AI/ML Fundamentals for Product Managers
Summary
TLDR本次演讲由Darris Kumari主讲,讨论了产品经理学习人工智能和机器学习的重要性,以及如何在现有产品和组织中实现AI。演讲涵盖了AI的基本能力、如何作为产品经理理解AI的内部工作原理,并通过实例讲解了监督学习、无监督学习和神经网络在AI领域的作用。强调了AI在提高用户体验、产品差异化和推动收入增长方面的关键作用,同时探讨了实施AI技术的战略方法和潜在风险。
Takeaways
- 🚀 产品管理者应学习AI和ML,以保持竞争力并最大化客户和用户体验。
- 💡 创新的七个关键力量包括流程需求、行业变化、市场变化、新知识、潜在的不一致性和用户期望。
- 🌟 AI和ML可以提供额外的价值,帮助组织获得溢价收入和产品差异化。
- 🛠️ 实施AI技术时,组织可以选择自建或与云供应商合作。
- 🔄 对于AI和ML的采纳,组织结构上可能需要一个集中的平台团队来支持其他产品团队。
- 📈 产品管理者应从用户痛点和旅程出发,思考AI和ML在产品中的应用。
- 🥄 开始AI和ML项目时,建议采用分步方法,包括创意、概念验证和产品化。
- 🔒 在实施AI和ML时,需要考虑法律、数据所有权和用户数据安全等问题。
- 📊 AI和ML的成功需要业务的大力支持,包括领导层的支持和资源分配。
- 🎯 产品管理者应关注AI和ML技术的最新发展,以便在产品中实施创新用例。
Q & A
为什么产品经理需要学习人工智能和机器学习?
-产品经理学习人工智能和机器学习可以帮助他们掌握新技术,最大化客户和用户体验,提供差异化和有吸引力的AI驱动功能,从而使组织保持竞争力,避免被更敏捷的初创公司颠覆,并推动最大的收入回流到组织。
如何将AI集成到现有产品和组织中?
-可以通过与云服务提供商合作,利用他们的API和AI能力来增强自己的产品功能。对于大型组织,可以利用已有的AI技术平台和产品团队。对于早期AI应用,建议从云服务提供商开始,而不是从头构建模型。
人工智能和机器学习的关键能力包括哪些?
-关键能力包括回归、分类、神经网络、强化学习、自然语言处理和计算机视觉等。这些能力可以用于预测分析、用户旅程中的决策支持、文本和图像生成、语音识别和内容提取等。
如何开始AI和ML的创新之旅?
-作为产品经理,可以从识别用户痛点和旅程中的AI应用开始,然后通过概念验证和用户验证来测试和验证想法。最后,将这些功能产品化并推向市场。
在实施AI技术时,产品经理可能会遇到哪些法律风险?
-可能的法律风险包括数据所有权问题、客户数据训练的合法性、第三方云AI服务提供商的数据安全和隐私问题,以及确保遵守相关的数据保护法规。
如何确保AI模型的训练数据质量?
-需要对训练数据进行适当的预处理,去除个人身份信息(PII),确保数据安全,避免使用有偏见或非法获取的数据集,并确保法律要求得到满足。
产品经理如何识别AI在其产品中的应用机会?
-产品经理应该了解自己的产品领域、用户以及他们试图解决的用例。然后,可以根据这些信息思考AI和ML在现有产品和业务流程中可能发挥作用的地方。
什么是神经网络,它如何工作?
-神经网络是一种模仿人类大脑神经元功能的机器学习模型,由多层节点(神经元)组成,通过权重和激活函数处理输入数据,进行决策和分类。
生成性AI技术有哪些应用?
-生成性AI技术可以用于文本生成、图像生成、音乐创作等领域。它们通过预测下一个最高概率的单词或图像来生成新的内容。
产品经理如何使用AI技术提高自己的工作效率?
-产品经理可以使用AI技术来生成产品文档的大纲、测试用例列表、研究问题、路线图项目和功能等。此外,AI还可以用于生成演示数据和进行市场分析。
Outlines
🤖 人工智能与产品管理概述
本段落介绍了人工智能和机器学习对于产品经理的重要性,强调了学习AI的必要性,并概述了AI如何融入现有产品和组织中。提到了AI的能力以及可用的工具,并以简单的例子解释了AI的工作原理,特别是监督学习、非监督学习和神经网络。强调了产品经理需要了解新技术,以最大化客户和用户体验的价值,并帮助组织保持市场竞争力。
🚀 创新与产品管理中的AI应用
这一部分讨论了为什么产品经理应该学习AI和ML,并探讨了AI在产品管理中的应用。强调了创新的重要性,提到了Peter Drucker的七大创新力量,并讨论了AI如何提供额外的价值和差异化。同时,也提到了组织如何通过云供应商来实施和采用AI技术,并建议产品经理从了解组织内部已有的AI技术开始。
🛠️ 组织结构与AI实施
本段内容讨论了组织结构在设计AI和ML能力时的重要性。对于小型组织和大型组织的不同情况,提出了集中化平台的概念,并讨论了产品团队如何与中央团队合作。还提到了产品管理过程中的创新旅程,包括创意、概念验证和产品化。
💡 AI技术的实际应用
这一部分深入探讨了AI技术的实际应用,包括如何开始创新旅程、组织结构的设计、AI实施的技术方法和组织结构。讨论了如何通过云供应商来增强产品能力,以及如何利用现有的AI技术进行投资。同时,也提到了在产品管理过程中,如何通过理解用户需求和用例来寻找AI的应用机会。
🧠 探索AI的工作原理
本段内容深入探讨了AI的工作原理,包括监督学习、非监督学习和强化学习等不同类型的机器学习。通过具体的例子,如篮球投篮预测和神经网络分类,解释了这些算法是如何工作的。强调了产品经理需要理解AI的基本概念,以便更好地应用到产品中。
🌐 利用AI进行文本和图像生成
这一部分讨论了AI在文本和图像生成方面的最新进展,如GPT和Dali等生成模型。解释了AI如何通过预测下一个最可能的单词或生成图像来工作。同时,也提到了产品经理如何利用这些技术来提高工作效率,例如通过AI生成产品文档、测试用例和研究问题等。
📚 学习资源与未来展望
最后一部分提供了一系列的学习资源,包括在线课程、新闻通讯、YouTube频道和报告等,以帮助产品经理深入了解AI和ML。同时,也强调了产品经理在考虑将AI集成到产品中时需要考虑的因素,如数据收集、法律合规性和用户许可等。最后,鼓励产品经理学习AI和ML,以应对市场的挑战和机遇。
Mindmap
Keywords
💡人工智能
💡机器学习
💡产品管理
💡创新
💡监督学习
💡非监督学习
💡强化学习
💡神经网络
💡自然语言处理
💡计算机视觉
💡生成式AI
Highlights
为什么产品经理应该学习人工智能和机器学习
如何在现有产品和组织中实现AI
人工智能的组成能力及可用工具
理解AI黑箱中的监督学习、无监督学习和神经网络是如何工作的
作为产品经理,如何通过AI和ML提高用户体验和产品价值
创新团队和新产品零到一市场推广的经验分享
如何识别并利用AI和ML在产品和行业中的用例
AI成熟度初期,推荐使用云供应商AI而非从头开始构建模型
在大型组织中,AI产品领导如何与中央团队合作
AI技术策略、组织结构和执行过程的考虑因素
AI和ML在产品管理过程中的风险和挑战
AI成功需要业务的大力支持和执行层面的支持
AI和ML技术的最新进展,如计算性能、存储能力和成本效益
如何使用AI进行文本生成和图像识别
AI在产品管理中的实用工具和资源推荐
Transcripts
hello and welcome to today's
presentation on artificial intelligence
and machine learning fundamentals for
product managers my name is Darris
Kumari and I'm a director of product
management at a Fortune 500 Enterprise
SAS tech company and I've had the
experience of leading both Innovation
teams as well as net new products zero
to one in the market that take advantage
of AIML capabilities
in today's presentation I will cover
simple topics such as why you should
learn AI as a product manager how you
might operationalize AI within your
existing products and organization the
capabilities that make up artificial
intelligence or the tools that are going
to be at your disposal and then finally
simple examples and to actually
understanding what goes inside the black
box of AI how does supervised
unsupervised and neural Nets work in the
AI landscape
as a product manager it is key to enable
yourself on new technologies like Ai and
ml in order to maximize your own
customers and your own users experience
and the value they get out of your
products to help your organization stay
competitive in the market and not face
disruption by smaller startups with more
agility and importantly help Drive
maximum Revenue back into your
organization by offering differentiated
and compelling AI driven capabilities
beginning with the topic of why to learn
AIML as a product manager it's important
to highlight the benefits of innovation
now you may be familiar with Peter
drucker's seven key forces of innovation
that can drive opportunities for
Innovation but specifically to AIML
you're going to find improvements in
process needs industry and Market
changes new knowledge that has
formulated into these algorithms and
potential incongruities with user
expectations
to simplify the explanation AIML will
allow you to unlock a different
capability or experience that'll provide
additional value to your customers
employees and other users to allow your
organization to get premium Revenue
capture and product differentiation this
is extremely critical to the longevity
not only of your product but to your
larger organization to continue to
innovate onto its core competencies
now if we overlay this conversation of
innervation on top of a popular book and
concept of the innovator's Dilemma it
goes to tell the following story which
is that in many cases your existing
organization the existing product you
have may not have AIML capabilities and
that's because you've rightfully been
focusing on your customers demands and
the requirements of the market where
they wear today and you've gained a lot
of market share as a result what ends up
happening though is that the pace of
Academia the pace of what's potentially
possible through technological growth
and in this case what Ai and ml can do
with the latest trends in gen Ai and
other such Technologies is you see the
potential exceeding or performance
oversupply the demand of the market and
what ends up happening is smaller
startups that are a bit more agile that
are focusing on more of this bleeding
edge technology coming out of Academia
and doing research they're able to take
this new technology to Market productize
it and eventually match the demand of
the market because the expectation is
we're seeing today of AIML is no longer
something that's future facing it's
something that's becoming table Stakes
for every product manager to add into
their products to stay competitive and
to stay relevant and so the challenge to
you is how are you able to add your own
knowledge of AIML and the own use cases
to your domain to your industry to
identify opportunities to shift your own
organization right on this Innovation
scale
and so now that we have a good
understanding of the importance of AIML
as it goes to your own organization's
Innovation strategy let's talk about how
your organization can actually Implement
and adopt AIML technology starting with
as this is a technology the landscape of
where to invest as with any technology
or product problem and solution you'll
