Speech Transcript - Interpretation of DNA evidence as a paradigm for speaker recognition

0:00:06	thank you for the uh
0:00:07	invitation to be here uh i it did come as a
0:00:11	surprised because as you know uh
0:00:13	uh immediately appreciate are not uh
0:00:15	uh a voice
0:00:16	or language recognition
0:00:19	person uh but right from day one i realise that there are lots of issues uh circulating here but uh
0:00:24	related to things that we've had to uh
0:00:27	struggle with in connection with yeah
0:00:29	i'm not even a D N A evidence
0:00:31	person mainly i work in a kind of medical genetics context and my main uh brighton but work is
0:00:37	you know looking for disease genes and cool you know 'cause
0:00:40	the fixed
0:00:40	tween
0:00:41	genes an interesting uh phenotype
0:00:43	uh but i've long had an interest in the interpretation of
0:00:47	D N A evidence
0:00:48	and uh
0:00:49	try to contribute a uh
0:00:51	a lot to the developments in the field over there
0:00:53	oh the yeah it is and i'm pleased to say that
0:00:56	we have made a
0:00:57	a lot of progress
0:00:59	um it's also clear that uh people in this
0:01:02	uh community here have made a lot of progress in trying to get
0:01:06	uh the
0:01:07	the field on
0:01:09	what i would regard as a more rigorous footing in terms of the
0:01:13	interpretation and i i'm thinking
0:01:15	uh in a second frames the context of all the evidence
0:01:18	that will be
0:01:20	comprehensible and meaningful in whole
0:01:23	uh and so i've done a little bit of background reading of uh
0:01:28	uh oh
0:01:29	interesting our work in the field by F and what team and several other people here
0:01:34	and so uh
0:01:35	it's clear that i you know don't
0:01:36	have
0:01:37	much to say about the
0:01:39	the basics but what i thought
0:01:41	i would um
0:01:42	do is take a some slightly contrary in position and uh i would say there seems to be uh uh
0:01:48	a kind of sense i got from the reading that the that the
0:01:51	that the grass is greener in the next field uh
0:01:54	that uh everything is solved and works very well
0:01:57	for deny evidence and and uh i'm going to tell you that that's not the case uh it's
0:02:02	very complicated and messy there are some compromises that
0:02:05	that's all for work
0:02:07	uh
0:02:07	we are saved
0:02:09	to a large extent by the
0:02:11	evidenced by the by the fact that D N A evidence is in general very good
0:02:14	evidence uh
0:02:16	and very powerful
0:02:17	uh and so even if you make a mess of the interpretation
0:02:21	uh the ultimate outcome might not be the wrong one but that's not always the case uh
0:02:25	and
0:02:26	actually the reality in courts today of the presentation of D N A evidence
0:02:31	it's still pretty dismal uh
0:02:33	and it doesn't matter when i get to the end are we talking about the
0:02:37	the latest generation of low template D N A evidence
0:02:40	where
0:02:41	very small amounts of D N A lots of stochastic
0:02:43	effect
0:02:44	uh and lots of
0:02:45	complications
0:02:46	so
0:02:47	um
0:02:48	those of you who are paying careful attention my
0:02:51	recognised some of the
0:02:53	whatever it in some of the children's change my tackles
0:02:55	uh
0:02:56	somewhat instead of talking i've talked about comparisons here instead of recognition so we we have the same debate in
0:03:02	indeed evidence that we shouldn't talk about deny identification because
0:03:06	identification is
0:03:07	is not possible and not the business of the scientific expert
0:03:11	um
0:03:13	personally i'm a bit more
0:03:14	lack some light on happy to use words that make sense the general public even if
0:03:18	uh we have to be careful about understanding uh
0:03:20	what they really mean but anyway you know
0:03:22	acknowledgement here i put voice comparison might happen
0:03:25	but in the end although my goal was to try and think about relationships
0:03:29	between
0:03:30	D N A evidence and uh and voice evidence
0:03:33	all the basic work has already been done by people here and i didn't feel i had very much the
0:03:37	way
0:03:37	so i'm really just gonna restrict myself to talking about the you know evidence
0:03:41	some of the problems that we've had
0:03:42	some of
0:03:43	my views on how well come them
0:03:45	uh and then we leave for the discussion uh
0:03:48	uh
0:03:49	the possibility for people to really raise parallels and you go advised me not to leave any time for discussion
0:03:54	because it's a very controversial area but uh i'm going
0:03:57	i'm going to try and take the risk
0:03:59	uh but uh
0:04:01	in fact
0:04:02	i packed quite a lot of uh
0:04:03	stuff into my slides and sounds to me a bit louder again
0:04:06	down
0:04:07	just move it down a little bit
0:04:08	um
0:04:09	the um
0:04:11	and i i wouldn't have time to get through it all properly
0:04:14	but
0:04:14	the um
0:04:16	you have the luxury of knowing that
0:04:18	uh you don't have to really get to grips with all of this material i just wanna give you the
0:04:22	the flavour of
0:04:23	things
0:04:24	the problems
0:04:24	we worry about
0:04:26	and uh historical
0:04:28	perspectives on the go back right to the sort of beginning of time so to speak of this whole um
0:04:33	weight of evidence academic literature a lot of it springs from this famous case uh
0:04:38	in california in the nineteen sixty eight
0:04:41	uh i could define how about all the papers written about how to interpret the evidence in this case correctly
0:04:46	it would
0:04:47	would go up to the roof and it wouldn't reach a conclusion the famous uh uh saying about the colours
0:04:52	because
0:04:52	uh it it is a very very interesting case uh i i mean i think you get the details just
0:04:56	from there that
0:04:57	numbers one made up
0:04:59	uh that might be associated with frequencies for various traits
0:05:03	that the defendants possessed
0:05:05	and
0:05:06	uh it was claimed the uh the true criminals also possessed
0:05:10	uh and you know it's lots of fun you can give this to students 'cause there's lots of things wrong
0:05:14	with this you know obviously those probabilities of just made up obviously independence is a problem
0:05:19	but this sort of more fundamental issues uh every packet
0:05:22	wave my magic wand and get rid of those problems and if those really work through probabilities and they really
0:05:27	were independent
0:05:28	um
0:05:29	what would the number you get by multiplying these probabilities
0:05:33	together what would it be
0:05:34	uh and
0:05:36	how does it relate
0:05:38	to the juror or the finder of fact
0:05:40	problem
0:05:40	of deciding whether the computer
0:05:42	well i'm i'm not gonna answer that problem
0:05:44	for you here entirely but uh
0:05:46	it's an interesting and and uh and difficult problem
0:05:50	but
0:05:50	certainly one version
0:05:52	to answering it
0:05:53	uh
0:05:54	and and one branch
0:05:56	of the academic literature
0:05:58	but i'm sort of merging things here that slightly uh
0:06:01	there there was
0:06:02	uh
0:06:04	um
0:06:05	developed slightly differently but the the the sort of canonical problem for D N A evidence is uh
0:06:10	we've got
0:06:11	a sample left
0:06:12	sign
0:06:13	yeah
0:06:13	crime scene
0:06:14	uh we've already had some discussion of the notion of a match
0:06:18	is meaningful for D N A evidence not always
0:06:21	um not follow template in a of and
0:06:23	stochastic
0:06:23	fixed
0:06:24	not for the older form of D N A profiles that was in use
0:06:27	in the early nineteen nineties and is still occasionally
0:06:30	crops up
0:06:31	but let's just
0:06:32	he rarely all the idea that you know there is a notion of a match of a yes no answer
0:06:37	uh and we've got some frequency information again um
0:06:40	let's not worry about where this frequency information comes from just believe it for the moment
0:06:45	um
0:06:46	and so
0:06:47	you know how
0:06:48	how convinced
0:06:49	should you be i think the sort of fallacy that uh many people has already been alluded to here is
0:06:55	to think well one and it's pretty small
0:06:57	so he must be guilty
0:06:58	a what you know what is that sort of logic there and
0:07:02	that was a bit of a
0:07:04	and academic literature a fun discussion that went on
0:07:07	a quite a few years um
0:07:09	it's uh
0:07:11	in retrospect the answers
0:07:12	seems very easy and you wonder how we manage to argue about it for quite a number of years but
0:07:17	uh of the uh
0:07:18	but uh anyway that's what economics uh and therefore and uh finding out
0:07:22	uh problems to argue over and
0:07:24	i can't consensus
0:07:26	eventually emerged around the
0:07:29	you know what
0:07:30	from uh
0:07:31	and orthodox
0:07:32	uh
0:07:33	bayesian position would be a kind of standard and straightforward response
0:07:37	you should be using
0:07:38	base there but uh and i put one version of it there um
0:07:43	you can so
0:07:45	uh introduce some notations C is the name of the person who committed the crime
0:07:50	or let's say of course
0:07:52	being the source of the D N A is not logically equivalent to committing the crime will just suppose it
0:07:56	is here
0:07:56	um and S it's the name of the
0:07:59	of the
0:07:59	offended
0:08:01	uh
0:08:02	and
0:08:03	there are some subtleties and difficulties built into here i'm gonna spend a little bit more time and
0:08:08	talking about
0:08:09	uh and that revolves around the idea of
0:08:12	of what is the
0:08:13	alternate hypothesis
0:08:15	uh
0:08:16	so
0:08:18	in um
0:08:20	as i i've already uh mentioned earlier this idea that i i think there's a bit of an impression
0:08:24	that the you know the grass is greener in the next field and things are easy if the D N
0:08:28	A evidence but one of the things that's uh
