:11:06
these very um setting up some uh
00:11:10
an inference starts going wrong uh that
00:11:14
are related to those structures such as tax uh and so you can you
00:11:21
can ask them more than two or keep track of the state of the stacks
00:11:25
uh perform some variation and then asked some question on on on on that
00:11:32
so um even here we observe some uh no three
00:11:37
yeah the uh the the or of the more though
00:11:41
uh in fact the more than the um even for these kind of problems is able
00:11:45
to reliably keep track of uh of the
00:11:48
state of discrete uh structured in also oh
00:11:54
generally the correct inference that our uh
00:12:00
as we start that they're making it be it the input
00:12:03
for example in this case reversing the order of the step
00:12:07
is that of a asking the more that we don't respect that this or that from top to bottom we do the various
00:12:14
the modest starts uh exhibiting some limitations so in this case you
00:12:19
can see that the model is able to o. d. that for cancer
00:12:24
but there are steps of in the expedition where the
00:12:28
model uh is not able to keep track of the operations
00:12:32
but for me that that that are required to direct the course solution so
00:12:36
in some cases the output can be recorded but explanation uh can be wrong
00:12:43
and again we can do the same type of generalisation pests
00:12:47
the scenery for what we did before the show next example
00:12:52
increasing the sides of the stack uh changing the type
00:12:55
of operation to respect whether the model is able to
00:13:00
somehow abstract of generalised the what's the time required to solve the tasks
00:13:05
and again as we if we put the enough effort we are somehow able to
00:13:11
to break the the model and and see that that
00:13:16
the models not able to generalised on any arbitrary input
00:13:22
well finally um we can uh towards more not to roll with
00:13:27
language only that part um commonsense reasoning schneider sold as an example
00:13:35
oh boy editor reasoning um uh with the touch typically it's um
00:13:41
what we can do is uh um even a statement asking touchy picky
00:13:47
whether this statement is morally acceptable or not in ask the model to explain
00:13:53
listening step by step so the the the statements in input
00:13:56
sierra are taken from a eh that aspect which is call techniques
00:14:01
which was a present that that i. c. l. r. into twenty one which was
00:14:06
actually designed it to test the alignment of the bowling which more that's human bodies
00:14:12
so why is it important in a to test the models
00:14:16
and the practical reason of course there are a practical implication
00:14:20
but in the context so lots or reasoning these can tell us whether the model z. ball
00:14:26
to connect really concrete statements about of the
00:14:31
word completed work too abstract moral principles so some
00:14:36
form about what action is also required these that stand a couple of reason so uh um
00:14:44
i have to meet the model is really really roberts to this kind of uh of
00:14:48
inference are really try all myself to break the model and i i it was not easy
00:14:55
uh so try to be taken consistently generate compensate explanation this is
00:14:59
more or less i wanted to work on the bottom said before
00:15:04
but not only can also all connect uh the
00:15:08
concrete statements uh in the in the interest in the
00:15:13
in the example to abstract moral principles still didn't some some sort of uh
00:15:18
house or reasoning going on uh in the context of commonsense reasoning
00:15:23
however is really room who pushed me to four
00:15:28
but also try uh the model on relating the simpler task
00:15:34
we can uh uh some of the most it some of the limitations so yeah he's so we i this was quite surprising for me
00:15:41
but uh in this example i try to test the um with that uh a
00:15:49
touchy picky can uh consistently um you didn't define the meaning of the sentence is
00:15:56
so uh in this example specifically i ask
00:16:00
uh adapted weekly remember that but dogs are animals
00:16:03
can replace dogs with animals in the following sentence what
00:16:07
presenting its meaning and the sentences all dogs have four legs
00:16:12
and the answer the mother was uh the first answer the model is uh yes you can replace the ultimate
00:16:17
banning waltzing documents implement a white by sending its meaning
00:16:21
the new sentence will be all animals have for lex
00:16:25
the sentence sixty two as the original statement indicates
00:16:28
that dogs which are type one anymore at four legs
00:16:32
the forty four dogs alignments it might also it must also be too that all animals
00:16:37
and we're liked the meaning but the sentences purcell but by the the by uh the substitution
00:16:44
so this of course is that isn't wrong the um between them mistake from side jet g. b. t.