have a build by or partner opportunity
available to you and what we've seen out
there in the market is it's very hard to
build an entire AIML competency in-house
it's expensive to not only hire the
talent required for data science
resources but also there's a significant
infrastructure cost from a capex
perspective to get the necessary servers
for the training and for the investment
and as a result what we're recommending
what we've seen in the market is what
we're calling the cloud vendor AI so
these are your big players like
Microsoft with what they're doing with
Azure with Google with Amazon on with
IBM they're offering these API inference
calls as essentially
transactions and so your organization
simply needs to identify where in the
workflow you need what AI capability
identify the vendor that has that
capability send the API call in and take
the result and so you're essentially
augmenting your own product capabilities
with these third-party Cloud vendor AIS
to add that AI capability now if you're
in a larger organization or maybe your
organization already has proprietary
in-house AI technology built on its own
platform and made available to its own
product teams that's fantastic that's
the first place to start as a product
manager is learning about what is
available already in my organization and
what technologies can I re-leverage that
maybe other teams have already
implemented
but that's really the the playing space
where I'd recommend making AI technology
Investments it's a bit less desirable to
try to build your own models from
complete scratch using something like
open source libraries it's more
recommended even to reuse something like
from a hugging face something that's
been published that's open source but
what I'm hiding is building something
from scratch isn't the best opportunity
especially if you're early on in your
AIML maturity and so to recap you can
begin your Innovation Journey as a
product manager down AIML capabilities
not requiring necessarily any AI
technology in-house but beginning at
looking at vendors Cloud vendors for
opportunities to partner
now
that we've touched briefly on your
technology approach it's important also
to look at Classic organizational
structures when designing how to
actually scale out and building your
organization for AIML capabilities now
if you're a smaller organization maybe
you're a single product manager and just
one product team with a few products
under you you may not have this problem
of scale and in which case you're going
to be working directly as a single scrum
team or a set of scrum teams on your
product
looking towards Innovation and we'll
touch on that in a moment but generally
what we see in large organizations is
there's going to be a centralized
platform body that is providing the AI
essentially as a service to other
product teams within the organizations
you could call this The Hub and then out
in your different business units or the
different domains of your product
applications the different lines of
business you may have ai product leads
that are interfacing with that Central
team with either requirements but more
importantly just with a knowledge
expertise from a product owner
perspective on the domain as well as the
AI capabilities and in application
development perspective from
architecture and implementation and
there you'll be able to drive the key
requirements roadmap and vision that
ultimately get executed on by each
individual scrum team that might be
supporting your product again it could
be one team it could be multiple teams
but take uh take a moment to pause the
video if you'd like to ingest what's on
the slide if organizational structure is
something of Interest otherwise we're
going to dive forward and actually talk
to the actual journey of thinking about
as a product manager where do I even
begin now that I know maybe I'll use a
third-party technology maybe there's
something in-house for AI so what's next
for my product
and so for that I definitely recommend a
split approach in terms of ideation
proof of Concepts and actual
productization to get a AIML capability
out to Market
and what I mean by that is as with uh
really any robust product management
process you should always start with the
pain points you're looking to solve
start with the user Journal you have in
mind and ask yourself where in my
product where in my user Journey today
is there any form of Intelligence coming
into play it could be a systems already
getting a decision for your users it
could be a users in the loop calculating
something or identifying something or
making a decision or even following a
specific business process as part of
that product or maybe consuming
something in it as a user and more of a
b2c concept the idea is AI is going to
be able to improve and augment those
existing processes and capabilities to
proactively serve your users insights
serve your users informations make
decisions take actions on their behalf
but it all begins with understanding
your domain your user and the use cases
you're looking to solve for and we'll
get into what AI capabilities you could
think about in a bit later in today's
presentation
but fundamentally as I highlighted it
begins with you as the product manager
thinking about where in the process
today AIML can fit in to essentially
improve and augment an existing
decision or intelligence uh step within
a process product and capability
and that's where you'll be able to take
that requirement in more of a proof of
concept hackathon informal fashion go
work with your scrum team to actually
balance out a little lightweight proof
of concept of
can this idea that the product manager
brought to us on implementing AI can it
even work am I getting the results I'm
expecting is the user experience
something that's explainable and it
makes sense and so before you even think
about productizing something actually
bringing it to the market to your users
is tremendously valuable to test it in a
proof of concept and even if you have
some customers available to either in a
design partner manner or just some close
customer contacts you've worked with as
a product manager it's a great
opportunity to take that proof of
concept and to do additional validation
that this isn't just using AI to use AI
which is something I'm going to call out
AI is just another tool in your
capabilities not every problem not every
product will need AI necessarily now ai
could become part of other points in the
user experience interfacing with your
product maybe to purchase it maybe on
the support they receive but keep in
mind your product itself you as a
product manager will be best equipped to
identify and understand is there
opportunities for AI in my product
and so later in today's presentation
hopefully we'll share some ideas that
will help you as a product manager
Identify some thoughts and capabilities
add to your toolbox on what AIML can do
and so once you've done the ideation
your team has done the POC you've
validated uh technically and from a
business value perspective that it is
improving the user value then you can