0:08:30	that's not easier in some uh is
0:08:33	this uh
0:08:33	specifying of the alternate hypothesis and
0:08:35	and the uh
0:08:37	can be
0:08:37	why uh
0:08:39	i difficulty logically
0:08:40	uh in a number of ways
0:08:43	uh but yeah
0:08:44	i'm going to assume so so in the
0:08:47	the
0:08:48	this
0:08:48	the speech recognition literature people have been happy to just posit that the
0:08:52	the null hypothesis the prosecution hypothesis if you like the same source
0:08:56	that the uh
0:08:58	uh queried rick
0:08:59	uh voice and the suspect's voice uh
0:09:02	come from the same individual
0:09:03	or different source
0:09:05	um
0:09:06	four
0:09:07	you know evidence at least
0:09:08	it can be that simple
0:09:10	and uh i've chosen to break down the whole time to type up
0:09:13	this year
0:09:14	into
0:09:15	a number of hypotheses of the form X did it for various things
0:09:20	uh but
0:09:21	for more complex problems there are different ways uh to break down the evidence
0:09:26	uh the alternative hypotheses
0:09:28	for example if there uh multiple
0:09:31	uh do you know samples which is often the case there are lots of alternatives around
0:09:35	different contributors to the different samples you know just uh
0:09:39	often it's just assumed implicitly that there's a single contributor
0:09:42	but for mixed samples
0:09:43	that's not at all straightforward
0:09:45	uh there's different alternative hypotheses a around uh
0:09:49	relatedness
0:09:50	and there's different alternative hypotheses around the number of contributors to the sample
0:09:55	um
0:09:56	but um
0:09:58	in this form here
0:09:59	i'm just thinking about breaking down the alternate hypotheses into all the individuals who it could have been
0:10:04	and we have to add up
0:10:05	over these uh evidence
0:10:07	um
0:10:09	logically you have to add up of everyone on a
0:10:12	uh and uh
0:10:13	and that this idea came up actually in a court
0:10:16	S and the judge was horrified at the idea that he has to sit there and think yeah
0:10:19	uh about every person on a one at a time
0:10:23	uh but i wanna emphasise this point that logically you
0:10:26	have to
0:10:26	um
0:10:27	the
0:10:28	if you want to prove
0:10:30	that particular individual is the source of your D enable the source of your voice recording
0:10:35	logically that means that everyone else on the
0:10:38	is not the source
0:10:39	and also alternate hypotheses around
0:10:42	uh you know synthetic um voice fabrication on these kind of things all of those hypotheses
0:10:47	uh have to be ruled out in order to establish the one type of
0:10:51	do you
0:10:51	care about
0:10:54	um
0:10:55	so a little bit of uh
0:10:58	manipulation we can right now
0:11:01	formula
0:11:02	like this
0:11:03	in uh again this is kind of just
0:11:05	uh
0:11:05	classic uh
0:11:06	way of breaking down the evidence
0:11:08	uh uh breaking down the calculation in bayes theorem
0:11:11	and the idea is to introduce some notation here i put a whole the likelihood ratio
0:11:16	and again i want to emphasise that isn't one likelihood ratio there are many and we count
0:11:22	that is difficult problem of how to combine the like
0:11:24	ratios
0:11:25	uh although we we we
0:11:27	we'd like to
0:11:28	um
0:11:29	i'm thinking
0:11:31	in terms of the
0:11:32	D N A evidence being interpreted last
0:11:34	and so
0:11:35	this other ratio the prior
0:11:37	i'm thinking all those incorporating all the other evidence i mean there's no
0:11:41	logical reason
0:11:42	for doing it that way around of course that's
0:11:43	it's a nice
0:11:44	coherence property of
0:11:46	of the bayesian analysis you don't
0:11:48	two one
0:11:50	uh you get the same answer which
0:11:52	whichever order you analyse the other
0:11:54	um
0:11:55	and in order to get it in this form you need to make a
0:11:58	uh assumption uh
0:12:00	various uh uh independence
0:12:02	some some
0:12:03	we could argue about that scene
0:12:04	generally reason
0:12:06	here
0:12:08	some putting the um
0:12:11	the
0:12:12	no getting to be able to write
0:12:14	uh weight of evidence in this form in a kind of forensic setting
0:12:18	was a pretty big uh
0:12:20	step for all that it took yeah
0:12:22	many years and lots of arguments and so on but it's
0:12:25	you know pretty much accepted amongst a bigger
0:12:28	community nowadays and
0:12:30	it overcame a lot of problems that people
0:12:32	uh struggled with i mean i've been in the field so long now that it's
0:12:36	sort of hard to remember how difficult some of these troubles were
0:12:40	um
0:12:41	but uh you know this basic idea that i nations you to well one in a million is really small
0:12:45	he must be guilty
0:12:46	uh
0:12:48	it's not it's not true and people didn't know how to think about that
0:12:52	uh until we were able to formalise the problem in this way
0:12:56	uh and now it seems pretty easy to think about it for
0:13:00	again this is a simplification in the general problem is not that simple but one way to think about is
0:13:05	how many alternate suspects
0:13:06	there are
0:13:07	uh and
0:13:09	you under some simplifying assumptions
0:13:12	you were essentially add up the likelihood ratio of a your alternative suspects
0:13:17	and so
0:13:18	a likelihood ratio
0:13:20	or one over a million
0:13:22	isn't convincing uh if
0:13:24	the number of alternate suspects
0:13:26	is larger very
0:13:28	into the time that there's you know no fundamental logical problem here about this uh
0:13:32	nice uh
0:13:33	distinction about between the role of the experts and the role of the
0:13:37	of the of the finder of fact
0:13:39	oh come back to that but but
0:13:41	uh you know this is certainly an only true under some simplifying assumptions
0:13:45	and whenever if i kind of present this idea in court
0:13:48	i have to sort of be careful about wording like if you choose to assume that all the alternate suspects
0:13:53	are equally likely
0:13:54	uh then you come up with a formula like this
0:13:57	of course nowadays um
0:13:59	likelihood ratios
0:14:01	a much
0:14:02	bigger or smaller whichever way you do them around uh and one million
0:14:05	uh and so uh
0:14:08	tip in typical cases the problem has
0:14:11	vanished but again i wanna emphasises lots of cases out there would mix profile small amounts of D N A
0:14:16	complex relatedness
0:14:18	we're all these issues still matter
0:14:21	uh
0:14:22	the role of relatives so that was a
0:14:25	again much confusion about this than in the past
0:14:29	uh
0:14:30	this nice uh
0:14:31	formalisation in terms of bayes theorem
0:14:34	i want i might slip into this
0:14:36	language but another point four
0:14:38	for discussion is i don't think
0:14:40	what i'm doing is fundamentally bayesian i tend to avoid
0:14:44	the label bayesian the way i'm just using bayes theorem and its
0:14:46	theorem probability that all uh
0:14:49	or uh or light here it's the model
0:14:51	mathematical probability except
0:14:53	uh in fact i would say my approach is fundamentally non bayesian in ways that i will point out uh
0:14:58	want to later
0:15:00	uh i just remembered now i forgot to put a slide on this there was a mention of it uh
0:15:04	that there was a big uh
0:15:05	court case um
0:15:07	in the U K a number of years ago where
0:15:09	the uh
0:15:11	that was strong D N A evidence implicating implicating a defendant
0:15:15	but it was quite a substantial amount of evidence in his favour any particular
0:15:19	uh the victim of this crime gave a good description of the defended
0:15:22	uh of the of the attack uh and the defendant didn't match you know gross mismatch
0:15:27	between the description and what he looked like
0:15:29	but she also said in court this does not resemble them and
0:15:33	you know he does not resemble the man
0:15:35	attack
0:15:35	since i wasn't interested it doesn't resemble
0:15:37	the man that attacked me and
0:15:39	so uh and he had an alibi and
0:15:42	wasn't near the scene of the crime at the time and so
0:15:45	um quite a complicated case went to summary trials and uh
0:15:49	the um i wasn't involved in that case but the the defence expert actually
0:15:54	uh proposed at all the jurors through
0:15:56	a bayes theorem calculation
0:15:58	uh with likelihood ratios for
0:16:01	the wood
0:16:01	description on likelihood ratios for the uh
0:16:04	for the D N A evidence
0:16:05	uh and likelihood ratios for the alibi evidence
0:16:08	and suggesting values and the jurors were asked to multiply them together
0:16:12	and the judge got quite enthusiastic about this and ordered somebody to go out and buy doesn't calculators for the
0:16:17	jurors to uh
0:16:18	multiply the numbers together but uh the judge kept getting zero and uh when he tried to do the calculation
0:16:24	himself
0:16:25	uh
0:16:25	anyway the um
0:16:27	uh
0:16:28	the guy was uh was convicted
0:16:30	but it went to appeal and the appeal court was sort of
0:16:32	horrified absolutely horrified about this uh complicated mathematical stuff that these uh wise all judges didn't understand
0:16:39	uh was uh about having this in court and so
0:16:43	the judgement was very severe that uh
0:16:45	you know bayesian methods one not to be introduced in U K courts that uh uh because you know ask
0:16:51	right here at all judges don't understand that and uh you know it's all lots of a power thing no
0:16:56	worried about losing that
0:16:57	losing that how about i just thought it was a sort of amusing
0:17:00	idea that any form of reasoning is is allowed there is no other
0:17:04	role as far as i know there's not
0:17:06	oh all reasoning is allowed in a british court except
0:17:10	the form of reasoning that's been established to sort of be logical and read and rational and reasonable that's the
