Machine learning (ML), a pivotal area within the field of artificial intelligence, has permeated various aspects of our daily lives. While not always aware, we engage with ML algorithms routinely – from web searches, targeted advertising, and spam filtering to driving autonomous vehicles, analyzing complex bioinformatic data and asking ChatGPT for advice. In forensic science, the adoption of ML is rapidly expanding, offering innovative approaches to a range of forensic challenges. However, there exists a knowledge gap: many forensic scientists are not fully versed in the capabilities and limitations of ML, while ML specialists may not be familiar with the unique demands of forensic applications. This presentation attempts to bridge this gap by introducing ML methods and their application in forensic DNA analysis, with a focus on the HID area. We will explore how ML methods can streamline the manual analysis of intricate DNA data, ensuring both accuracy and reproducibility, delve into enhanced prediction accuracy of phenotypic traits for forensic intelligence, address the challenges in transparency and validation of ML methods prior to casework implementation, and discuss other pertinent topics. The integration of ML in forensic DNA analysis represents a significant advancement, promising more efficient, accurate, and objective analytical methods. Nevertheless, this integration also presents challenges that must be addressed to garner trust and acceptance within both the forensic community and the legal system.
You Might Also Like
0:00
[MUSIC PLAYING]
0:05
Hello, I am Chantal Ragh.
0:14
I am a software and algorithms developer
0:17
at Thomas Fisher Scientific.
0:20
The field of artificial intelligence and machine
0:23
learning is not only of great interest to me personally,
0:28
but it is also an important development
0:30
to focus here at Thomas Fisher Scientific.
0:34
And so I'm very excited to introduce our next speaker,
0:39
Dr. Mark Bresch.
0:42
Dr. Mark Bresch is an associate professor and co-director
0:47
of the forensic science program at San Jose State University
0:53
with over two decades in the field,
0:56
including 10 years of operational experience.
1:00
He has significantly contributed to the field
1:04
of forensic DNA analysis in teaching, research, and case
1:10
work.
1:12
Dr. Bresch's research interests ban
1:15
multiple disciplining areas, particularly focusing
1:19
on employing machine learning and bioinformatics
1:23
to predict physical characteristics from DNA samples,
1:28
showcasing the powerful intersection
1:32
of artificial intelligence and forensic genetics.
1:37
He will be talking about unlocking justice,
1:40
the AI revolution in forensic science.
1:44
Great, everyone.
1:45
And thanks to San Jose State Scientific for the invitation
1:48
to give a talk at Heeds 2024.
1:50
My name is Mark Bresch.
1:52
And I'm a professor and forensic science program coordinator
1:54
at San Jose State University.
1:56
In this talk, I would like to introduce you
1:58
to the topic of artificial intelligence and machine learning
2:01
to make sure that you are aware of the AI revolution already
2:05
happening in forensic science, and particularly
2:08
in the field of forensic DNA analysis.
2:10
Imagine walking into a room where a call case
2:13
left unsolved for decades is laid out on the table.
2:16
The evidence is there, fingerprints, DNA profiles,
2:20
photos of the scene.
2:21
But for years, these puzzles remain unsolved.
2:24
Now, imagine we have a new key to unlock this mystery,
2:28
a key forged from the advanced algorithms
2:30
of artificial intelligence.
2:32
In the future, where forensic science
2:34
reveals the story of behind every trace found at a crime scene.
2:39
This future is much closer than you might think.
2:42
AI and machine learning are not just revolutionizing
2:46
the way we live, work, or think.
2:48
They are transforming the very fabric of forensic science.
2:53
Today, I'm going to take you on a quick journey
2:55
into the heart of this transformation.
2:57
But first things first.
2:59
AI is a distinct branch of computer science
3:02
and engineering that focuses on creating technologies
3:05
capable of predictive analysis by learning and executing
3:09
cognitive tasks that typically require
3:11
human intelligence.
3:13
These tasks include decision making, visual perception,
3:16
language understanding, and more.
3:19
At the same time, machine learning
3:20
is a range of powerful computational algorithms
3:23
representing a subfield of artificial intelligence
3:27
capable of generating predictive models via intelligent,
3:31
autonomous analysis of relatively large
3:33
and often unstructured data.