actually launch and productize and as
you go to productizing an AI capability
I will say there's a bit of red tape you
might run into as a product manager
there is the legal aspect who owns the
data can you train on your customers
data where's the data coming from even
if you're using a third-party Cloud AI
are you allowed to send your customers
data to that third-party for processing
so you might have some contractual
updates and of course on an I.T side if
you're doing it in-house you may need to
make sure that your it infrastructure
your data center is equipped for the
right server types to do the expensive
training processes not as much of a
concern if you're using Cloud AI
so by taking into account your AI
technology strategy the organizational
structure and your execution process you
can begin envisioning your strategic
approach to AI which is traditionally
split into three phases first you could
think of it in from a systems
architecture and engineering approach in
terms of how are you going to get your
data what are the algorithms you're
going to use where is the human machine
teaming going to come into play is your
compute in place for storage and what's
the governance like for robust and
responsible Ai and from there you can
Define the Strategic principles that
will educate that around the vision for
your product as a product manager
incorporating Ai and what that ideal
state looks like the governance culture
infrastructure that needs to be in place
to accomplish that Vision as a product
manager and the AI Talent you may need
on your team to help execute against
this again if you use cloud vendor AI
not so important to have the talent the
key thing is you as a product manager
being aware of the technology in use
cases you can implement which brings us
to part three so you've thought about
your persona you've thought about the
key pain points in use cases and where
to apply AI what use cases do you Target
first and for there there's a very
simple diagram that can be used for
human machine augmentation which is
taking your highest confidence decision
with the lowest risk and Outsourcing
those to the machines first and that's
where we can talk about the reimagining
of work and the age of AI reimagining
how your product works and the processes
your products to support where they'll
still be human only activities around
leading empathizing creating and judging
but you'll start seeing these human
machine teams and compliments around
training explaining sustaining
amplifying interacting and embodying and
machine only interactions where
Automation and AI will actually take
over a lot of that human workload around
simple transactions iterations
predictions and evolving and adapting
and actually learning based on an
environment and so AI in the short term
and for simple use cases very very well
suited for delivering recommendations
and insights in your products and humans
May remain in the loop to make judgment
and decision based on that information
now with any new technology and new
technical process you will run into
risks and AI of course comes with its
own set of risks so as you go about
ideating and potentially launching proof
of concepts of AI capabilities there's a
couple key areas to look for for risk
and to mitigate risk one is of course
you need data to train your AI models or
data to even send in to get a result
from a response from an AI endpoint and
so you need to make sure that you
properly pre-process your data strip out
pii information if you're housing that
data keeping your customer data safe and
making sure there isn't any biased data
use in your data sets illegal data
acquisition make sure the legal
requirements are in place and make sure
the infrastructure is both secure and
scalable enough to house the training
data and to train on the data as
necessary
as you go into building testing and
deploying make sure that there is Gates
and checks and you have an ability
organizationally to disable capability
if required because there may be for
self-improving algorithms uh there may
be a risk of ml poisoning there may be
vendor risk where if you're deploying uh
depending on a third-party Cloud vendor
that vendor goes down for a period of
time what's the impact on the user
experience is there a fallback for a
non-ai capability making sure your
product can handle that and then finally
having a feedback loop in and this could
just be on within the product collecting
some csat data from more of a plg
perspective but also collecting feedback
on the performance from a Telemetry
perspective of your models how many
times are users taking recommendations
how many times are users changing data
after AI changed it so you know it's
been a uh inaccurate prediction or value
and so having your own Telemetry as well
on the performance of your product will
be key as you go into implementation and
so up to date we've discussed kind of
the benefits of innovating with AI we've
discussed some organizational structure
we've discussed a little bit on how to
think about structuring the AI process
beginning as a product manager thinking
about use cases pain points going into
pocs validating customers and then
finally We Touch briefly on these
technology risks and the last thing I'm
going to highlight is AI to be
successful
requires significant support from the
business as well I'm sure all of you as
product managers AI in your backlog
might be something that historically had
been more future facing I mean today
with the hype around chat gbt and gen AI
has discussed and maybe something your
leadership is asking you for today and
so the biggest problem in the past was
lack of executive support we're seeing a
lot of that happening now in
organizations they're getting the
executive support they need from the CPO
from their own customer demands giving
them the funding and resourcing they
need to Pilot and test those pocs around
AIML capability
and so I'm very excited to Now cover
what are the capabilities to think about
as a product manager the new tools if
you will in your toolkit to think about
actually implementing in your product in
those pocs whether it's in-house or
whether it's using a cloud vendor Ai and
first of all before we do so it's
important to highlight that the reason
we have so much AIML power available to
us today is significant strides on both
compute in terms of the performance to
compute the amount of storage we're able
to have the models we can train and then
finally the cost has come down to a
point where it's very very uh efficient
and available to organizations to uptake
from a vendor perspective and so
artificial intelligence as we dive into
the capabilities here is it's a very
large State they call it suitcase term
you know a broad field of study a lot of
subsections really working to make the
the systems and machines perform human
tasks and so in this diagram here I've
listed out some of the key types of AI
that you're going to run into and you're
going to hear in the market and as a
product manager and it's important to