0:17:15	only thing you're not allowed to present in a
0:17:17	in a british court uh
0:17:19	so that so that was a bit of an aside but that's why i sort of avoid the label
0:17:23	bayesian and i think it's uh it is uh irrelevant and um
0:17:28	uh
0:17:28	to what we are doing
0:17:30	uh and of course i don't have a
0:17:32	explicitly introduce mathematical formalism
0:17:35	whenever i'm giving expert witness
0:17:37	but i do try and talk to jurors through this kind of thing and say
0:17:40	imagine how many
0:17:42	close relatives there are
0:17:44	of the defendant
0:17:45	what is the match
0:17:46	what is the matter probability for then
0:17:48	uh and imagine how many unrelated
0:17:50	people what's the matter probability you gotta combine the total weight
0:17:54	uh to come up with a
0:17:56	and if that combined weight is norman
0:17:58	negligible
0:17:59	then
0:18:00	you've got reasonable doubt about
0:18:03	having the right guy
0:18:04	uh unless there's other evidence uh implicating those
0:18:07	uh oh or not
0:18:08	taking the brothers
0:18:10	four
0:18:11	uh
0:18:20	uh
0:18:22	yes many of these features of course i'm gonna be talking you want really relevant to voice recognition but i
0:18:27	think some of them will be
0:18:28	uh i'm talking uh
0:18:30	i've got a sneak in the label here now i only mean genetic uh
0:18:33	ideas of uh
0:18:34	of ethnicity here
0:18:36	um
0:18:37	relatedness matches the do you know i haven't
0:18:40	uh
0:18:40	and
0:18:41	this is really the same issue is
0:18:42	as
0:18:43	close relatives
0:18:44	but it's just relatedness
0:18:46	on a more distance scale
0:18:49	uh so the relatedness of people in an isolated uh
0:18:53	uh geographical or religious group
0:18:55	someone
0:18:57	bay yeah
0:18:58	compared to relatives
0:18:59	they're relatedness is less
0:19:01	relatives like cousins and so forth
0:19:03	they're relatedness is less than is typically more of them
0:19:06	and so they
0:19:07	you know they
0:19:08	i kind of plausibly balance out
0:19:10	um
0:19:11	i'll come back to this
0:19:13	uh really really important of these issues are around uh and and a real fundamental difficulty that i don't think
0:19:19	we really have sold
0:19:20	is what to do about
0:19:22	lab
0:19:23	air uh
0:19:24	labelling errors
0:19:25	uh and outright
0:19:27	evidence for award
0:19:28	um
0:19:30	but at least the bayes theorem
0:19:31	paradigm
0:19:33	tells us how to think about the problem and what
0:19:35	relevant issues out but
0:19:36	but what it tells us is um
0:19:39	uh
0:19:39	is a little bit worried i mean first of all one thing that's not wiring is that some
0:19:44	critics of D N A evidence we're going round saying well you know i know
0:19:47	no human activity has an error rate less than about one in a thousand
0:19:51	and therefore
0:19:52	numbers like ten to the minus six ten to the minus seven ten to the minus eight
0:19:55	come up in uh
0:19:57	in in connection with dean evidence a completely meaningless
0:20:00	so that reasoning is invalid
0:20:02	this is not the probability of any error that matters
0:20:04	but only a narrow that generated the data that we observe
0:20:08	so
0:20:09	uh
0:20:10	there's a famous story that uh feel
0:20:12	david uh
0:20:14	pointed me to from
0:20:15	price i think
0:20:16	probably eighteen century english philosopher to discuss this point that
0:20:19	uh
0:20:20	a printing error in the newspaper is more likely than you winning the lottery
0:20:25	uh but nevertheless
0:20:26	if you see a number printed in the newspaper as the winning lottery number
0:20:31	you don't through the newspaper out and say a probably a printing
0:20:34	uh
0:20:35	fig
0:20:35	because uh
0:20:37	the uh because it's not any printing error that matches the printing a rather generated your number
0:20:42	is much less likely that you winning the lottery and therefore you do
0:20:45	through the paper up and run down the lottery office to claim your prize
0:20:49	um but
0:20:50	this is a fundamentally a problem
0:20:53	that uh
0:20:54	some of the
0:20:56	more reason critics of D N A evidence i don't think we can easily get away with this
0:21:00	that uh evidence tampering
0:21:03	doesn't involve this problem because evidence
0:21:05	prob tampering doesn't generate
0:21:07	the evidences of
0:21:09	um
0:21:10	and
0:21:12	she quickly
0:21:14	a reasonable persons
0:21:16	view of the probability that the police or somebody else
0:21:19	tempered with the other
0:21:20	in some way
0:21:21	it's gonna be much greater than a match than a match probability or likelihood ratio
0:21:25	in connection with you know
0:21:27	um
0:21:28	and
0:21:29	logically i think it is true
0:21:31	but because of this
0:21:33	the
0:21:34	this typically will swarm
0:21:36	the uh the significance of the match probability
0:21:39	D N A evidence
0:21:40	so that
0:21:41	if you do get a good you know profile match
0:21:44	the actual number connected with it is pretty meaningless it's
0:21:48	it's virtually impossible and the only way down now too
0:21:51	thinking about these kind of alternatives
0:21:54	but what i nets but can do about that in court
0:21:56	is quite uh
0:21:58	it's quite difficult you could you you can't even
0:22:01	consider putting numbers on this kind of thing of course uh but uh
0:22:05	but
0:22:06	ideally
0:22:07	you should be alert injurious to this possibility
0:22:09	uh and uh
0:22:11	that it should be
0:22:13	you know wait
0:22:13	in combination with the
0:22:15	with the match probability for the D N A
0:22:23	and
0:22:23	i don't have time to go this here this is also the start of stuff but this was a sort
0:22:27	of fun
0:22:28	debate will not find really 'cause i did get a bit tedious it just went on and on and on
0:22:31	and it still goes on this uh argument about the uh
0:22:35	effect of the evidence i know some of your read some of the literature so you'll be aware of these
0:22:38	issues but uh
0:22:40	have some of you want be one
0:22:42	i imagine case number one
0:22:45	uh i just say you know he matches and it's one in a million probability of a match
0:22:49	uh case number two i tell you those two facts but also tell you all by the way
0:22:54	uh i found him by looking through our database of D in a profiles
0:22:57	he was the only match
0:22:59	now
0:22:59	in which case
0:23:00	is the evidence stronger case one okay stew
0:23:03	the classical statistical viewpoint is a case too is uh evidence trolling
0:23:08	uh you've gone through the uh
0:23:10	uh you know you're going out fishing for hypotheses and we all know about multiple testing and one for any
0:23:15	corrections
0:23:15	kind of thing
0:23:16	uh evidence is much weaker
0:23:18	if you go fishing for hypotheses
0:23:21	um
0:23:22	so
0:23:24	uh if you've the defendant has been identified through a search in a database
0:23:28	or
0:23:29	D N intelligence database
0:23:31	of known uh
0:23:32	previous offenders
0:23:33	uh the the data
0:23:36	uh
0:23:37	we can
0:23:37	then in a standard
0:23:39	and of course uh
0:23:41	that's
0:23:41	completely wrong uh this i i think the uh the standard uh
0:23:45	statistical reasoning is just uh inappropriate here certainly
0:23:49	all of that you know there's a classical argument about frequentist and bayesian views or
0:23:53	in the literature but
0:23:54	but often the frequentist in the bay seems get to roughly the same place in the end
0:23:58	that is one example where they get to very different places
0:24:01	and
0:24:02	when it because of the strong logical foundations of the bayesian foundation
0:24:06	whenever the two of them disagree
0:24:08	uh
0:24:08	it's pretty much always the bayesian view that's right
0:24:11	uh and
0:24:12	and it is
0:24:13	yeah what
0:24:15	people who are worried about the evidence trolling idea on that weakening the evidence
0:24:20	the problem with their approach is they are not
0:24:23	being critical enough
0:24:24	in the first
0:24:25	because even if i just
0:24:27	even if there was no
0:24:28	um
0:24:29	evidence trolling even if there was no database search i still have to logically
0:24:34	to prove that this guy committed
0:24:35	crime i have to prove that every other person on a
0:24:39	didn't commit the crime
0:24:40	those two
0:24:41	formulations are equivalent statements of the problem that's a really tough
0:24:45	task and nobody else
0:24:47	nobody would have dared to
0:24:48	even contemplate that in the past because it was unthinkable that you could prove that everyone else on the didn't
0:24:53	do it
0:24:53	but you can now would be an evidence you can think about an ugly
0:24:57	arguably due
0:24:58	and so
0:24:59	any amount of
0:25:01	fishing or trolling for hypotheses
0:25:03	is it doesn't change that fact
0:25:05	you still have to prove that everyone else on a didn't commit the crime
0:25:08	and in fact
0:25:09	uh
0:25:10	it makes life
0:25:11	better because everyone else in the database didn't match
0:25:14	so that helps even your task
0:25:16	proving
0:25:16	that everyone else on a didn't commit the crime
0:25:18	'cause you've got a whole lot of people that you've shown not to that a profile doesn't match
0:25:23	um
0:25:24	that
0:25:24	is
0:25:25	but latter argument is an argument about uh
0:25:28	hypotheses rather than so this
0:25:31	issue about probabilities of evidence was probably
0:25:33	posses
0:25:34	um
0:25:35	i know it's a bit more about in a moment we like to separate the two but i
0:25:40	insist that fundamentally it's not uh
0:25:42	it's not possible to
0:25:43	achieve that ideal in many situations
0:25:46	that that's not really uh
0:25:47	what
0:25:48	yeah
0:25:50	um so
0:25:51	yeah
0:25:51	some of the things right feel that the
0:25:54	the
0:25:54	the the
0:25:55	formulation
0:25:57	of the problem