3:35
Machine learning algorithms are designed to train
3:38
a computer program how to learn from experience
3:41
and improve over time.
3:43
In a way, similar to how a child
3:45
learns new skills from their everyday experience.
3:50
AI's impact spans across sectors,
3:53
penetrating almost every aspect of our lives.
3:55
And forensic science is no exception.
3:58
Like the private sector, many federal agencies
4:00
are expressing interest and investing in AI-enabled
4:03
technologies that may be capable of optimizing resources
4:07
and expanding capabilities across many industries.
4:10
For forensic science service providers,
4:12
AI-enabled technologies represent
4:14
a significant opportunity to improve the way they identify,
4:18
analyze, and reach conclusions on forensic evidence.
4:21
However, its integration here is nascent,
4:25
often limited by the gap between forensic experts
4:28
acquaintance with machine learning capabilities.
4:31
Yeah, the potential is a mess from enhancing evidence
4:35
analysis to unlocking call cases.
4:38
So let's first demystify what machine learning really is.
4:42
Imagine training dog, you reward four correct actions.
4:45
It learns.
4:46
This is similar to supervised learning,
4:48
where an algorithm learns from labeled data,
4:51
often called training data.
4:53
Unsupervised learning on the other hand
4:55
is like observing birds in nature
4:58
to understand their behaviors without professional guidance.
5:02
Semi-supervised learning mixes both using a little labeled
5:05
data to guide the learning from a larger pool of unlabeled data.
5:09
And finally, reinforcement learning
5:12
is akin to teaching a child to ride a bike
5:15
learning from trial and error.
5:17
Now, most AI systems are machine learning base,
5:20
where machine learning algorithms use inferences
5:23
derived from data to find correlations and importance
5:27
that allow them to make predictions about similar new data.
5:32
To better understand the basic principles
5:34
behind different methods of machine learning,
5:36
let's think of machine learning as a versatile shaft
5:40
in a kitchen full of ingredients, our data,
5:43
where each dish represents a task or a problem.
5:47
Obviously, our machine learning shaft
5:49
would require a different cooking technique for each dish.
5:53
So, classification problem is like soaring ingredients
5:57
into beans, fruits into one and vegetables in another.
6:01
An example of classification approach would be
6:04
distinguishing handwriting styles where the data set
6:08
for classification includes examples of handwriting
6:11
and the prediction categories are individual letters or words.
6:15
Regression predicts how long it takes to bake a cake
6:19
based on past cake baking experiences.
6:21
An example is an estimation of the post-mortem intro
6:24
by analyzing decomposition rates based on correlations
6:28
found in previously collected data.
6:31
Clustering aims to group similar things together,
6:34
like sweet fruits or savory spices.
6:37
And for instance, would be clustering hair samples
6:40
based on morphological features.
6:43
Dimensionality reduction simplifies recipes
6:46
by removing unnecessary steps while keeping the dish delicious.
6:50
In the area of forensic analysis,
6:52
dimensionality reduction is often applied
6:54
in order to predict the bi-geographical ancestry of a person.
6:59
So here, highly dimensional data can vary from a few dozen
7:03
single nucleotide polymorphisms to over a million
7:07
snips per sample.
7:08
These data are used to two or three dimensions,
7:11
which explain most of the variance in the data
7:14
and can then be analyzed or visualized much easier.
7:18
At the same opportunity, I want to mention a few advanced
7:22
machine learning approaches to these problems
7:25
and explain them using highly specialized
7:27
master-shaft analogy.
7:29
So natural language processing approach can be visualized
7:33
as a chav that understands and follows a recipe
7:36
with written in any language.
7:39
Generative models is a highly experienced chav
7:43
that invents new recipes after tasting
7:46
and studying countless dishes.
7:48
These master-shaft isn't just copying.
7:50
They are innovating by introducing highly creative dishes
7:54
that could easily get a Michelin star.
7:57
We'll get back to these generative models shortly.
8:00
Now, let's see how AI is already transforming
8:04
the landscape of forensic investigations.