highlight that there's a lot of
different capabilities related to what
humans can do today with their
intelligence and their reasoning that we
can now have these systems in many cases
outperform Us in in conducting and so
the basic machine learning could be
something like regression which is
calculating given an input what a given
kind of output is going to be a more of
a data set you know a continuous data
set you have classification which is
more discrete you're trying to identify
something in fix buckets it could be
identifying a color identifying a shape
identifying a category so fixed classes
and then you have the neural next and
we're going to get into neural networks
and how they work at the end of today's
presentation but that's the capability
that's given us powerful text and image
generation such as the chat gbt you know
from open AI or Dali and that's type of
models and you also get things like
reinforcement learning you know when
you're seeing the robots that are able
to walk themselves like happening at
Boston Dynamics it's a great example of
reinforcement learning where the systems
are given rewards you know they have to
achieve a certain outcome and they're
given parameters that they can
manipulate to try to get that outcome
and then you have of course a within
those flavors of supervise just meaning
you have structured data structured
inputs and outputs and unsupervised
learning where it's a bit more
unstructured there's a lot of data and
the AI is essentially telling you the
pattern telling you the Insight in terms
of that data very popular for things
like clustering or finding groups
now there's also simpler things like
simple decision trees you know I think
your call center it's not high-tech
technology but it's technically an
artificial intelligence making decisions
and then you have as an extension of
that call center technology and
basically just speech and communication
the whole speech and text branches of
natural language processing natural
language generation natural language
understanding intent entity extraction
and in all the capabilities that are
famous with things like Siri and Alexa
around speech to text text to speech
translation content extraction and even
summarization which are some of those
use cases that you're going to get there
and then Vision you know it's
self-driving cars very popular
pioneering work done by the team over at
Tesla but a lot of organizations now in
the space but there's the need for
computer vision and all the other
applications on what you can do with uh
identifying footage whether it could be
geospace information whether it's camera
information from a security perspective
and image recognition and so maybe
you're getting a stop sign getting the
speed limit or so forth and so more on
the OCR side
but fundamentally these are a number
this is just a small list by by all
means there's a lot of additional
contacts especially as you get into
things like Robotics and other domains
there's a lot of extensions deep
learning and so forth Beyond these
simple Concepts but for most product
managers and especially for those that
are just beginning their Journey on AIML
Concepts this is a relatively good high
level view that I found to introduce
some of the capabilities to think about
may be part of your user Journey from
activities or inputs outputs they're
expecting in their user experience
now that we have a shared sense of some
of those basic capabilities on AI and ml
it's time to look behind the covers and
see what's actually going on in the
black box so you have a bit more of an
appreciation for how these algorithms
work at the end of the day it's math not
Magic
so let's take a look at some of these
simple Concepts and to remind you
machine learning can be defined as a
study of computer algorithms that allows
computer programs to automatically
improve through experience so there's a
concept of improvement there's a concept
of it figuring out a solution on its own
without having every case every
permutation coded ahead of time and we
talked about a few of the use cases in
the capabilities slide but other
examples include identifying people and
self-driving cars identifying emails if
they're spam or not predicting home
prices and temperatures teaching robots
to move and sort uh packages as they do
at Amazon and play games and understand
and come up with text images and sounds
a lot of the deep fakes you see or the
image and text generation you see today
now let's begin though getting into a
little bit more of the details around
some of these common types of AI and ml
beginning with what we called previously
supervised machine learning now when you
hear supervised machine learning it's
important to highlight that we're really
just High uh explaining that there's
labeled data involved and you're looking
to predict or map a given set of inputs
or features as they're called to a
corresponding set of outputs and
something important to highlight here is
there's generally a correct output or a
correct answer for a given input and so
given the number of Dimensions you have
you're going to have a line essentially
driven to make decisions but in a simple
two-dimensional example here we could
look at classification it's drawing a
decision boundary and you could see two
classes then being allocated to these
points on the graph depending on where
they lie in the boundary or a regression
based model which is more for that
continuous data set so predicting your
home prices versus predicting your
temperature there you may have an input
of how big is it you know what's the
square footage and then your output of
the price and that's again
plotted on that line as the prediction
and so supervised learning is all about
having that labeled data and labeled
data set whether it's hey all these
pictures or cats all these are dogs or
the attributes hey all these
temperatures correspond to the following
cities on the following days and so
forth regression but label data is the
key piece here for supervised machine
learning
now the next type of classic machine
learning is what you'll call
unsupervised machine learning so whereas
with the supervising machine learning
you kind of had labeled data set you
knew what the input was and what the
output was on those examples with
unsupervised machine learning you're
just feeding a lot of data into this
system here and then it's coming back
with what it detects as generally
clusters or patterns commonalities and
so maybe it's doing it based on the
color and it's giving you a grouping on
color maybe it's doing it based on the
height maybe it's doing it based on the
number of edges or the shapes and so
here there isn't one right answer right
there's in a fixed output to get based
on the inputs they're looking for it is
more of an observation that you're
predicting on and so clustering is a
very very popular case here where you
take in a lot of raw data and maybe you
want to assign some kind of a label or
attribute based on a cluster
and so that's uh the second piece of our
machine learning puzzle we had the
