Transcript
This transcript was autogenerated. To make changes, submit a PR.
Hello, I'm Parvin Ghasemzadeh.
I'm a senior software engineer at Amazon.
And in this talk, I'll be talking about the role of prompt engineering in ethical
AI and how we can mitigate bias and reduce hallucinations in AI systems.
So here is the agenda for today's talk.
We'll start with a brief introduction to LLMs Then I will explain two main
Challenges we face in this area by some house nation in AI systems And we'll
then discuss about methods to address these challenges and I will wrap up
with some key takeaways Let's start with the definition of large language
models And I thought it's better to ask the definition of one of the LLM to see
how they define itself So here I used
So basically it says LLM is a type of AI model designed to understand
and generate human like text.
And these models are trained on vast amount of data which enables
them to predict and produce response based on the inputs they receive.
So a few keywords here are important to note and remember.
Generate human like text and trained on vast amount of data.
So just let's keep these two phrases in mind, and we'll come
back to this to understand how it impacts the outcome of the LLAMs.
In general, LLAMs are powerful tools, but their response depends on the
data they have trained on, which can lead to issues we'll discuss shortly.
So let's discuss a bit about the background.
So the transformer architecture is the background of the large language
models, which allows AI to learn language patterns and generate text.
So in 2017, Google researchers introduced the first transformer model, which is
described in the paper, attention is all you need paper, which is added here, which
light the groundwork for the modern LLMs.
So before transformers, all those who relied on the sequence, like
a processing, takes word by word, however with transformers, it scans,
transformers look at the entire sentence, and, To understand the context better.
here the model uses a special technique called self attention, which allows to
look at the, all the words in the sentence at once, to understand the context and
understand the relationship between the words, to predict the next word.
after this transformer model, OpenAI created its own model.
First, generative pre trained transform model in 2018 and 2020.
They created the GPT 3 model, which is one of the largest language model created for
now, and it has 175 billion parameters.
And after that, so like many other models with billions, even trillions of
parameters created by different companies.
And the main thing to remember here is that all these models are trained using
the data from like the public source from books, from websites, wikipedia,
and like public forms like reddit.
And let's keep in mind that like These are the, not the accurate source of
the information, so it might have, some misinformation, some, bias,
within the data, which could, impact the outcome of the LLMs, which we will
discuss shortly in the upcoming slides.
So now let's talk about a bit on the ethical part of the AI.
So what do we mean by the ethical AI?
It's mainly about like developing an AI stems that align with the
core ethical principles and values.
fairness, transparency, accountability, privacy, safety and inclusivity.
in short, ethical AI systems should treat everyone fairly, regardless of their race,
age, gender or any other characteristics.
They should also make their process clear, keep personal data secure and safe,
and respect the diverse perspectives.
So the two key issues we are facing with ethical AI are bias
and hallucination, which I will go into details in the next slide.
So let's start with the bias.
So bias in AI mainly refers to the systematic errors that cause
certain groups or individuals to be treated unfairly by the model.
So for example, if let's think about the example for if the face recognition
application has been trained on, a lighter skinned individual, it may
not accurately identify the darker skin tones, which can lead to the
error rates for certain groups.
So this, like a bias, can come from multiple sources.
And the training data, is, this is the main source as we discussed before.
So the, all these metals are trained using the, the information
from the public source.
And this could impact the bias on the output of the as.
So this, like the data might be imbalanced, so the model design
could have some limitations.
So just feature selection is also important.
Explaining some.
Features could impact the AI's decision.
Labeling might also be subjective here.
So if the annotators, who, manually label some training data, have some
bias on specific topics, might also impact the, decision of the AI.
And also like even the biased input from the users could
also impact the final output.
So if this data is used for retraining the models.
And also the ethical impact of bias is significant.
it can lead to the unfair treatment of certain groups and ultimately
people lose trust in AI systems.
And for the hallucination, so this is the mainly commonly
common term used in AI world.
when the model generates incorrect or fabricated information.
So this can also, happen for several reasons.
So similar to the bias.
So training data is also plays a huge role here, so low quality data also,
cause some, AI models to hallucinate, to provide a fabricated and misinformation.
So lack of sufficient context also cause the, issues here.
Or, a complex language, inputs, Sarcasm, like cultural references, could be
difficult for AI models to understand, and in that case AI models would just fill
the gaps with some fabricated information.
And also, this can also lead to the misinformation and reduce reliability
in AI systems, which is why it's important to address all these issues.
let's see what methods can be used to mitigate the bias.
here are the few strategies listed below.
as we discussed, the source of the issues is coming from the training data.
using diverse training data sets that cover the range of
demographics is a good idea.
key starting point here.