8:07
In the realm of forensic science disciplines,
8:09
machine learning algorithms can be trained
8:11
for automated identification, classification
8:14
and verification of fingerprints,
8:17
shoe prints and bullet striations.
8:19
Additionally, speech recognition algorithms can transcribe
8:22
and analyze audio recordings, identifying speakers
8:26
or detecting keywords, while video analysis algorithms
8:29
can automatically recognize faces, track movements,
8:33
perform gate analysis and even analyze the behavior
8:36
of individuals in a footage.
8:38
To perform the predict analysis of this complex data,
8:41
machine learning models must be trained on very large
8:45
data sets to accurately recognize unique features
8:48
and importance enabling more objective analysis
8:51
of impression and pattern evidence
8:53
based on quantitative and qualitative data.
8:57
While tools like Chad GPT showcase AI's utility
9:01
in summarizing complex information
9:03
and generative models and natural language processing
9:05
offer new ways to approach forensic evidence,
9:08
the true potential lies in AI's
9:11
augmenting human investigators effort.
9:14
AI can essentially be a part of the investigative team
9:19
collaborating with human partners in a brainstorming process
9:23
helping to uncover new leads and perspectives
9:27
that might have been overlooked.
9:29
AI algorithms can extract comprehensive information
9:32
from various types of forensic evidence
9:34
by analyzing multi-dimensional data points
9:37
and finding hidden patterns
9:39
leaking between traces, crime scenes and people.
9:42
Notably, the application of the tools offers an opportunity
9:46
not to only arrive at a decision as to whether the crime scene
9:51
evidence matches the references,
9:53
but also to quantify the similarity between items
9:56
and calculate the likelihood of that match
9:59
being true.
10:00
This would improve standardization
10:02
and reduce the inherent subjectivity of forensic pattern
10:06
analysis.
10:07
And perhaps most importantly, this exemplifies
10:09
a truly holistic approach to forensic evidence,
10:12
a critical missing link that has been highlighted
10:15
in new research studies.
10:17
I know what you're thinking.
10:18
No, AI is not going to totally replace human investigators.
10:22
AI can indeed process and analyze vast amounts of data
10:27
at speeds unattamed by humans.
10:29
At the same time, human investigators
10:31
bring critical thinking, intuition, creativity
10:34
and investigative experience that AI cannot replicate yet.
10:39
Therefore, the ideal approach is synergistic,
10:42
where AI provides insights and suggestions
10:44
based on data analysis, while human investigators
10:47
can then dull deeper into these leads,
10:50
using their expertise to verify, expand and act
10:53
upon the AI generated insights.
10:56
In fact, I'm 100% sure that in a couple of years,
11:00
we'll see most law enforcement agencies
11:02
making their own generative AI systems
11:06
capable of consolidating big data
11:08
from hundreds of thousands of criminal cases
11:11
along the AI detective to find hidden importance,
11:15
linking multiple crime scenes and people,
11:18
kind of a 24/7 personal shallow homes or her cool poor
11:21
or whoever's your favorite detective.
11:24
No doubt, this approach will revolutionize
11:27
the investigation process.
11:29
At the same time, it is important to remember
11:31
that even though AI helps reveal hidden importance in data,
11:36
it does not ultimately reveal why these data
11:39
may be related in a way people can easily understand.
11:43
We'll get back to these problems shortly.
11:45
Now, let's dive into the core of our discussion
11:48
machine learning applications in forensic DNA analysis.
11:51
So, current applications of machine learning
11:54
can be arbitrary divided into several areas.
11:57
Applications related to human identification,
12:00
applications related to forensic intelligence
12:02
and those related to increasing the evidential value
12:06
of biological evidence.
12:08
Now, as you know, the contemporary process
12:10
of SCR geotyping and data interpretation
12:13
consists of numerous types, while the complexity
12:16
of separating the air alleles from background nose
12:20
and artifacts is essentially the basic problem
12:22
in forensic DNA analysis.
12:24
Currently, the process of SCR geotyping is carried out,
12:27
semi-automatically using dedicated expert software
12:31
such as gene member.