0:25:59	even though it's just a sort of standard was no
0:26:01	bayes theorem it wasn't obvious for about
0:26:04	twenty or thirty years after the commons case when all this academic literature was piling up
0:26:08	people didn't get to this position
0:26:10	of just writing down bayes theorem and seeing its implications in the way described in them
0:26:15	uh nowadays the majority of people even in the field don't
0:26:18	uh so i don't succeed in understanding the evidence this way but there's a big enough community obvious that do
0:26:23	uh that that isn't the problem
0:26:25	but um
0:26:26	there are many many uh
0:26:28	problems that remain uh and uh
0:26:32	i have already so stressed this one but this is one of my key points about
0:26:36	the difficulty is that uh
0:26:37	you know it's nice to think about a competition between the prosecution hypothesis
0:26:41	and the defence side
0:26:42	sis and uh
0:26:43	you know many lawyers have argued with me that is fundamentally
0:26:47	is what the whole legal system is based on the competition between two hypotheses
0:26:52	and a sorta reject my idea
0:26:54	but i've represented at but i
0:26:57	claim that this is just a straightforward logical situation that in order to
0:27:01	uh establish
0:27:02	hypothesis one the prosecution hypothesis
0:27:05	you must
0:27:05	proof that every other competing hypothesis is false whether or not the defence puts forward and of course and
0:27:11	most uh legal systems the defence downtown
0:27:13	to put forward any story at all of course
0:27:15	uh
0:27:15	and even if they do put forward a story
0:27:18	uh
0:27:19	judges
0:27:19	usually advise
0:27:22	the court in a
0:27:23	that um the jurors uh in now
0:27:25	setting
0:27:26	that um
0:27:27	they don't necessarily
0:27:29	uh even if they disbelieve the defence story it doesn't necessarily need the defendant is guilty these a separate question
0:27:36	to be
0:27:37	and sit separately
0:27:38	uh
0:27:41	it's inevitable that the forensic scientist has to make
0:27:44	subjective
0:27:45	judge
0:27:46	about
0:27:47	for example implausible hypotheses
0:27:49	uh so
0:27:51	uh in D N A evidence you've always got i was my
0:27:54	identical twin
0:27:56	story which makes D N A evidence uh
0:27:58	completely uh
0:27:59	useless
0:28:01	uh it's actually quite remarkable
0:28:03	how rarely that is used i think that everyone would just laugh it out of court actually identical twins are
0:28:08	not rare
0:28:09	and it's very hard to prove that you don't have an identical twin
0:28:12	uh so if any of you do commit a serious crime and rub on court would again errands i do
0:28:16	recommend you try the story that uh and uh
0:28:19	uh i i i think logically it's hard to be uh um
0:28:24	the uh but nevertheless in practice acting queueing sixties unfortunately that uh
0:28:28	uh
0:28:29	the
0:28:30	but also in the evidence we have the number of contributors do we deny sample even if there's no more
0:28:35	than two little that every locker
0:28:37	it doesn't follow that there's only one contributor
0:28:40	there's no what the bound on the number of contributors
0:28:42	uh and
0:28:44	i'm the involved um right in the middle of a court case uh you know that i was giving evidence
0:28:48	um fried enough to go back and continue mild evidence tomorrow
0:28:51	uh and then the uh i it looks like one contributed
0:28:55	the crime sample
0:28:56	and i did some calculations one contributor to contributors
0:28:59	and of course that the that i'm not advising the prosecution in this case usually on divine advising the defence
0:29:04	uh but the defence of course it jumped up and said you haven't done any calculations for three contributors
0:29:09	and and of course i said well
0:29:11	you know there's no sign of even to contribute to so three contributors is ridiculous and they say a but
0:29:15	you cannot rule out the possibility of three contributors and i have
0:29:18	can see the uh
0:29:19	but i can't you know that's a subjective judgement
0:29:21	right
0:29:22	uh and uh
0:29:24	i kind of transgress is this idea of trying to
0:29:28	key
0:29:29	a clear logical distinction between the likelihood ratio
0:29:32	uh and the uh
0:29:34	and the probabilities of hypotheses
0:29:36	uh
0:29:37	and if you can read this
0:29:39	uh
0:29:40	in this respect scene is under what is is completely unavoidable that you can't uh you can't avoid
0:29:46	making judgements about probabilities of hypotheses
0:29:48	uh but nevertheless we should maintain the goal right just behaviour which is to try and avoid as far as
0:29:53	possible
0:29:54	any assumptions about the hypotheses and
0:29:57	you know to be aware of them and make them explicit as far as we can
0:30:02	the um oh i didn't mention this one as well contamination rates also an issue here
0:30:08	there is um
0:30:09	sometimes it's easy to get confused with discussions of
0:30:12	of priors because there is a
0:30:14	prior on the hypothesis
0:30:16	uh that he's guilty
0:30:18	uh for example
0:30:20	and that's very clearly not the business of the expert
0:30:23	uh
0:30:24	and this is the
0:30:25	and is the master of the
0:30:27	a finder
0:30:27	fact and it's for you know we have to be very careful in our wording to avoid any suggestion that
0:30:31	we're expressing a view
0:30:32	uh on the probability that he's guilty either before or after the evidence
0:30:36	but of course we
0:30:37	we include priors for other quantities all the way along a particular rate of contamination
0:30:42	so with low template do you know profiles
0:30:44	it's just amazing how difficult it is to get rid of contamination our environment is entirely covered with D N
0:30:50	A you know for four meters around me
0:30:52	there is my D N A staff it everywhere from my bread uh uh
0:30:56	and i know touching things leaves your D N A
0:30:59	uh it's a very kind of shocking
0:31:01	so when you think about you know it is room is entirely covered with D N A
0:31:06	uh the um
0:31:07	but uh very thin film obviously
0:31:09	and
0:31:11	we cannot ever exclude that
0:31:14	there are some of the illegals we see in a mixed profile got there
0:31:18	not through any of the main contributors that we're thinking about is the offender
0:31:22	crime
0:31:22	but some environmental
0:31:23	combination
0:31:25	uh that's a really serious issue with low
0:31:28	low template D N A profiles but in any case any assessment
0:31:31	about contamination rate
0:31:33	is it is effectively a prior judgement
0:31:35	um
0:31:36	based on you know there is
0:31:38	is it
0:31:39	so
0:31:40	into that the
0:31:42	um
0:31:45	okay
0:31:46	i yeah
0:31:48	as us
0:31:49	one that got too much stuff here i want to say very much i haven't said anything really about the
0:31:53	technology of the of the you know profiling
0:31:56	um i'd out there's a little bit there are those of the you don't know
0:31:59	it's just that uh
0:32:01	he short tandem repeat profiles
0:32:03	a little words of D N A the repeated a number of times
0:32:06	and the number of repeats affects the length
0:32:09	uh and the current technology dist it's still not sequence based even though
0:32:13	this may change in the future but there's so much investment in this technology now
0:32:17	time to think about changing it
0:32:19	we don't actually read the sequence we just
0:32:21	measure the length of it in a fragment
0:32:23	and the length is measured by running uh these fragments through a gel
0:32:27	and there's a laser
0:32:28	i detector at the finish line and
0:32:30	time
0:32:31	taken
0:32:31	uh her response
0:32:32	the length
0:32:33	fragment
0:32:34	usually we can interpolate
0:32:36	the number of repeats so you might have seven copies
0:32:39	of the repeat someone promise on the nine on the other
0:32:41	so you would unit i would be represented the seven nine
0:32:44	but uh unfortunately for this nice story partial repeats
0:32:47	do
0:32:48	okay so this doesn't mean nine point three to decimal number it means nine copies
0:32:53	all the four base pair repeat and then
0:32:55	three base pair
0:32:57	uh
0:32:57	fragment
0:32:58	a repeat
0:32:59	and
0:33:00	but uh
0:33:01	nevertheless it is
0:33:02	pretty much
0:33:04	possible to
0:33:05	to say yes no whether the um
0:33:08	whether the fragment lengths match
0:33:10	uh and this is sort of idealised view
0:33:12	of the
0:33:13	electra fairground
0:33:15	basically a time series plot as these freshmen
0:33:17	past the finish line
0:33:18	uh there they are
0:33:20	there are
0:33:21	dies you know coloured eyes you can think about
0:33:24	uh that distinguish
0:33:26	the fragments from different loci
0:33:28	uh and then
0:33:29	different loci have
0:33:30	fragments
0:33:31	and it in a different length ranges
0:33:33	so that enables you in one test you
0:33:35	uh
0:33:36	to uh
0:33:37	and the lies
0:33:38	channel twenty uh genetic loci
0:33:41	and
0:33:42	in the current technology we all done i mean we'd love to be able to take into account heights of
0:33:47	these peaks
0:33:48	uh but we don't we we just it was of the binary yes no
0:33:51	uh there is a piece here
0:33:53	um
0:33:54	and i'll come back to that and a little bit if i have
0:33:56	time because with small amounts of D N A that
0:33:58	problematic
0:34:00	um
0:34:02	so there's a lot of problems with these issues of where
0:34:05	does that where do the probabilities come from
0:34:07	um
0:34:09	and
0:34:10	uh people have mentioned to me here and it's sort of true that in um
0:34:14	you know in D N A
0:34:16	it's easy 'cause we've got population genetics theory which generates uh
0:34:20	um
0:34:21	uh which generates probabilities and of course
0:34:24	in a larger population genetics theories
0:34:26	based on one of bruno's famous sums of course mental who is a uh here and that