So for example, so if you are building a assistant like AI assistant to answer
medical questions, we should use a data set that reflects diverse patients like
demographics and conditions to avoid like a bias recommendation from the other side.
So another method is fine tuning and debiasing where we retrain model.
with specific data that helps to reduce the bias.
So doing some bias audits and transparency about how these
algorithms function are also important.
And finally, updating the models regularly, keeping them
aligned with social norms would help to minimize the bias.
And similar to the bias, in order to reduce hallucinations, using high
quality training data is important.
So the more accurate the data, the better the model's output it is.
So contextual training is another approach that we can use to reduce the
hallucination here to where we provide the model with additional background
information for specific tasks.
For example, if we train the customer service chatbot on
domain specific language, it will give more relevant answers.
So fine tuning the pre trained models on this focus dataset also increase
the reliability of the system.
for example, I guess AI system design for the medical advice should The
fine tuned on the verified health care data have an irrelevant output.
So integrating with the external knowledge sources like a database
is going to help verify the facts and provide additional context and
help to reduce the hallucination.
And also continuous learning is important.
Which allows the model to keep improving based on the user interaction here.
So all the previous methods we discussed for both minimizing the bias and reducing
hallucinations is not scalable and cost efficient as they require retraining, fine
tuning, which are any costly operation.
So now we will discuss a few, prompt engineering methods, which will be
easy and quick to implement and try.
we'll discuss zero shot and few shot learning, chain of thought, and
modular reasoning, knowledge, and language, in short, Miracle Framework.
So let's start with the zero shot and few shot learning.
So mainly zero shot and few shot learning methods are useful in areas where we
don't have a lot of, labeled data.
And zero shot means, allows the model to apply its existing knowledge to, new
tasks without any specific examples.
So it relies on the training data.
Where in the few shot learning, we give model a few additional examples.
Helping it to understand what we expect, for example.
if we are building an AI system to, let's say, classify emails, we
can show, show it a few examples, which will adapt, more effectively.
here, in the, actually, in this screenshot shown here, shows the three examples,
zero shot, like a glitched email.
It doesn't provide any examples of one shot which provides a single additional
example and a few shot which provides multiple examples that shows the
translation of word from English to French and in the one shot it provides a single
example to show what the translation of word from English to French and the
few shot provides multiple examples.
Which increases the accuracy of the result of the AI system.
another method is the chain of thought approach.
So this is the one, from the paper created by the Google resource team.
and this, I'll add the link here.
this helps the method to think step by step and break down the complex
problems by reasoning through.
like at each step, so instead of jumping directly to the answer, so it just
asks the model to think step by step.
this is the screenshot, this is the example from the paper itself.
Make sure this is a good example that shows the combination of one shot
example and the chain of the tiles.
So in each, like in the standard prompting and the chain of the tile prompting, so
it provides an additional example before asking the, example to the model itself.
So in the first one, It provides the one example with math problem
and in the answer part it just directly provides the answer.
However, when it asks the new math problem it just couldn't
find the answer correctly.
In the second one, the chain of 10 problem, prompting part, so it provides
a math, same math problem, but in the answer, so it just gives a, like a step
by step explanation more, explain the reasoning of the answer here, which
helps the, model to respond correctly.
So in the, in that case, so it's able to correctly find the math,
answer of the math problem.
And the last method is the Miracle Framework created by the AI21 Labs.
basically it suggests to enhance the capabilities of language
models by integrating with external tools and knowledge sources.
Some examples are like, using it for, integrating with the additional source to,
to, check the weather, real time weather, or for the financial applications, maybe
integrating with the financial source to check the real time, stock prices.
Could be a good example here and in this screenshot below actually shows also like
another use case for the math problems, so which is able to find the answers for
the simple math problems but struggles to, calculate the complex math problems.
So the calculator tool or calculator application is another good use case here.
So basically it just.
integrates with the calculator application and parse the user inputs
and pass the parameters to the Miracle framework like through the APS which
calculates and responds back to the correct result for the math problems.
So just to wrap up, here are the key points to remember
as part of this presentation.
So ethical AI is important for creating systems that are fair,
transparent, and accountable.
We should also remember that prompt engineering is a powerful
tool to address issues of bias and hallucinations, which makes the AI
system more reliable and transportive.
It's also important to remember that the ethical is a continuous
journey, so we need to continuously, evaluate, update and take the diverse
perspectives into account in order to tackle new ethical challenges here.
And finally, Organizations must commit to ethical practices in AI development
to ensure technology benefits society.
So here are the list of references I used throughout the slides
and thank you for listening.
Please feel free to connect me on LinkedIn for further discussion and questions.