12:32
These software separates a little pics from nose
12:35
and artifacts with the help of built-in algorithms
12:38
and utilizes a number of validated thresholds
12:41
resulting in the detection and removal of studies,
12:44
pull-ups and other artifacts.
12:46
The resulting DNA profile can be compared
12:49
with a reference profile or further analyzed
12:52
by probabilistic gene type and software.
12:54
However, the overall process is tedious
12:57
and prone to significant variability
13:00
in interpreting complex DNA mixtures.
13:02
Furthermore, with the rapid implementation
13:05
massively parallel sequencing,
13:06
the generated bioinforming data becomes even more complex
13:09
requiring trained personnel and significant resources.
13:13
The problem of accurately classifying
13:16
such a diverse range of variables
13:19
might be better addressed by a more sophisticated
13:21
machine learning approach, deep learning.
13:25
Deep learning methods utilize variations
13:27
of a hierarchical organization of artificial neurons
13:30
with connections to other neurons
13:33
similar to the brain structure.
13:35
These neurons pass a signal to other neurons
13:38
based on received input,
13:40
a process that can be repeated multiple times
13:43
which ultimately creates a complex network capable
13:46
of intuitive learning and artificial neural network.
13:50
Using the culinary world analogy again,
13:54
think of deep learning as a master shaft
13:57
who trains a team of specialized sous shafts,
14:00
our neural networks to handle complex
14:02
and layered recipes, our data.
14:05
So each sous shaft focuses on a specific task
14:08
like chopping or baking,
14:11
learning from every dish they make.
14:13
As dishes be passed through the kitchen,
14:15
each shaft adds the expertise
14:17
resulting in a sophisticated final dish.
14:20
Similarly, in the artificial neural network,
14:23
each layer of neurons performs transformations
14:26
using weights, biases and activation functions.
14:29
These process continues until the data reaches
14:31
the output layer where final predictions are generated.
14:35
Another puzzle well suited for machine learning approach
14:39
is the interpretation of complex DNA mixtures,
14:42
especially if one or more of the contributors
14:45
is represented by a low level portion of that mixture
14:48
and if the DNA molecules are degraded.
14:52
In such a common forensic situation,
14:54
a true allele can be confused with various artifacts.
14:57
All these variables may be influencing
15:00
one of the key parameters in DNA mixture interpretation
15:03
estimating the number of contributors.
15:06
Now machine learning offers the tools
15:08
to piece this puzzle together,
15:11
distinguishing between the true alleles and artifacts
15:14
resulting in identifying even low level
15:17
individual contributors with unprecedented accuracy.
15:20
First, instead of using static analytical
15:23
and stochastic thresholds,
15:25
the machine learning algorithms do this
15:26
by utilizing either a dynamic threshold
15:29
or refrain from applying any thresholds.
15:32
At the same time, they can efficiently recognize
15:35
various artifacts and remove them.
15:37
This way, machine learning methods can utilize
15:39
all the available information in the raw EPG
15:42
or sequencing data to learn and make informed predictions
15:47
maximizing the information obtained from the DNA evidence.
15:51
One such example is PACE.
15:53
This machine learning software has demonstrated
15:56
remarkable accuracy in predicting the number
15:58
of contributors for single source samples
16:01
and mixtures with up to four contributors.
16:04
Furthermore, by combining various methods of machine learning,
16:07
it is possible to build even more powerful models
16:10
that could be used to analyze a variety
16:13
of very complex incoming information.
16:16
An example of such an approach is generative models
16:19
which I mentioned earlier.
16:21
Generative models often incorporate layered networks
16:24
that utilize multiple levels of non-linear data
16:28
processing to extract and transform features of interest.
16:32
One popular example is generative adversarial networks,
16:35
GANs. The key innovation behind GANs
16:38
is their two network architecture consisting of a generator
16:42
and discriminator networks.
16:43
The generator creates synthetic data
16:46
and the discriminator evaluates whether the data is real
16:50
or fake, leading to a competitive training process
16:53
where the generator becomes increasingly skilled
16:56
at producing highly realistic data.