0:34:31	uh you did is a work on the P Z here um
0:34:35	and but nevertheless
0:34:38	although
0:34:39	mentors
0:34:39	labels as applied here
0:34:41	ah
0:34:43	near enough to being objective fact
0:34:46	uh there's a lots of elements of
0:34:47	theory that subjective
0:34:48	uh
0:34:49	strength of D N A evidence
0:34:51	it's all about related
0:34:53	uh and
0:34:54	these questions of independence all questions about relatedness
0:34:57	how you model relatedness
0:34:59	is you know a typical story in complex scientific evidence
0:35:03	you think about all the people in this room we've got hugely complicated
0:35:07	of relatedness through all think of my all my lineage as
0:35:11	mother father for grandparents great grandparents you know go back five generations where
0:35:16	where up to large numbers of ancestors and then any other individual in this room
0:35:20	every has got
0:35:21	you know same so many lineage is up to sixteen great grandparents
0:35:25	and
0:35:25	every
0:35:26	yeah all those lineage is one of my sixteen great grandparents someone of your sixteen great grandparents
0:35:31	all meet in a common ancestor at some point
0:35:34	past and
0:35:35	unless you think i i'm an alien from another planet but more or less the uh
0:35:39	that's pretty substantial evidence that we all have common ancestors
0:35:42	so
0:35:43	the
0:35:45	fully detailed model would specify all the patterns of relatedness for every individual on a
0:35:51	and of course that's ridiculous
0:35:52	the complicated
0:35:53	so we have to make simplifying assumptions
0:35:55	uh
0:35:56	and most of the models
0:35:57	kind of break
0:35:58	relatedness down into three levels known relatedness
0:36:01	which is usually you know just one or two generations in the past
0:36:04	uh relatedness future unknown shared ancestors but understood to be
0:36:09	on a relevant relatively recent time scale and
0:36:12	and how you define recent is how these theories
0:36:15	very
0:36:16	uh and then the completely unrelated case
0:36:18	is an idealised
0:36:19	case where the ancestors of so far back
0:36:21	but it really doesn't matter
0:36:22	we can just
0:36:23	assume independence
0:36:25	um
0:36:26	the
0:36:27	they are kind of good enough models even uh
0:36:29	he too few people really understand how they work
0:36:32	now but you know the great
0:36:34	vocational reality
0:36:35	and i just wanna emphasise the uh
0:36:38	this objective miss of the underlying model of these models
0:36:41	um
0:36:43	there's a lot of argument over the years about independence various independence assumptions that go together
0:36:48	and that's of course important
0:36:50	four
0:36:50	for you guys as well uh
0:36:53	we did this is where we do have an advantage that uh in in the
0:36:57	the only dependence that matters
0:36:58	is due to relatedness
0:37:00	uh
0:37:01	the other important point i want to make it causes that um
0:37:05	is whether or not to think you know to kind of meaningless thing to say a is independent to be
0:37:09	in a in a kind of a general real world objects that
0:37:12	uh independence is all about what information you condition on and if you get the conditioning right
0:37:17	things
0:37:18	are typically independent to a good enough approximation the example i've used this uh
0:37:23	is uh
0:37:24	reading ability and shoe size in children are not independent
0:37:28	the
0:37:29	the bigger the better readers have big F eight
0:37:32	uh and uh that's a very well established fact then you can look at the correlation it's quite strong
0:37:37	uh of course they depended because of the varying ages both of those things are correlated with age
0:37:42	if you uh
0:37:43	you condition on age
0:37:44	the dependence goes away uh and uh
0:37:47	similarly
0:37:48	uh
0:37:48	if you condition if you do the right conditioning for D N A evidence
0:37:52	uh you
0:37:53	and make a reasonable assumption of in
0:37:55	and that sort of course you know if you
0:37:57	if you want to take a contrary imposition which of course defences in court so sometimes do you can never
0:38:01	rena rigorously prove anything to be independent
0:38:04	um
0:38:07	so
0:38:08	the
0:38:09	what
0:38:10	what matters fundamentally in the match probability is a statement like this
0:38:14	at a single locus the probability that an unknown individual acts
0:38:17	as gina type A B
0:38:19	even
0:38:21	yeah
0:38:22	oh
0:38:25	i
0:38:27	but
0:38:28	okay good
0:38:29	just a speck
0:38:30	um
0:38:31	and
0:38:33	so
0:38:34	what matters is affected by this conditioning and of course the probability that this guy's got a be given that
0:38:40	this guy's got a bee
0:38:41	depends on the on their relatedness
0:38:44	what doesn't matter
0:38:46	and they argued about at great length
0:38:48	is the dependence or otherwise of the two labels
0:38:50	within a block us
0:38:52	so cold hardy weinberg equilibrium i mean again uh
0:38:55	um
0:38:56	much discussion about this
0:38:57	it's relatively unimportant
0:38:59	i've had
0:39:01	overemphasised in my writing this condition because that's what i see is the important one
0:39:05	but of course there's a lot of other stuff in the conditioning as well as all kinds
0:39:08	assumptions and
0:39:09	background data
0:39:10	uh that you are relying on and i'll i'll say more about that moment
0:39:15	now i see i'm going to take uh
0:39:17	i didn't intend to
0:39:18	uh following because advice and uh use up all my time and not leave any for discussion but
0:39:23	uh there's a bigger
0:39:25	this is where i say that uh
0:39:27	what i've been doing
0:39:29	is fundamentally non bayesian although based on bayes theorem
0:39:32	because uh all of these theories require
0:39:35	uh parameter estimates
0:39:36	and and everybody likes to putting plugin estimates
0:39:39	uh
0:39:40	the simplest thing to do
0:39:41	but also
0:39:42	you know
0:39:43	you can think about what the different estimates are the different parameter estimates are change them
0:39:48	that the parameters for us
0:39:49	are they really all frequencies
0:39:51	uh and this population genetics parameter which is the average
0:39:55	relatedness in a community
0:39:57	uh we've got various estimates of these
0:40:00	um
0:40:03	we like to use
0:40:03	plugin estimates it has an advantage that it keeps the
0:40:06	the evidence specific to the case over here and all your training and background data that feed into apply guest
0:40:11	estimates over there
0:40:13	of course the ideal
0:40:15	and and again this or the bayesian
0:40:17	position would be to integrate out the unknowns
0:40:20	and
0:40:21	to in a sense
0:40:22	combining the data so uh and and
0:40:25	you know feel david in
0:40:27	london right
0:40:27	papers trying to do this where
0:40:29	the you know the
0:40:30	the days you conditional is not just the dependence
0:40:32	profile
0:40:33	but the defendant profile and all the profiles even the scene before
0:40:36	uh that uh that formula background information
0:40:39	um
0:40:40	so we are recognising this idea like to think it is just a bit too complicated
0:40:45	um
0:40:46	and so i have
0:40:47	donna sort of a good compromise
0:40:49	of using plugin estimate
0:40:51	but in recognition
0:40:53	uh all these uh problems
0:40:57	the about the um
0:40:59	the expectation all the hype how i can be much greater than the power the expectation
0:41:04	so this is why putting in plugin estimates
0:41:07	at that uh something like maximum likelihood estimates
0:41:10	can be really really misleading uh because you know the out the effect of uncertainty is not symmetric
0:41:16	uh when you've got high powers and product
0:41:18	um
0:41:19	so
0:41:21	that's right you know there's a lot of again this sort of boston vast amount of wasted literature in this
0:41:25	field like there is in any academic field so you have people talking about how to do
0:41:29	maximum likelihood estimates of these plug in parameters
0:41:33	and it's just a complete waste of time because the maxima like to estimate or anything like in any kind
0:41:37	of sensible estimate in the middle of the distribution is hopelessly wrong
0:41:40	uh because of this problem here
0:41:42	um
0:41:43	so
0:41:44	but i haven't really got a very good solution i just say well we want something new the top of
0:41:49	the plausible range like a ninety eight or ninety percent or something like that
0:41:53	uh although of course i haven't really got any formal
0:41:55	just
0:41:55	cation
0:41:56	doing that
0:41:58	um
0:42:01	right
0:42:01	a lot more to say what should i choose to include
0:42:04	i can't resist talking a little bit i talked about
0:42:07	the the probability is coming from
0:42:10	uh series
0:42:11	uh
0:42:12	uh population genetics theories which sound very brandon i can easily put them past just get you know a courtroom
0:42:17	who never sort of question me about any of these things but ultimately when you looking them
0:42:21	it's all full of subjectivity and judgements and
0:42:24	and i chosen this theory and not the theory and so on
0:42:27	um
0:42:29	of course many people are happy with that kind of subjective element and they want to sort of rigorous
0:42:33	and one way and again the sort of classical statistical position
0:42:37	to get a kind of rigorous probabilities
0:42:39	uh is to put it in the context of random sampling so that a lot of literature out there and
0:42:43	a lot of
0:42:44	um critical thinking which is based around the idea
0:42:48	that the suspect as being chosen randomly in a population
0:42:51	uh now i've already talked about evidence tampering and uh relatively high probability that the police could fiddle with the
0:42:57	evidence
0:42:57	but the possibility that the
