Page 18 - CCFA Journal - Third Issue

加中金融
CCFF2020/加拿大中国金融论坛 2020
问答环节 Q&A
主持人：谢谢各位嘉宾的建议。现在进入问答环节，第一
个问题，金融机构广泛采用人工智能最大的障碍是什么？
韩玫博士，你可以先回答吗？
韩玫博士：在传统金融领域里，人们真正关心的是解释能
力，就像卓女士提到他们关心特征工程一样。对我们来说，
模型有时候就像一个黑匣子。我们正尝试将新的深度学习
与传统的特征工程相结合。我认为如何将传统的机器学习
方法与新的数据驱动方法结合起来会存在一定的困难。
主持人：Touyz 博士，您有什么补充的吗？
Touyz 博士：对我很同意韩玫博士的观点。我想补充的是
我们在很多会议上看到很多有趣的应用，尤其是一些深度
学习技术。真正的挑战是如何定义在哪里使用商业问题。
所以我认为参加这些会议然后分析比如金融领域的一个应
用是非常有意思的事情。我们确实在一些领域创造性的应
用了这些技术，但除了解释能力，主要障碍的其中之一是
怎样提升更先进的方法。
主持人：您认为任何一个组织结构都可能成为障碍吗？
Touyz 博士：我现在想的是在大多数金融机构里，很多高
官和领导已经意识到将数据融合到系统的重要性。所以第
一步是要能够阐明数值。当然，美国、欧洲和加拿大一些
机构已经开始建立研究实验部门，专门研究如何将这些想
法引入金融领域。从一个组织的角度来看，它最终会走上
正轨。现在在我看来，问题是如何在一个组织中以一个更
有条理的方式应用这些技术。
主持人：卓女士，您有什么想补充的吗？
卓女士：银行业是一个受高度管制的行业。举个使用人工
智能模型来批准信用卡申请的例子，有一家人申请信用卡，
妻子的信用卡批准的额度比较低，而丈夫的额度却比较高。
那我们如何在机器学习模型中检测到这种偏向呢？所以在
财务决策过程中，我们需要先人工解决这个问题再去使用
机器学习模型，这是个难点。
主持人：谢谢卓女士！下个问题是如何在疫情期间保护数
据隐私？我们先从韩玫博士开始吧。
韩玫博士：这个问题很好。在医疗行业，获取正确的医疗
数据是非常关键的，从一个研究实验部门来看，我们花了
很多精力来开发保护数据隐私的工具。即使是在使用谷歌
街景时，我们也做了很多工作去模糊车牌、人脸和确切的
街道地址。数据隐私尤其在金融领域很重要，我认为从研
究的角度来看我们正在开发工具、试图开发新的算法来确
保数据对推出的功能和保护客户隐私方面是及其重要的。
CCFA JOURNAL OF FINANCE
Page 18
加中金融
Moderator: Thank you for everyone for your input. We can
now enter the Q&A session. First question, what do you
think is the biggest roadblock for the wide adoption of AI
for the financial institutions? You first, Dr. Han.
Dr. Han: In the traditional financial field, people really care
about explanatory abilities. Just like what Ms. Zhuo mentioned,
they care about feature engineering. For us, models are
sometimes like a black box. We are trying to combine the new
deep learning trends with traditional feature engineering. I think
there are roadblocks in how to combine the traditional machine
learning approach with the new data-driven approach.
Moderator: Dr. Touyz, do you have anything to add on that?
Dr. Touyz: Yeah, I completely agree with Dr. Han. I think what
I would add to that are a lot of really interesting applications we
see in conferences, especially some deep learning technologies.
The challenge is really being able to define the business
question on where they can be used. So I think it's really
fascinating to attend some of these conferences and then be able
to parse out maybe an application within finance. There are
definitely areas that I think that we're applying these things
creatively but indeed, I'd say that that in addition to explanatory
ability, some of the central roadblocks is to elevate some of
these more sophisticated methods.
Moderator: Do you think any organization structure can be
a roadblock?
Dr. Touyz: So, I am just broadly thinking about most financial
institutions and what's really been exciting about the field is that
a lot of executives and senior leaders have realized that there's
an importance (to integrate data into their systems) and so the
first step in terms of being able to do that is to articulate the
value. Certainly U.S., European and a couple of Canadian
institutions have started to set up research labs specifically
dedicated to how we can bring on these ideas within finance.
So, from an organization point of view, it would definitely be
going on the right track and now I think the question is, how
can we apply these methods in a more principled way in your
organization.
Moderator: Ms Zhuo, do you have anything to add on that?
Ms. Zhuo: You know banking is a highly regulated industry and.
So there is a case out there about using AI models to approve
credit card applications, where the family members who applied
for credit card, and the wife getting approval for a lower limit
while the husband got much higher limit. So, the question is
how we can detect this type of bias in our machine learning
models. It is a problem that we can always solve first before we
implement the machine learning models in the financial
decision process.
Moderator: Thank you. How do you balance data privacy in
this period of time? For all three of you, let’s start with Dr.
Han.
Dr. Han: That's a very good question. In the medical industry,
it is a very critical to obtain the right medical data. From my
point of view as a research lab, we put a lot of attention to
develop tools to protect data privacy. Even when using the
Google Street view, we do a lot of work to blur out the license
plate, the human face, and the street address. Data privacy is
important, especially in the financial field. Now, I think, from
the research point of view we are developing tools, and we try
to develop new algorithms to make sure the data is statistically
important for the feature to be launched while protecting (our
customers’) privacy.
MARCH 2021