17:00
One of the most intriguing applications of ANs and GANs
17:04
is the prediction of an individual's visual appearance
17:07
from a DNA sample producing investigative leads
17:09
where traditional means are unsuccessful.
17:13
This task is exceptionally challenging due to
17:16
really limited knowledge of the underlying genetic
17:19
and epigenetic architecture of these complex visible traits.
17:23
Therefore, GANs models are expected to lead to more accurate
17:27
predictions of these traits, especially facial appearance
17:30
as demonstrated by a few recent proof of concept studies.
17:34
Machine leering also comes to the rescue
17:37
by imputing missing genotypes in partial DNA profiles.
17:40
These imputation approach may be especially helpful
17:43
in forensic genetic genealogy because these processes
17:47
often involve generating and interpreting SNP data
17:51
and estimating genealogical relationships
17:53
in circumstances where the resulting SNP profile
17:56
often contains numerous errors or gaps.
18:00
Moving to the forensic microbiology area,
18:02
machine leering can help analyze the complex microbiome
18:06
formed on a decomposing body to estimate
18:09
the postmortem interval more accurately.
18:12
Additionally, machine leering algorithms can now confirm
18:15
the biological tissue source and estimate biological age
18:19
from epigenetic data producing valuable forensic intelligence.
18:23
These information would add layers of context
18:26
to the evidence and help to produce an insights
18:30
into how DNA was deposited on the object
18:33
assisting with activity level reporting.
18:35
And finally, AI can be utilized in forensic labs
18:40
to streamline the genotyping pipeline logistics
18:43
and reduce the overall cost and time of analysis
18:46
by incorporating automatic intelligent trash and stab
18:50
helping the analysts to decide which samples
18:53
are worth processing further and which would most likely
18:56
be a waste of resources.
18:58
So by now, I'm sure you can agree that
19:01
into integration of AI machine leering clearly represents
19:05
a paradigm shape from traditional forensic methods.
19:09
It has the potential to revolutionize
19:12
the analysis of forensic traces,
19:13
streamlining the process and even more importantly,
19:16
making it more accurate and sanitized.
19:20
However, as with any powerful tool,
19:23
there are several limitations related
19:25
to AI implementation that the forensic law enforcement
19:29
and legal communities should take into consideration.
19:32
By then nature, forensic applications demand
19:35
unparalleled accuracy and reproducibility areas
19:38
where current AI tools, despite their potential,
19:42
present significant challenges.
19:43
A core issue with these technologies
19:46
is their black box nature.
19:48
This complexity possess a particular challenge
19:51
for non-computer science specialists,
19:53
legal professionals and jury.
19:55
Moreover, the data used to train these tools
19:59
often encapsulates another layer of complication.
20:02
Issues such as non-representative datasets
20:05
can introduce systemic biases,
20:07
skewing results and potentially undermining
20:10
the availability of these methods in forensic analysis.
20:14
Therefore, one of the essential conditions
20:16
for efficient learning of these systems
20:18
is the large size and diversity of a training set.
20:21
Traditionally, these data often require extensive processing
20:25
for a more efficient classification via supervised learning.
20:28
As a result, the training data must be curably produced
20:31
and maintained because biases, inaccuracies
20:35
and lack of diversity in the training set
20:37
will be reflected by, that's right, bias,
20:40
poor accuracy and spurious correlations in the output,
20:43
as we all know, garbage and garbage are.
20:46
A promising avenue in addressing the black box dilemma
20:51
is the concept of explainable artificial intelligence.
20:54
In short, this approach offers a visualization
20:58
of the inner processes aiming to make the explanations
21:02
as intuitive and informative as possible.
21:05
This method significantly improves the explainability
21:09
of predictions by ensuring AI predictions
21:11
are grounded in realistic and plausible scenarios.
21:15
Using vast, well curated databases,
21:19
developing more transparent and interpretable,
21:22
in other words, wide box algorithms
21:24
and their extensive validation can increase confidence
21:28
in the results produced by these systems
21:30
and ensure their acceptability in legal context.
21:34
At the same time, it's equally important
21:36
to maintain privacy and ethical standards,
21:39
particularly when dealing with sensitive genetic information
21:42
and personal data.