0:42:59	police are capable of uniform random sampling is completely ridiculous i don't to accuse them of that of course all
0:43:05	the
0:43:05	all the X
0:43:06	that's in the world can't do uh in a find it very difficult to do uniform random sampling
0:43:11	uh and so
0:43:13	you know many people think
0:43:14	that
0:43:15	uniform random sampling idea
0:43:17	uh
0:43:18	click
0:43:18	field on a rigorous footing because these are objective probabilities
0:43:22	uh and i would say yes
0:43:24	objective that is clearly nonsense uh the police haven't
0:43:26	sample size
0:43:27	X randomly
0:43:28	this is just a completely made up assumption
0:43:30	uh which is uh
0:43:32	uh
0:43:34	i i mean i'm probably overdoing it here i mean it is the kind of assumption that
0:43:37	that people make and for good reason in some settings
0:43:40	i don't think we need to make it here
0:43:42	and it doesn't lead to lots
0:43:44	problems and in particular
0:43:46	the sort of
0:43:47	endless endless arguments about in which population has suspect being randomly chosen
0:43:53	and i say
0:43:54	here that you know because there is no such
0:43:56	sampling and there is no such population is like arguing over the number of weenies on the two very that
0:44:01	uh
0:44:02	yeah that that you know there is there is no such object so there's no point arguing about the properties
0:44:07	um
0:44:08	but
0:44:08	you know there is a real fundamental problem here that the more now really define the population
0:44:13	the better it is for the defendant
0:44:15	and we usually try to sort of
0:44:17	leaning defendants direction but the only logical endpoint of this
0:44:21	is uh
0:44:22	the population of size one that includes the defendant and have a hundred percent frequency for the uh for the
0:44:27	dependence profile which of course is a useless uh
0:44:29	position
0:44:30	and of course
0:44:31	so you get
0:44:32	you lose all these nice advantages
0:44:34	of the
0:44:35	bayesian formulation because this friend sampling hypothesis
0:44:38	i don't suppose you could do it
0:44:39	i mean you just can't do it in a bayesian way because it's just a ridiculous hypothesis that's got nothing
0:44:43	to do with the
0:44:44	with the with the evidence
0:44:46	um
0:44:46	and all this stuff which works well in the
0:44:49	framework i've been telling you about
0:44:51	hard to do in this
0:44:52	in this setting
0:44:55	right
0:44:55	i
0:44:56	this is an old topic of
0:44:57	mine and i will skip over this one that the the the U S
0:45:01	national research council did a report maybe fifteen years ago it still and hold absolute sway in the U S
0:45:06	uh
0:45:07	it's all
0:45:08	it's all based on this random and hypo
0:45:10	sis and it's all
0:45:11	kind of
0:45:12	riddled with errors but it's interesting in the sort of social psychology of the feel
0:45:17	we had huge arguments about D N A evidence and an early nineteen nineties
0:45:21	any nineteen ninety six the mood was just right
0:45:23	the kind of settled on a compromise
0:45:25	and so the authority of the
0:45:27	national research council in the U S was
0:45:29	was such that everyone kind of lead on this
0:45:32	uh and
0:45:33	in some kind of prey
0:45:34	consensus it sort of worked you know D N A evidence
0:45:37	based on this is gonna be does a lot of people in the U S and they're probably all guilty
0:45:41	but the fact that it's a completely riddled with misunderstandings and errors and uh and
0:45:46	and the evidence is being devastated
0:45:48	in almost every
0:45:49	court case in the U S
0:45:51	involving deny evidence the evidence is routinely overstated because of
0:45:55	oh
0:45:57	you know the truth is that the evidence was probably pretty strong anyway and this is why we haven't had
0:46:01	the kind of gross
0:46:03	miscarriages of justice
0:46:04	coming to light
0:46:05	uh
0:46:06	that would just a
0:46:07	channel these floors
0:46:09	so i won't go into that but all these things i've been talking about they
0:46:12	they use on the stored
0:46:13	but really the important thing was this uh population genetics
0:46:17	where
0:46:18	what we
0:46:19	care about is the conditional match probability
0:46:22	but what they cared about was just the marginal probability
0:46:25	and everything all the population genetics
0:46:27	issues are in this conditioning
0:46:29	uh and so
0:46:30	by leaving that out
0:46:32	they had a whole population genetics experts on this comedian they had big chapters on population genetics and completely missed
0:46:38	the point
0:46:39	uh and gave completely misleading and recommendations
0:46:43	oh
0:46:44	now
0:46:44	i yeah i want you that i tried to have a
0:46:47	too much of material in this talk
0:46:48	and um
0:46:50	i've withdrawal these topics but i just wanna bring you up to date with some of the uh
0:46:54	let 'cause everything i've been talking about today i could've
0:46:56	talked about years ago it's sort of a what the arguments of the nineties
0:47:00	uh and uh really
0:47:01	uh two thousand
0:47:03	but
0:47:03	but what's really
0:47:05	really come to a crunch this year in particular is what to do about this low template D in a
0:47:10	way down to
0:47:11	getting D N A from samples of just two or three cells and so this huge stochasticity in the results
0:47:17	um
0:47:18	the uh and of course many
0:47:20	jurisdictions just say this is way too complicated and
0:47:23	we don't want to touch this
0:47:25	uh but
0:47:25	more and more particularly uk that more and more people are
0:47:29	and it ended in and and uh it is potentially you know it doesn't mean
0:47:32	that just from the slide
0:47:34	touch
0:47:34	it's rather than collect a fingerprint
0:47:36	uh it's
0:47:38	it's
0:47:38	can be strong evidence to collect
0:47:40	D N A from this way
0:47:41	phone
0:47:42	i think
0:47:42	great
0:47:43	uh
0:47:45	but we get all these kind of stochastic features
0:47:48	i've got some slides yeah one huh
0:47:50	time to
0:47:50	but um
0:47:51	these peaks that i showed you about you get
0:47:54	so the top half as we could be in a
0:47:56	good amount of D N A
0:47:57	and this is with a sort of moderately low amount of D N A and you get all these features
0:48:01	like uh
0:48:02	peak imbalance but most one really
0:48:04	complete drop out of any of the labels either two peaks there
0:48:07	but there's only one showed up here
0:48:09	and that's
0:48:10	because
0:48:11	the the P C R reaction that underlies
0:48:13	the whole thing with such
0:48:15	with so few cells involved it can just completely fail if there's some uh
0:48:19	uh you know mutation the primer or something else goes wrong
0:48:22	um
0:48:23	and you can get dropped in the contaminant really owes you
0:48:26	you would have thought that
0:48:27	these high tech uh
0:48:29	le bar trees could keep
0:48:31	the land in a free but it's absolutely impossible even just you know the plastic where
0:48:36	that uh people use it
0:48:38	uh
0:48:39	it's full of D N A and you just because our in denies everywhere in our environment
0:48:43	uh it's impossible to keep it out
0:48:44	hi
0:48:46	um
0:48:47	so the little bit here about the
0:48:49	the
0:48:50	the various
0:48:51	so uh
0:48:52	where draw so these thresholds
0:48:54	uh that are being used you can see that
0:48:56	the way the evidence is
0:48:57	analysed is quite true
0:48:59	uh but this threshold means anything below this doesn't count
0:49:03	so this he he is very strong evidence could be against individual but that peak is now we have
0:49:08	so tall
0:49:09	because it's uh because of the thresholding affect
0:49:11	but this is
0:49:13	the threshold
0:49:14	for where there's a single peak about this
0:49:16	we assume there's enough
0:49:18	D in a block
0:49:19	that part hasn't dropped out and there's only one only able to true hamas i get
0:49:23	but a single peak below this such as that one there
0:49:26	um
0:49:28	the black one uh it doesn't have a partner
0:49:30	but because it's below the threshold
0:49:32	it's considered the dropout all of this is sort of a battery in very unsatisfactory but it's about this
0:49:36	where at at the moment
0:49:38	uh i would say so much about that case now
0:49:41	'cause i'm running out of time but this
0:49:43	is a zeromean on what the electorate rhymes actually look like
0:49:47	and with these low announced at dinner
0:49:49	it's quite noisy this thirteen liam turned out to be quite important
0:49:53	and at that time and this court case i
0:49:56	on this axis was regarded as the threshold and you can see the audio
0:49:59	the team
0:50:00	reached a peak height of fifty four on this one run up the dozens and dozens of reruns of different
0:50:05	samples from the crime scene
0:50:07	that was the only time that it reached about fifty
0:50:10	but i counted
0:50:11	as a
0:50:11	a full a leo and this
0:50:13	because it's a rare really all turned out
0:50:15	the strongest evidence against this guy
0:50:17	so you can see what
0:50:18	this page here much bigger than that one
0:50:20	is of no evidential value that's just an experiment
0:50:24	a cold start
0:50:25	uh
0:50:26	this
0:50:27	one yeah
0:50:28	and this one here are assumed to be just background noise
0:50:31	uh and so you can
0:50:33	see that it's not quite as sensitive issue about whether that's a real big
0:50:37	um but uh nevertheless it was counted as such
0:50:40	um
0:50:41	and in that
0:50:42	case
0:50:42	there was some three
0:50:45	a labels that shouldn't be in there if the defendant really was
0:50:49	the contributed
0:50:49	sample but one time we have a lot of argument about
0:50:52	how to deal with this
0:50:54	um
0:50:55	and
0:50:57	the standard
0:51:00	um
0:51:01	way of analysing this problem is a kind of version of the random anything you work out you would the