supervise labeled done supervised
unlabeled
now the next common type of machine
learning is something very cool it
referred to as reinforcement learning
now the unique thing about reinforcement
learning is that what you're defining
ahead of time is really just the
objective of the AI what they call an
objective function which is the goal and
then the AI is learning through multiple
simulations and this doesn't have to be
in the real world generally you could
use software to do the simulations and
it could do simulations in parallel to
learn from it but it does these
simulations and sees given this fixed
set of inputs and the outputs that
happen as a result of those inputs is it
getting closer to this objective
function scoring higher so to speak and
it's using that to guide its decision
making and so in this video example you
could see a robot is trying to get that
green circle into the red circle and the
graph you see in the lower right is the
result of the objective function the
closer it gets into that Circle the
higher the score of the objectives uh
function and so
the robot is essentially learning how to
move the green cylinder into the red
circle because it's maximizing the
reward it gets based on how it moves its
arm and how it moves this green circle
in space relative to that red circle and
so there's a lot of cool examples about
this and papers published on this online
around AI agents playing hide and seek
learning to walk learning to navigate
obstacles and it's a very cool example
of machine learning
solidify some of these Concepts what
we're going to do now is walk through
two
size machine learning examples in
Greater detail so you can understand how
the algorithms work and generally what's
going on in those black boxes of all
those AI capabilities we listed earlier
we're going to look at two examples one
prediction and one classification for
the prediction example we're going to be
given a certain set of data and we're
going to understand a measure of
correctness which is going to be making
a basketball in the hoop and we're going
to be able to train a prediction
function to say given wherever our
individual or in this case this robot is
in space we're going to be able to
predict the power and angle which it
needs to throw the ball in order to make
the hoop and so that's going to just use
a simple linear regression on the other
hand we're going to get a bit more
complex by looking at how we can do
classification however this time using
the concept of a neural network and
neural Nets or deep learning Nets as
well as multiple layers that's something
we're going to spend a lot more time on
towards the end of today's session
because it's an extremely important and
Powerful concept that powers a lot of
modern AI
so let's walk through this first
prediction example of using linear
regression to identify how to throw a
basketball given any position on the
court
and so given this example we essentially
have one simple input and one output the
input is the distance away from the
basketball hoop
and the output that we're after is how
much force they should use to throw it
now we're going to assume that their
angle is going to be fixed for this just
to normalize uh on a single Dimension
here what we're predicting and what our
input variable is but keep in mind in
practice you could have multiple
variables all happening at once
now the goal is to predict of course
what force we need to use to actually
get that ball into the hoop and so when
the algorithm is just starting out
without any data it has no idea how hard
it needs to throw the ball to actually
achieve the goal of making it in the
hoop
and so what you need to do is collect a
ton of data points to understand what
that success looks like and if you have
the data already it's a matter of just
labeling and collecting the ones that
were successful so you can see here this
is actually the balls turn green anytime
they go in the hoop indicating a
successful data point but otherwise
you're going to need to go out and
actually collect this data so you have
those sample examples of success and
linear regression as an algorithm is a
simple y equals MX plus b line and so
it's great for 2D representations of a
single input a single output but as I
highlighted you could get into
quadratics and multiple variables with
actual applications of this similar
technology
in our example though we're looking to
collect the training data of our shots
that went in and the algorithm is going
to then be trained to learn off of what
that success looks like
so what we'll be able to see here is
that we can plot out all the rows of
these data points where you have a given
X distance and a given y uh force that
you need to use based on that distance
and a simpler linear regression just
calculates the optimal y equals MX plus
b line minimizing the squared error
which is essentially the distance any
point is away from the line so how far
off was your prediction and so it
calculates that line that minimizes the
distance any point is away from the line
on average and so we reduce that error
rate in most machine learning algorithms
have a concept of an error rate that
you're looking to minimize to get an
optimal line and while in this case
we're very uh sure that it's going to be
a comprehensive data set of different
points and the Court's not going to
change the hoop height is not going to
change it's important to be mindful of
this concept of overfitting so in the
real world not all environments are the
same and so this model trained for this
hoop at this height we could have a
model that's over fit for just this
scenario and doesn't take into account
say the player height or other Hoops or
the type of ball you're using and so on
and so forth and so overfitting just
means your line has been optimized for
your training data and the data you have
but not necessarily the real world
but in this example we're quite
confident and happy with the result of a
line that maps to the input force based
on the distance we have to get us a
successful hoop in the basket
now to get a bit more technical on this
whole concept of mapping a a set of data
against the minimal error rate it's
important to throw out a word out there
called gradient descent this is a very
very popular term and it essentially is
just a graph that highlights your error
rate at different values in your
algorithm and in the input so in this y
equals MX plus b we're testing different
M's we're testing different B's and
we're trying to figure out at what point
of each do I have a relative minimum or
the bottom of that gradient descent
array curve and so we can see here as
the line continues to fit closer and
closer to the points the error rate
continues to decrease closer and closer
to ideally zero so the lower the error
the better and that's what helps get us
our optimal y equals MX plus b line or
our optimal prediction based on any X or
position we are on the court what that y
or the forces we need to use to make
that basket
and just like that we have just seen how