21:44
Therefore, collaboration between AI developers,
21:47
forensic experts and legal professionals,
21:49
is essential to develop transparent,
21:52
equitable and ethically sound AI tools
21:55
for forensic investigations.
21:57
So, let's summarize.
21:59
The integration of AI and machine learning methods
22:02
in forensic science represents a powerful
22:04
and irreversible progression.
22:06
AI's ability to integrate and analyze
22:09
multi-dimensional forensic evidence
22:11
offers a revolutionary approach to DNA,
22:14
mixture interpretation,
22:15
important evidence analysis and more.
22:17
This symbiotic approach promises to enhance
22:21
investigative strategies offering new insights
22:24
and interpretation of evidence.
22:27
At the same time, we should be aware
22:29
of the potential consequences of AI implementation,
22:32
which may lead to overreliance
22:35
on the AI-dominated evaluation of evidence
22:38
and even letting the AI tools reach conclusions
22:41
without human experts oversight or intervention.
22:45
As our official intelligence continues to shape
22:48
the landscape of forensic science,
22:50
the forensic science service providers' leadership
22:52
should stay informed.
22:54
All these emerging tools develop training resources,
22:57
consider partnerships with academia
22:59
and plan to use cases and admissibility.
23:03
And finally, I wanna finish with a call for action.
23:08
As we stand on the brink of this AI revolution
23:10
in forensic science, let us embrace AI
23:14
and machine learning not just as tools,
23:17
but as allies in our quest for scientific truth and justice.
23:22
Let's be the architects of this new era
23:25
and actively contribute to the development
23:27
of transparent ethical AI tools
23:30
often a beacon of hope for unsolved cases
23:33
and people who were wrongly incarcerated.
23:36
Thank you.
23:38
- Thank you so much for this super interesting presentation.
23:41
This is such an exciting topic.
23:43
I have a few questions that the audience
23:46
might also be curious about.
23:48
So what kind of AI-proof skills
23:52
would the next generation of forensic professionals
23:55
have to equip themselves with so they won't stay behind?
23:59
- Yeah, thanks Chantal and thanks for this question.
24:04
It's a great question.
24:04
It's very important that the next generation
24:06
of forensic scientists, and I would say even
24:09
the current generation, really continues their education
24:12
and learns at least the foundations of machine learning
24:17
and AI to keep up with all the developments.
24:22
And these area is developing so rapidly as we see.
24:26
So it's really important to learn the basics
24:32
of machine learning, the foundations
24:35
and measure that we humans remain relevant
24:41
with these AI revolution because that's something
24:43
that I hear a lot from my colleagues that are kind of afraid
24:48
that well, AI gonna just replace that very soon.
24:53
And as I said in my talk, I don't really think
24:55
that's gonna be the case because we still, by we,
24:58
I mean humans still have these skills like creativity, right?
25:03
And experience, which I don't think AI has
25:10
and will have at least in the near future.
25:14
So things like critical thinking, creativity
25:19
and obviously a bind from X skills,
25:25
these are really, really critical in these journeys.
25:28
So it's really important to learn these skills
25:31
to ensure that we keep up with all these developments.
25:36
- Sounds good, thanks a lot.
25:37
That means a lot of friends.
25:39
Given the significant advancements in machine learning,
25:43
how do you foresee its integration evolving
25:46
in forensic DNA profiling,
25:49
especially regarding SDR genotyping
25:52
and DNA mixture analysis?
25:55
- Machine learning duties, very powerful capabilities
25:58
to handle this very large data sets,
26:02
very diverse data sets, I think could be very useful
26:06
in very different areas, but specifically
26:09
we've talked about S years,
26:10
that's analysis of complex mixtures, right?
26:13
So this, this area is one of the most controversial areas
26:18
I would say in forensic DNA analysis.
26:22
And especially, you know, estimating the number
26:26
of contributors in those mixtures
26:28
and especially providing a statistical way
26:33
to the low level contributors, right?
26:36
So machine learning methods can not only streamline
26:40
these processes, right, to make it easier for human analysts
26:44
but also sanitize these processes to make sure
26:48
that, you know, different labs essentially produce
26:51
the same output, right, which is very important.