0:51:06	probability
0:51:07	there are a a guy chosen at random in the population would be excluded by the seven
0:51:12	and there's a huge amount of problems with this
0:51:14	you probably gathered i'm not a fan at all
0:51:16	all this approach
0:51:17	uh
0:51:18	and i got hollis
0:51:19	here things that are wrong but i'm sort of rushing out the end of my talk
0:51:23	uh so i won't go
0:51:24	in any uh did how but other
0:51:26	other than that to say that in the
0:51:29	the whole idea of inclusion and exclusion don't apply anymore when we've got a small amounts of D N A
0:51:35	uh
0:51:35	and uh
0:51:37	but
0:51:38	just one of many uh problems with this approach
0:51:41	uh
0:51:42	and how we're gonna talk you through a little bit
0:51:45	of the
0:51:47	how to work through a likelihood ratio in this problem in the way that i would
0:51:51	fig is at least uh
0:51:52	somewhat acceptable
0:51:53	but i want to
0:51:55	i won't go into that so they're all these issues about modelling dropout
0:51:58	um
0:51:59	but i'm going to skip over
0:52:02	um
0:52:04	the
0:52:06	quite important would be low level cases usual masking that you often have D N A from a victim which
0:52:11	is of high level
0:52:12	uh and it could be masking nearly all from uh
0:52:15	from the true uh
0:52:17	perpetrator
0:52:18	um
0:52:19	so we need to take that into account
0:52:21	drop in
0:52:22	uh i've got just some little
0:52:24	simulation results here that showed no matter how much you feel this
0:52:27	to pee wee which is part of the
0:52:29	of the random and I D and always claim to be conservative
0:52:33	it's not
0:52:34	so these probabilities under various
0:52:35	assumptions
0:52:36	sorry likelihood ratios which are smaller than the likelihood ratio one to that
0:52:40	two P rule
0:52:41	uh and i'll skip
0:52:43	all of that oh i see this one is quite interesting if i
0:52:46	oops
0:52:47	it's quite interesting
0:52:49	this is about
0:52:50	what happens
0:52:51	if the crime scene profile is now
0:52:54	and if the defendant is pictures i guess
0:52:57	uh i'd say that's like evidence against
0:52:59	such a big so
0:53:01	the typical position of almost everyone in the field would be to say that if the crime scene profile is
0:53:05	now i'm is nowhere
0:53:06	we can ignore it
0:53:08	uh i say that's why incriminating
0:53:11	because
0:53:12	if you didn't see anything it's more likely that the offender was hedges i guess
0:53:16	and so if you would defend the dispatchers i guess that's like evidence against him
0:53:20	it's like evidence in his favour if he's homozygous
0:53:22	but if there's masking
0:53:24	it can be dramatic yeah evidence in favour of
0:53:26	and
0:53:26	and that sometimes so
0:53:28	not appreciated
0:53:29	um
0:53:30	and
0:53:32	i will
0:53:34	do you have
0:53:36	hesitate just like maybe on this case because
0:53:39	it's sort of remarkable the idea
0:53:41	is sometimes suggested that uh all the problems are solved
0:53:45	in the uh
0:53:46	in the D i haven't
0:53:47	field and this is an example
0:53:48	about what seems to me the most kind of scandalous uh
0:53:52	uh miscarriage of just
0:53:53	as i understood the case
0:53:54	um
0:53:56	the uh
0:53:57	there was another who contributed to the sample
0:53:59	uh and the case revolved around whether this whether or not this
0:54:03	the stuff
0:54:03	fig
0:54:04	it's suspect it contributed was actually true contributor
0:54:07	this is what was seen in the crime scene profile you see several dashes he means nothing was observed
0:54:12	so both contribute is we're very low levels of the you know i
0:54:16	and we have a substantial amount dropped out
0:54:18	this was the sort of random and not excluded probability reported in court
0:54:23	totally one in ninety six thousand
0:54:25	it seems to be convincing enough to the guy to get
0:54:27	i did
0:54:28	uh but if you start looking closely at this
0:54:31	this
0:54:32	and there's some really uh
0:54:34	scandalous things going on here
0:54:35	uh look at this twelve and thirteen that was in the in the crime scene sample
0:54:40	it's exactly the same as the G the type of the node contributed
0:54:43	so arguably this is no evidence at all it's just reflecting the known contributed doesn't tell us anything
0:54:48	uh
0:54:49	but the relevant likelihood ratio used for that locus was six point five
0:54:53	uh because that's what you get from this
0:54:55	random and not excluded for me which is completely illogical
0:54:58	uh and completely miss rate presents the evidence
0:55:01	and uh when i
0:55:03	applied
0:55:04	the sort of likelihood ratio based theory
0:55:07	that i'm talking about so i got some criticisms of the methods here
0:55:10	um
0:55:11	i could
0:55:13	modify the random and not exclude formula to be a bit more reasonable instead might
0:55:18	two thousand i would've got eight
0:55:19	uh but when i did a likelihood ratio calculation
0:55:22	that allows for example evidence to favour the depend
0:55:25	some loci with less than one
0:55:27	um
0:55:28	i come up
0:55:29	instead of ninety
0:55:30	X thousand with a like a racial too
0:55:32	uh this is
0:55:33	you know virtually useless and
0:55:35	i'm the worst uh
0:55:37	uh we can study and i haven't i don't come across this really hardly any
0:55:41	information in in italy three labels
0:55:44	in all of this
0:55:45	that are attributable
0:55:46	to this person and not to that person
0:55:49	so it's really uh kind of uh
0:55:51	shockingly weak evidence
0:55:53	uh
0:55:54	completely
0:55:55	misunderstood and misrepresented in court
0:55:58	uh and the guy was found guilty
0:56:01	so
0:56:02	my uh
0:56:03	conclusion as i said i had hoped
0:56:05	the come back to draw more explicit parallels with
0:56:09	voice problems but i'd i didn't really feel confident uh to do that
0:56:13	uh i
0:56:14	one
0:56:15	tell you that uh
0:56:17	there's a lot of progress being made with D N A evidence situation is much better than it used to
0:56:21	be
0:56:21	uh well as the previous case just shows that a lot still wrong
0:56:25	um
0:56:26	and uh
0:56:27	much remains unsatisfactory
0:56:30	and
0:56:31	much
0:56:32	there are some fundamental problems with the logical approach
0:56:36	that uh
0:56:37	that
0:56:38	to which they i don't think there's ever going to be really sad
0:56:40	actually solution but it nevertheless
0:56:42	provides the most useful framework for
0:56:44	thinking
0:56:46	so i should stop
0:56:46	yeah
0:56:47	area
0:56:47	never
0:56:48	to
0:56:58	very much
0:56:59	you could be going on
0:57:01	or
0:57:02	oh
0:57:02	the i think we can
0:57:04	some
0:57:04	a few minutes to
0:57:06	okay
0:57:06	hmmm
0:57:07	scription
0:57:14	um
0:57:15	i work with
0:57:16	and
0:57:16	consuming no pollution
0:57:18	automatic systems
0:57:20	um
0:57:22	usually we select
0:57:23	no
0:57:23	hmmm
0:57:24	speaker comp
0:57:25	from cool
0:57:27	randomly
0:57:28	one
0:57:29	recuse work
0:57:30	uh_huh
0:57:31	oh
0:57:33	always
0:57:33	yeah
0:57:34	so
0:57:36	so when we do
0:57:38	hmmm
0:57:38	one
0:57:39	to me oh cues
0:57:40	see
0:57:41	solution
0:57:41	uh_huh
0:57:42	fig
0:57:43	equipment we have
0:57:45	from
0:57:46	selection
0:57:47	different speakers
0:57:49	uh
0:57:51	oh
0:57:51	well obviously it's um
0:57:53	it's difficult to get it right and it seems to me that this is you you have to do some
0:57:57	uh
0:57:59	some version of this calibration on the basis of man
0:58:01	no
0:58:03	um
0:58:04	speakers but
0:58:05	the
0:58:05	but but let me see comes to your question i mean the problem is about to leave the limited
0:58:10	selection of comparison
0:58:12	is that what you see
0:58:13	not mutation but um
0:58:16	uh usually
0:58:17	whatever
0:58:19	okay
0:58:19	the question is
0:58:21	should we only
0:58:22	use
0:58:23	different speaker comp
0:58:24	and
0:58:25	you know
0:58:25	evaluations
0:58:26	all speakers
0:58:27	who sound simple
0:58:29	because
0:58:29	the keys
0:58:30	right
0:58:30	the case
0:58:31	no obvious
0:58:32	man
0:58:33	comes to us with
0:58:34	two totally different sounding speakers
0:58:36	also
0:58:37	expertise
0:58:39	uh i see that um
0:58:41	well i i can i i
0:58:43	um
0:58:44	do you
0:58:45	you know the issues that i had to worry about a a quite a distinction is some overlap but there
0:58:49	are fundamental uh
0:58:50	uh
0:58:51	differences and um
0:58:53	i would
0:58:54	that is that's obviously
0:58:57	somewhat uh
0:58:58	unsatisfactory but nevertheless i can see that it's going to sort of
0:59:02	bias you in a difficult
0:59:04	in in a bad direction because this is most
0:59:07	this is the most challenging situation
0:59:09	to distinguish the similar sounding voices
0:59:11	uh
0:59:12	and um
0:59:14	any and by biasing harrington that should be a good bye
0:59:17	i would
0:59:17	oh
0:59:17	it makes it more difficult for you to um
0:59:20	establish
0:59:21	i
0:59:22	or whatever get produce evidence for identity
0:59:26	but what about all this
0:59:27	summation of all
0:59:28	accuracy and precision
0:59:29	please
0:59:31	and
0:59:33	yes
0:59:34	speaker so maybe we want to
0:59:37	during to use
0:59:38	mm
0:59:39	hmmm sounds
0:59:40	maybe
0:59:41	oh yeah
0:59:42	more easy
0:59:44	yes
0:59:45	um
0:59:46	but if you watch trying to distinguish
0:59:49	same source
0:59:50	from different sources if the different sources that you use
0:59:53	i'm different but similar
0:59:56	that makes that a hot a comparison not easy so
0:59:59	that's what i was suggesting that um
1:00:02	there should be a
1:00:03	yeah that's this
1:00:04	it's
1:00:04	it's
1:00:05	it's good that you you have to do one
1:00:08	something like we would like what you were doing it would be nice to have
1:00:12	probably well designed experiments where you have uh
1:00:15	speakers that are similar and