we've created together an algorithm that
is successfully making any basket and
every basket no matter where it is out
there on the court and we did that
through a collecting the data be
labeling what success looks like and
then see training a model that reduces
our error rate and as we highlighted
this was a very very simple the simplest
example using just a two-dimensional
space y equals MX plus b in the real
world you have different inputs which I
apply as different dimensions and
different variables and so instead of
just a simple linear function you get
into quadratic or logistic or other
function types based on the environment
and the data
now where something like linear
regression really pulls a lot of its
source from statistics the next example
we had of classification utilizing a
neural network is taking more from
biology and this is extremely
fascinating if you haven't learned about
this a neural network is a way to teach
the computer how to process the data
mimicking the way the human's brains
neurons function in our own biological
processing it's really a type of machine
learning that sometimes also they call
it deep learning because it's multiple
layers of neurons that fire in a in a
sequence to interconnected nodes that
they call neurons with different
activator functions and we'll get into
all this in detail in the upcoming
slides but the takeaway is it's actually
built to replicate in a way the human
mind and human intelligence using these
artificial neurons
so let's take a look at how neural
networks work for classification
so let's take a deep dive into one of
these orange neurons in our neural net
what is a neuron really when it comes to
our mathematical algorithmic
representation that machine learning
uses really at the end of the day keep
in mind you have a series of inputs
coming into this decision maker of a
neuron which then spits out its outputs
to additional neurons or potentially to
your final output layer and so what the
neuron is doing is generally a it's
accepting inputs from other neurons or
from your raw input and it could be
applying weight to those different
inputs and that helps it understand
maybe importance of an input in its own
calculation
from there it generally has its own
summation function to take into account
all those weights with potentially a
bias function that offsets the summation
based on some type of an attribute and
from there we go into a very very
important step which is called an
activation function and the activation
function is useful because it helps
normalize
standardized and essentially keeps in
check the desired dimensions and traits
of the inputs to the next output or
prediction and so if you ever hear the
terms thrown around like sigmoid or relu
or softmax these are just different
activation functions that take all that
input data from your other neurons and
it makes a calculation based on it to
keep the output uniform or within some
type of a boundary
and so let's go into an example of the
classification let's say the difference
between a balloon and a tree how can a
model tell well you might use different
dimensions of decision making
one dimension let's say we just use
color okay well what if our balloon and
our tree are both the same color now we
need to add another dimension roundness
great now I have a green balloon and a
green tree but what if my tree was round
and so now we have to add another
dimension into our decision making which
it could be something like shininess and
so how these models work and in many
cases actually explainability is a
challenge in that you don't really know
what the model is looking at in terms of
these different dimensions but it's
fundamentally coming up with its own
decision making criteria in those
different layers in the neural net to
distinguish something like
is it a balloon or is it a tree and so
with enough Dimensions enough layers you
can accurately distinguish one object
from another and each Dimension is
essentially being checked the activation
function helping classify the object
based on those inputs it gets for each
of these different dimensions you know
different colors would have different
input values that then help us get to a
predicted output of is it a balloon or
is it a tree
now let's look at another classification
problem with a neural net and keep in
mind that a lot of the different
criteria that you use for decision
making would most likely be inputs that
are initially sent in to this neural
network model so say we wanted to
identify something as a dog or a cat
based on an image that came in first of
all computers can actually see anything
there's no images so there's a slight
asterisk here when I say we're looking
at an image of a dog really that image
would probably be turned into a matrix
of pixels and maybe the pixels are
assigned RGB values or other values
around their intensity or significance
and so what's happening in this case
with a neural net unlike our regression
example where you altered your M and
your B in that y equals MX plus b
formula with the neural net each layer
is playing around with the weights from
the previous layers it's playing around
with what it wants to assign
significance to and value to
until it gets to the next neuron in the
next layer and so
something to keep in mind as well is the
only thing you really have access to and
you see out of the black box is going to
be that output so while all those
neurons and and the weights and the
biases are being adjusted really all
you're seeing at the end of the day is
that final prediction or calculation
that output is it a dog or is it a cat
maybe in this case it's a probability
score of is it a dog uh or is it a cat
or something along those line that just
a class of dog or cat
and so extending on this cat and dog
image recognition problem and taking
into the context that your input layer
is feeding in the pixels of the image
you need to keep in mind that some of
the power of the neural net is the fact
that not all neurons will quote unquote
activate or send an additional input
into future neurons Downstream and the
reason that happens is through that
activation function we were talking
about earlier so you might have a set of
neurons higher up on a layer that feed
in given their weights into that
summation function but then that
individual neurons activation function
may decide if it's that one that zero or
what that value is that it's going to
continue to pass Downstream to further
layers and further neurons and so that's
how a lot of the decision making
actually gets made which is certain
signals so maybe it uses something like
the size of the animal or the teeth or
the shape to make a decision it'll start
embedding those as neurons either do or
do not fire to help it make a decision
down the stream
and so let's bring these classification
with neural net examples home with one
more flavor and explanation and so in
summary keep in mind that your inputs
into these neural net models they of
course have to be some type of
mathematical representation because the