26:54
And on the same token, I think it's very important
26:58
to make sure that these machine learning systems
27:02
are validated before they can be implemented
27:05
in forensic DNA analysis.
27:06
So that's one of the key issues
27:09
and to make sure that they're admissible in court.
27:12
So all these results should be admissible in court, of course.
27:15
And as I mentioned in my talk,
27:18
these tasks requires a collaborative effort, right,
27:23
of forensic professionals, academia and legal professionals
27:28
working together to make sure that indeed
27:30
we moving into the right direction and taking,
27:34
I would say, baby steps towards these, to this aim.
27:38
- So machine learning models, they can be complicated, right?
27:43
And they can introduce potential bias,
27:46
as you mentioned in your talk.
27:48
So what steps to believe are crucial
27:52
in maintaining reliability in forensic DNA analysis?
27:56
- Yeah, that's a very great, a very important question.
27:58
So the reliability of analysis based on machine learning methods
28:04
is essentially as you said, based on data
28:06
that we provide these systems.
28:08
So it's important that these data is diverse and large enough
28:13
to essentially encapsulate all the different types
28:16
of all the variability in these variables, right?
28:19
So, and another important thing is that
28:24
these datasets must be well curated, right?
28:27
So they must be updated regularly to, again,
28:31
include all the new things, all the diversity.
28:36
And only this way we can ensure that the output produced
28:40
by these machine learning systems is a reliable, right?
28:44
So we need to ensure that the data that is used
28:47
to train these systems is accurate,
28:51
is large enough and diverse enough.
28:54
So this way we can ensure that the output
28:58
would be indeed reliable and accurate.
29:01
- Thanks a lot.
29:02
I think this is an important topic.
29:03
It's really important.
29:04
So last question, machine learning
29:07
in forensic DNA profiling offers great potential, right?
29:11
I see half explained really well,
29:13
but it also comes with complexities.
29:15
So how do you propose that we address the challenges
29:19
of integrating these sophisticated technologies
29:22
into the forensic field without compromising
29:26
the integrity and reliability of forensic evidence?
29:30
- Thanks for these great questions, Chantal.
29:32
And let me answer by just saying three things.
29:36
So these systems must be reproducible,
29:40
independently validated and transparent.
29:44
I think these three things are the most important
29:47
to make sure that these, their results are accurate
29:50
and approved by the legal community.
29:53
As with any new tool, especially such a powerful tool,
29:56
it's really important to look into the ethical side of it.
30:00
First of all, at least from my perspective,
30:02
I think that's the most important part.
30:04
And apply these tools reliably and ethically.
30:08
So it may affect some communities
30:11
in very disproportional manner.
30:13
And, you know, we all know all these examples.
30:16
So it's really important that when before
30:18
we really implement these such powerful tools,
30:21
we really make sure that we know their limitations,
30:26
right? And we have also covered the legal side
30:30
and we know what to expect, right?
30:33
So, and we know that it won
30:35
disproportional effects.
30:37
People who already are affected by the criminal
30:42
legal system, by the criminal justice system, sorry.
30:47
- Thank you very much, Mark, for these great insights.
30:51
There's a lot of things to think and talk about in the future.
30:54
Thanks again.
30:55
- Thank you so much, Chantal, for this invitation.
30:57
It was my pleasure.
30:58
Thank you.
30:59
(upbeat music)
31:01
- With the seek studio flex genetic analyzer
31:08
and GMapper IDX software version 1.7,
31:11
the number of overall edits during analysis are decreased
31:14
due to the default algorithms.
31:16
This includes the auto spectral calibration
31:19
on the seek studio flex system
31:20
and the pull up detection in GMapper IDX software.
31:24
The seek studio flex uses the same dye information
31:27
to continuously update the spectral information
31:30
and reduce the number of pull up peaks that could be called.
31:34
GMapper IDX software then can either label or delete
31:38
those pull up peaks, reducing the number of calls
31:41
even further.
31:42
Analysts can then spend less time on artifacts
31:45
and more time on potential contributor peaks
31:47
in complex mixtures.
31:49
(upbeat music)