1:00:17	speakers that are more different
1:00:18	uh and you can see the range of differences um you know i have emphasised a lot roller
1:00:23	relatedness for D N A evidence but i don't know well
1:00:27	you know of relative you know distinguishing
1:00:29	brothers speaking for example whether that's harder than for unrelated individuals
1:00:34	um but it
1:00:35	um
1:00:36	so
1:00:37	ideally you'd like to
1:00:38	be able to consider all those
1:00:40	yeah
1:00:41	i
1:00:42	when you say um
1:00:43	you might be overstating the precision
1:00:45	but ultimately
1:00:46	you want to process your
1:00:51	uh
1:00:52	and if you've given yourself harder task
1:00:54	do
1:00:55	by having the different speakers being somewhat simple
1:00:58	i like it
1:00:59	actually
1:01:00	the other way around
1:01:00	thing
1:01:01	which
1:01:02	you should
1:01:03	two
1:01:03	however
1:01:03	false
1:01:04	you know
1:01:05	action
1:01:05	so we can be more
1:01:07	sure
1:01:07	no
1:01:11	yeah okay maybe i missed something the problem 'cause it does seem to me harder task if you had very
1:01:14	different sources you could distinguish them quite easily
1:01:17	uh and so that's an easy task
1:01:19	if you have similar sources trying to distinguish them is hard so you have given yourself harder task it seems
1:01:24	to me on this i've missed something
1:01:25	problem
1:01:28	maybe maybe we can chat a bit more later and i guess the bottom of this
1:01:39	and
1:01:40	i think if we don't
1:01:41	yeah
1:01:42	my question
1:01:43	tries to the conventional you should be using
1:01:46	or
1:01:47	so
1:01:47	yeah imagine we we
1:01:49	yeah but uh
1:01:50	speech lab
1:01:51	which analyze the highest you
1:01:53	to present evidence
1:01:54	well
1:01:55	and is
1:01:56	which would be that uh we have a
1:01:58	with the recording
1:01:59	maybe
1:02:00	along
1:02:01	um
1:02:03	one
1:02:04	um we have
1:02:05	we we are able to estimate multiple
1:02:08	like
1:02:08	reissues
1:02:10	yes
1:02:10	often
1:02:12	do you have
1:02:13	you know actually you know not
1:02:15	you would
1:02:16	imagine that
1:02:18	those
1:02:18	one
1:02:19	yeah
1:02:19	um
1:02:21	reasonable
1:02:23	uh
1:02:24	and correlation with them but you cannot
1:02:26	okay
1:02:27	independent
1:02:28	you like
1:02:29	to to make the problem but you
1:02:31	kind of
1:02:32	um
1:02:32	uh those uh
1:02:34	multiple issues
1:02:36	all of them
1:02:37	small values maybe ten
1:02:39	and
1:02:40	more than one thousand
1:02:41	um
1:02:42	but
1:02:44	the idea is
1:02:45	how to present that
1:02:46	i didn't see what would you have
1:02:47	multiple
1:02:49	different
1:02:50	uses
1:02:51	i mean it was
1:02:52	yes
1:02:53	and
1:02:54	do you cannot
1:02:55	the proof
1:02:56	yeah
1:02:57	this do you
1:02:58	time
1:02:58	hmmm
1:02:59	the independence
1:03:01	those
1:03:03	yes
1:03:03	well
1:03:04	okay that is it
1:03:05	an interesting and uh
1:03:07	difficult question and i
1:03:09	i feel
1:03:10	instinctively as i was saying earlier that um
1:03:13	you know what
1:03:13	given
1:03:15	the right framework and some independence assumption you know or to be uh
1:03:19	more or less reasonable and you can never prove independence it does always uh you know
1:03:23	like people
1:03:24	spent a long time trying to prove
1:03:26	independent self
1:03:28	or
1:03:28	different labels in D N A profiles
1:03:30	it's a
1:03:31	it's a few tile um
1:03:33	the size
1:03:35	all ultimately
1:03:36	but um
1:03:38	so but if you if
1:03:40	if
1:03:41	dependence
1:03:41	is a really serious problem i am M
1:03:44	i'm just trying to think
1:03:46	i need i need to understand what the dependence structure is to
1:03:49	so really help anybody dependence is a real
1:03:52	and
1:03:53	i'm
1:03:54	to mount a problem then i think you're stuck really i don't i really can't see how
1:03:58	um
1:03:59	to make use
1:04:00	of the
1:04:02	multiple level
1:04:03	because obviously you know if you did have independence you can multiply likelihood ratios and everything is uh
1:04:08	it's uh
1:04:09	is easy
1:04:11	the um or
1:04:12	oh
1:04:13	yeah
1:04:14	so
1:04:14	the analogy with um
1:04:17	with D N A evidence is that is that the relatedness is the right
1:04:20	the condition on
1:04:21	things become uh
1:04:23	uh
1:04:23	can be independent once you've got the right
1:04:26	um
1:04:27	uh
1:04:28	conditioning but in general
1:04:30	um
1:04:31	this
1:04:32	kind of model where there's some kind of latent variables so essentially
1:04:36	relatedness is a latent variable
1:04:38	uh and some kind of model you i
1:04:40	where
1:04:41	there is
1:04:42	a latent variable that
1:04:44	and
1:04:45	encapsulate the common features of the different recording
1:04:48	to generate
1:04:49	pen
1:04:50	um
1:04:51	if you can
1:04:53	uh condition on that latent variable and then integrated out in some way would deal with it in some appropriate
1:04:58	way
1:04:59	uh you know i feel as if there should be some modelling approach like that
1:05:02	that would work and allowing to
1:05:04	then make
1:05:05	and it depends
1:05:06	assumption i mean just you know in general modelling
1:05:08	dependent data disk
1:05:10	and of um
1:05:11	random effects models type things work
1:05:14	well
1:05:14	and i would go back
1:05:16	you have you you have to explore
1:05:17	to the extent where you can be
1:05:19	reasonably confident about independence assumption and give
1:05:22	some good arguments for it i mean i've always
1:05:25	no i can still never
1:05:28	ruth independent any of the independent sign
1:05:30	assumptions i make for the D N A profiles
1:05:32	but i just tried walking from uh
1:05:35	from reason that
1:05:36	you know relation is just you know
1:05:38	oh
1:05:39	and if we model that we should have sold
1:05:41	and
1:05:41	problem
1:05:42	and i i would have thought that some kind of a venue like that
1:05:44	the only option
1:05:46	well about can can you tell me briefly what is the cause of the dependence that uh
1:05:50	the deepens is what you see yeah we have
1:05:53	as for
1:05:53	two can be analysing different phones
1:05:56	syllables
1:05:57	personally depends
1:05:59	well then
1:06:00	have different characters
1:06:01	but a lot and come from the same source
1:06:03	yes
1:06:04	but once you conditioned on it being the same source
1:06:06	right
1:06:06	yeah
1:06:07	uh anyway i think there's a modelling answer but if you really can't
1:06:11	tackle
1:06:12	the dependence with some kind of modelling and so than that and i think it is you know you do
1:06:16	have a real fundamental problem because ultimately gonna say well
1:06:19	if there is dependent stay then how big could it be and uh
1:06:22	if you can't really quantify that in some way then
1:06:25	i don't think you can usefully give multiple likely
1:06:27	ratios and
1:06:28	the
1:06:29	court will figure it out
1:06:33	you've got to do the work
1:06:37	but you do with a
1:06:38	people like working tonight
1:06:40	for historical reasons B C we should emulate D N A yeah
1:06:44	huh
1:06:45	you've gone through to
1:06:46	oh you're all the problems of the yeah yeah
1:06:48	oh i i did i think you're doing exactly the right thing i wouldn't disagree with the strategy that order
1:06:52	that's what i wanted to ask them what you
1:06:55	yes
1:06:55	should we still be saying we should ideally i i
1:06:58	i think so yeah something where a lot for all these
1:07:01	i mean i have to say difficulties remain otherwise on out of a job
1:07:04	and the uh
1:07:05	um
1:07:07	uh about
1:07:09	it is true but we definitely much better off than we were ten years ago
1:07:12	and uh
1:07:13	it's a bit like
1:07:14	when you're teaching
1:07:16	well almost anything
1:07:17	in effect you tell the second year class to get everything we told you last year that was an over
1:07:21	simplified version of the problem here is that
1:07:24	here is the real
1:07:25	them and then in the third year you tell the students to get everything we told you last year that's
1:07:28	an over simplified version of the problem here is the real thing
1:07:31	uh and uh
1:07:33	i i
1:07:34	i mean i can remember
1:07:35	now there are some things where i do actually literally tell the students that and uh
1:07:39	and i think
1:07:40	that
1:07:41	um
1:07:41	you have to
1:07:43	you know get
1:07:44	except
1:07:45	it's from the community by focusing on the
1:07:47	on
1:07:48	simplified versions
1:07:49	uh
1:07:50	and
1:07:51	it is a step forward and then there will always be
1:07:53	you know you never going to
1:07:56	overcome all the proper
1:07:57	just
1:07:57	but is actually interesting that um you know the way these various complications that i've talked about many of them
1:08:02	i don't think you do have an awards for it in many cases you're better off
1:08:06	uh we i think
1:08:08	you suggested to me in conversation that
1:08:10	you know we have these population genetics models
1:08:13	and that so
1:08:14	and basically this thing i'm talking about that
1:08:16	all the dependence comes from relatedness and once we
1:08:19	concludes condition on the right level relatedness we can get rid of the dependence
1:08:22	and i agree that's a good point but there's a lots of
1:08:25	subjective ms
1:08:27	in those models
1:08:28	um
1:08:29	and
1:08:30	we have
1:08:31	many other
1:08:33	problems that i was describing to you that i don't think you do have any
1:08:37	and the analogy for so
1:08:39	in many ways i think
1:08:41	the grass is greener on your side of
1:08:42	fans i
1:08:45	yes
1:08:47	well i would like to
1:08:48	thing but again
1:08:51	i
1:08:56	we should do this

Interpretation of DNA evidence as a paradigm for speaker recognition

INVITED TALKS

Added: 14. 7. 2010 11:08, Author: David Balding (UCL Genetics Institute, London), Length: 1:08:57