computer itself doesn't quite understand
images but it can understand the data
and so an input image say of a cat or a
dog would most likely be of a certain
pixel's height and pixels width and what
generally happens is you translate that
into a matrix or a vector and the
different colors might have different
values to them and that's really the
basics also and basis of computer vision
is taking these different pictures and
translating them into these numerical
Matrix representations
and so when we talk about classifying
and deciding if something is a cat or a
dog we really talk about beginning with
that numerical representation of that
image and then feeding it into our given
uh neural net here and keep in mind the
number of inputs you could have can be
extremely large you see some neural Nets
with Millions uh if not even billions
for the case of chat GPT input
parameters that it then fires off into
its its neurons to get to that final
output and so
as we highlighted the images are pixel
values we're using mathematics in terms
of calculating between each of these
neurons here the weights between the
lines between one neuron and another the
biases on each neuron and then the
activation function being used in these
neurons to basically say am I a one or a
zero right am I activating or am I not
activating and that ultimately is going
to get us to that output of a dog and
what generally ends up happening is all
those different features that are used
for decision making they'll happen one
layer at a time so maybe you identify
the edges of a dog and then you're
getting a bit more granular into what
the edges are combining as and specific
features like ears and nose and teeth
until getting to that final output and
so
it's really a combination of multiple
inputs and a combination of multiple
layers that are activating to get us to
that given output
and the same technology keep in mind can
be used for things like uh image and
text recognition or or OCR of is it a
one is it a two is it a seven and so as
you could see through this diagram of an
example as the different numbers get
interpreted and keep in mind they would
be translated into a matrix of their
values you have different neurons that
are lighting up here in green because
the activation functions criteria has
been met and the different weight and
biases are highlighting what those
numbers are keep in mind in this
specific text example as well there
would have been a training set of
thousand excuse me hundreds of thousands
in this case even of images maybe
thousands of numbers hand drawn with
actual manual labeling being done saying
hey this is a one this is a two and
that's what's being fed into train the
model
and so really really powerful stuff
mimicking the human brain in this case
using a neural net to help classify
something into a value we saw a dog tree
cat uh number number different use cases
here
now I would be very remiss not to touch
on the latest hype of things like Chachi
PT Dali or Texas generation so we'll do
that briefly before ending today's
fundamental overview session so how do
AI systems generate text we just saw
something like predicting your Force for
basketball we saw a classifying uh image
as a dog or cat how does this all
translate into things like generating
text well the great news is more or less
a lot of the fundamentals you just
learned still apply here when it comes
to text we treat each different kind of
piece of text as something that's called
a token so the word my is a token
favorite is a token color is a token is
a token red is a token and the training
data for something like a text
generation model could be as large as
every piece of text available online and
in public uh scientific Publications and
how the model is actually training is
it's predicting the next token so given
a token or context of tokens what is the
highest probability of the next token so
in this example if I had said my
favorite the neural net is saying the
highest probability coming next is color
then the highest probability coming
after that is is then the highest
probability coming out of that is red so
when you get into these text generation
algorithms what's actually going on is
they're predicting token by token the
highest probability word it thinks is
happening next and so
training data you know is taking these
tokens taking all these text examples
running that neural network and then
getting at that prediction for text and
keep in mind what we're seeing in
industry is things like aging GPT Auto
GPT Lang chain where users are using
these models uh like a GPT but they're
extending capabilities Beyond simply
returning your text to actually do
things and take actions in the world
through those examples provided and so
as a product manager this is a very very
relevant area of generative AI where you
think about any UI your users have or
any process where there's an input in
this output or there's a recommendation
or help you'd like to provide them you
have this advantage of taking this
text-based processing and then how you
want to interpret that text you don't
have to leave it as text on the UI or
screen you can take a text result and
display it as a set of check marks
display it as a Playbook book display it
as a summary on the UI and so forth and
so a lot of powerful use cases coming
out of generative Ai and being able to
have the system come out with a text
result given a text input
and image generation is actually even
more interesting than what's going on
with those tokens and the trained
examples of text generation and
predicting the next token because what's
going on with this algorithm for image
generation it all comes back to of
course having trained data but what
they've actually done with image
Generations they start with a given
image say of this pyramid that's been
compressed and then they'll go ahead and
they'll continue to add noise to that
image and keep in mind an image at the
end of the day is of course just a
matrix that's represented as numbers but
they keep adding noise to that Matrix
until finally you have an extremely just
noisy image that to the human eye you
can actually really make it out or see
anything
how these generative image algorithms
work is they'll then given an input so
say you started with this say you know
two pyramids uh with a blue sky that was
your input and you provided that image
and so the system added noise what it
then does is it starts with an image of
complete noise and then given your input
it attempts to get back to that original
state by removing the noise or it's
called denoising and they use something
called unit for this and so maybe you
typed in give me a beach landscape and
so of course it'll have images it's
trained on of beaches of Landscapes that
it's added noise to and so that it
starts with complete noise and step by
step it iterates in this net neural net
and it attempts to remove the noise from
it to get back to that Source image and
in this case that's how it kind of
dreams up an image if you will because
it doesn't have that exact image ahead