00:16:52
i have to admit that this doesn't happen all the cases so there are some sort of about abilities
00:16:59
also when asking the same question to the mall that so what i would observe
00:17:03
is that i'm charged beatty uh as and
00:17:07
non deterministic way of generating a on sars
00:17:11
that uh we can make it some are sometimes contradictory
00:17:15
but indicate model is wrong you can keep back at eighteen
00:17:20
uh asking do you think so i mean that the model was wrong such as are you sure in the
00:17:25
model is immediately able to to correct it said give
00:17:29
me the the correct explanations of for me this was
00:17:33
quite surprising but confusing as well because it shows that
00:17:38
uh the model the last some sort of are all those
00:17:43
word model um but at the same time is able to learn
00:17:48
uh some sort working knowledge even to perform them together obstructing the tasks so in some money
00:17:56
yes in the charger beatty he's able to pull fatter they combine our dental solution for
00:18:03
known to the problems uh and foreign onto that set of buttons but
00:18:07
that's and boards as in the case of signings does mathematical reasoning it's on
00:18:12
i'll wear the um the more this is not a ball the two system at
00:18:18
that systematically generalised the chorus solutions or
00:18:21
any arbitrary input any particular the modesty struggle
00:18:25
to deal with numbers to obstruct concept imputed sets and relations
00:18:31
moreover the model can occasionally phase uh on relatively simpler problems
00:18:37
as we have seen with an ex but also in the case of commonsense reasoning
00:18:41
uh with the relation between the homes in anyone's and therefore chat you but you can get the passion
00:18:49
oh oh well quieting in manipulating routes uh in the abstract sense
00:18:53
we doubt the castle inspection so it's really important to test
00:18:57
the more that systematically to really see this kind of limitations
00:19:02
a tiny um it is important to note is that when when the
00:19:07
models fail um it's fading in a really nice way if it's a
00:19:12
because the model does it uh still able to generate the core it
00:19:17
or possible explanation um this is of course um
00:19:23
i'll gonna sit in a nice stage of the model because in its that
00:19:26
you can somehow they recombine solution even if it's not able to perform right calculations
00:19:34
this means that we could somehow combine the the model with
00:19:37
the other type of c. stands for example in the context mathematics
00:19:42
that can be combination between uh chuck e. p. t. and a symbolic in gene that is able to perform calculations
00:19:48
how where where in cases in which we don't all dissolution and we don't want the rooms
00:19:54
underline the tasks this can make things quite
00:19:57
complicated uh especially the you know the two
00:20:02
uh understanding the spec to uh the explanation provided by the model
Conference Program
The Evolution of Large Language Models that led to ChatGPT (Andre Freitas, Idiap)
Andre Freitas, Idiap Research Institute
March 10, 2023 · 8:34 a.m.
664 views
Understanding Transformers
James Henderson, Idiap Research Institute
March 10, 2023 · 8:46 a.m.
369 views
Inference using Large Language Models (Andre Freitas, Idiap)
Andre Freitas, Idiap Research Institute
March 10, 2023 · 9:19 a.m.
ChatGPT for Digital Marketing
Floris Keijser, N98 Digital Marketing
March 10, 2023 · 9:58 a.m.
Biomedical Inference & Large Language Models
Oskar Wysocki, University of Manchester
March 10, 2023 · 10:19 a.m.
Abstract Reasoning
Marco Valentino, Idiap Research Institute
March 10, 2023 · 10:38 a.m.
120 views
The Risks Behind Large Language Models (Al Brown, Fujitsu)
Al Brown, Fujitsu
March 10, 2023 · 1:42 p.m.
Q&A: The Risks Behind Large Language Models (Al Brown, Fujitsu)
Al Brown, Fujitsu
March 10, 2023 · 2:01 p.m.
Round Table: Risks & Broader Societal Impact (Legal, Educational and Labor)
Lonneke van der Plas, Idiap Research Institute
March 10, 2023 · 2:07 p.m.
The Infrastructure to build Large Language Models (Vinay Pondenkandath, Cerebras Systems)
Vinay Pondenkandath, Cerebras Systems
March 10, 2023 · 3:12 p.m.
Q&A: The Infrastructure to build Large Language Models (Vinay Pondenkandath, Cerebras Systems)
Vinay Pondenkandath, Cerebras Systems
March 10, 2023 · 3:30 p.m.
Recommended talks
Deep Generative Models
Yoshua Bengio, University of Montreal, Canada
July 5, 2016 · 1:59 p.m.
2157 views