of time it is is incrementally peeling
back the noise and coming up with what
it thinks the image should look like so
very very fascinating and it's ways you
can generate text generate image and of
course as we saw from the other AI
capabilities you can do things like
music you can do things like voice you
can do all kinds of prediction
algorithms whether it's text or images
or or categories or temperatures or
prices and so a lot of opportunities for
your product as a product manager to add
capabilities in now on the note of
generative Ai and Technologies like
Chachi PT I will say not only can you
apply these Technologies within your
products you as a product manager can
apply these Technologies to make
yourself more productive and so examples
that I've personally found uh use in as
a product manager includes writing
outlines or skeleton starting points for
your epics your product requirement docs
turning short descriptions of stories
into lists of test cases or splitting up
stories you give it into two different
stories if you have larger scope
defining the research questions you
should ask for your new product or your
new domain identifying roadmap items and
features and here's another great one
identifying use cases for new
technologies like AIML within your
product or industry if you don't know
where to start you've gained a lot of
information today but you don't know
where to start you can ask chat GPT hey
given the following industry uh
following product on my product manager
sure what type of capability should I
consider on my roadmap great starting
point here it's also useful for
generating realistic demo data for your
engineering teams your QE teams for
testing it can be used to identify jobs
to be done and personas it's been
tremendously powerful for us to come up
with research questions that you then
take into research exercises as
questions and then getting go to market
considerations a lot more prompts and
ideas that I'll put into the links or
that you can request access to this
PowerPoint uh just leave a comment
requesting access to this PowerPoint
it'll be including all the links here
that you can click today now other
interesting examples include as I
mentioned identifying use cases with AI
for your product or even explaining
Concepts about your products and there's
a lot of other Technologies I'll say
that are out there on the market today
across a number of different Tech Stacks
that get you optimization in your
workflow in your processes on how you
get work done using AI
now while today's overview was by all
means a very brief and a very quick look
at what it looks like to be an AI
product manager there's a number of very
very useful resources I'd like to leave
you all with and it covers the what the
why and the how of AI and ml for product
managers and really Business Leaders on
the side of the what I highly recommend
the Coursera AI for everyone course the
Deep learning.ai batch AI newsletters
the two minute papers YouTube channel
and Stanford's read the AI index report
which I recommend you read as action
items I recommend engaging in hackathons
with your team to incorporate cloud or
in-house AI capabilities to familiarize
yourself with internal process and
approved vendors if required in terms of
what it looks like to get an AIML
capability including legal including it
and any dependencies and of course
familiarize yourself with the internal
capabilities that might already be
in-house in your organization on your
platform in your company's toolkit that
have already been approved and used
Maybe by other product teams
once you've done that it's good to learn
the why of AI and ML and how it impacts
your business strategy MIT has a great
AI uh implication for business strategy
six-week virtual async course there's
also a AI strategy in roadmap systems
engineering approach to AI development
and deployment one week kind of live
full day course both of them relatively
cover the same content so I would say
pick one of those as your next step and
for those that are looking to get more
technical on the back end of how some of
these AI ml algorithms work I do
recommend the machine learning
specialization on Coursera co-produced
with Stanford it's a three-week online
virtual async it covers supervised
machine learning Advanced learning
algorithms and unsupervised learning
recommenders and reinforcement learning
there's also for the generative AI folk
all that hype there's a great Cloud
skills course that Google put out
available as well that one's free and
then you'll find two good books that
have been recommended in Industry
designing machine learning systems and
iterative process for production already
applications by chipwin and building
machine learning powered applications
going from idea to product by Emmanuel
Amazin and so a lot of additional
resources I'm leaving you with and again
if you're looking for the raw PowerPoint
do leave me a comment and I'll provide
you a link to a hosted uh document
and with that ladies and gentlemen it
takes us to to taste summary for our AI
product management fundamentals and I
want to leave you just with the
importance of learning as a product
manager Ai and ml capabilities to help
you come up with new Innovative use
cases protect from those startups that
are going to be attacking you with new
technology and to help provide greater
value and differentiation to your own
users to lead to Greater Revenue
recognition for your organization
some final thoughts as well include the
fact that not every problem in every
product truly requires AI it's it's
tremendously useful in a lot of
applications not all applications AI is
math at the end of the day it requires
data it's not magic and data is key to
effective Ai and so as you think about
long term adding in AI into your product
think first do you have the Telemetry do
you have ways you could collect data do
you have the legal in place to allow you
to use that data consume that data and
what's that strategy look like is it
going to be in-house AI or most likely
maybe a big cloud vendor Ai and you have
your customer permission to send their
data there and so a lot of opportunities
as product managers a very exciting time
and for especially even as users as
customers a lot of exciting time where
you're going to experience your own
products in the market and the demand of
the market evolved to incorporate a lot
more AI ml capabilities over the coming
years so with that I thank you all for
listening and I wish you the best of
luck in encompassing and learning more
about AIML into your own Industries and
products
تصفح المزيد من مقاطع الفيديو ذات الصلة
How to Break into AI Product Management without experience
The ONLY 4 Ways to Become an AI Product Manager with No Experience
AI Product Manager
Demis Hassabis - Scaling, Superhuman AIs, AlphaZero atop LLMs, Rogue Nations Threat
Leveraging AI in Supply Chain Management: From Data to Delivery
Ilya Sutskever: OpenAI Meta-Learning and Self-Play | MIT Artificial General Intelligence (AGI)
5.0 / 5 (0 votes)