Introduction to Generative AI n explainable AI

Summary

Annie Thomas discusses advancements in AI, focusing on machine learning and natural language processing. She introduces generative AI, capable of creating new data like text or images, and explainable AI, which counters the 'black box' nature of ML by providing transparent decision-making processes. Thomas also touches on large language models like GPT and Transformer architecture, highlighting their applications in various tasks. She emphasizes the importance of local language processing and suggests research opportunities in summarization, spelling correction, and sentiment analysis.

Takeaways

• 🌟 Annie Thomas introduces herself as a speaker from India, focusing on advancements in AI, particularly in machine learning and natural language processing.
• 📈 She discusses the importance of generative AI, which can create new content like text, images, or audio based on training data, and discriminative models that classify and predict from existing data.
• 💡 Generative AI, also known as chain AI, operates on unstructured data and is exemplified by predictive text features on mobile devices.
• 🧠 Large language models are highlighted as a significant development in AI, capable of understanding context and generating human-like text.
• 🌐 The talk covers the rise of Transformers, a deep learning architecture that has revolutionized various AI applications beyond just language processing.
• 🔍 Explainable AI is introduced as a response to the 'black box' issue in machine learning, aiming to make AI decisions understandable and trustworthy, especially in critical fields like medicine and finance.
• 📊 Annie emphasizes the four principles of explainable AI: meaningful explanation, accuracy, knowledge limitation, and user assistance.
• 🔎 The script mentions the challenges of balancing interpretability and accuracy in explainable AI and the need for clear abstractions in explanations.
• 📚 Research opportunities in natural language processing are explored, including text summarization, spell checkers, and news article classification, with a focus on local languages.
• 📊 The importance of differentiating between extractive and abstractive summarization is highlighted, where the former extracts parts of the original text and the latter generates new text.
• 📈 The script concludes with the speaker's encouragement for researchers to explore various summarization techniques and semantic analysis to improve AI models.

Q & A

Who is Annie Thomas and what is her background?

Annie Thomas is the speaker of the keynote session. She is from the Line Street of Technology dirt, from Sattisgarh State, and from the country of India.

What is the main topic of Annie Thomas' keynote session?

The main topic of the keynote session is research prospects in the field of machine learning and natural language processing.

What are the two new aspects of AI in natural language processing mentioned by Annie Thomas?

The two new aspects of AI in natural language processing mentioned are generative AI and explainable AI.

What is the difference between supervised and unsupervised machine learning models?

Supervised models have a set of labels that are fixed and always existing, allowing for checking predictions for correctness. Unsupervised models do not have a class of labels and predict based on the model-generated data, which can be unstructured.

What is generative AI and how does it differ from discriminative models?

Generative AI generates new data based on the training provided and understands the distribution of data. Discriminative models, on the other hand, are used to classify, predict, and cluster and are trained on labeled data.

Can you provide an example of generative AI mentioned in the script?

An example of generative AI is the predictive text feature on mobile phones, which suggests the next word in a sentence based on patterns learned from previous inputs.

What are Foundation models in the context of generative AI?

Foundation models are large language models that work on unstructured data to generate new patterns and can generate new content such as text, images, or audio based on the training provided.

What is the importance of explainable AI in industries?

Explainable AI counters the 'Black Box' tendency of machine learning by providing explanations for decisions, which is crucial in domains like medicine, defense, finance, and law to build trust in the algorithms.

What are the four principles of explainable AI?

The four principles of explainable AI are providing meaningful explanations, ensuring accuracy, having a high knowledge limit, and assisting users in determining appropriate trust in the system.

What are the challenges faced in explainable AI?

Challenges in explainable AI include contrasting interpretability and accuracy, the need for abstractions to clarify explanations, and the difficulty of providing explanations that meet human accuracy levels.

What are some applications of natural language processing mentioned in the script?

Some applications of natural language processing mentioned are text summarization, spell checkers, news article classification, and semantic analysis of reviews.

Outlines

00:00

💡 Introduction to AI and Machine Learning

Annie Thomas introduces herself and the keynote session's focus on research prospects in machine learning and natural language processing. She distinguishes between generative AI, which creates new content like text or images, and explainable AI, aiming to make AI decisions understandable. Annie explains AI, machine learning as a subset of AI, and deep learning as a subset of machine learning with multiple hidden layers. She further discusses supervised and unsupervised learning, with the former using labeled data and the latter generating models from unstructured data.

05:00

📚 Deep Dive into Generative AI

The paragraph delves into generative AI, contrasting it with discriminative models. Generative models create new data based on training, understanding data distribution to generate examples. Annie provides an example with predictive text on mobile phones, illustrating how generative AI learns patterns to suggest the next word in a sentence. She also introduces foundation models and large language models capable of handling vast amounts of text, image, or video data to generate human-like text, with examples like GPT-3 and ChatGPT. The paragraph also mentions the rise of Transformer models post-2017, which are versatile for various tasks beyond just language processing.

10:02

🔍 The Emergence of Explainable AI

Explainable AI is introduced as a response to the 'black box' issue in machine learning, where decisions lack transparency. It's crucial for domains like medicine and finance to understand and trust AI algorithms. The paragraph outlines four principles of explainable AI: providing meaningful, accurate explanations within the designed operational limits. Challenges include balancing interpretability with accuracy and using abstractions for clarity. The categorization of explainable AI is discussed, including model-agnostic vs. model-specific and global vs. local explanations.

15:03

🌐 Local Language Processing and Applications

Annie discusses her work on local languages, focusing on text summarization. She outlines various types of summarization, including single and multi-document, and based on different criteria like informative or evaluative. The paragraph mentions different summarization techniques like extractive and abstractive, with the latter requiring more sophisticated models. She also touches on spelling and grammar correction for local languages and classifying news articles into predefined categories, emphasizing the challenges due to language diversity.

20:04

📊 Semantic Analysis and News Classification

The final paragraph discusses ongoing work on semantic analysis of social media reviews to improve sentiment analysis models. Annie mentions the application of data mining techniques to enhance accuracy. Additionally, she talks about classifying news articles into appropriate sections like national, international, business, etc., using discriminative models trained on the Hindi language. The paragraph highlights the collection of large datasets and the development of models to address the unique challenges of processing different languages.

Keywords

💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the video, AI is the overarching theme, with a focus on how it can make machines intelligent, efficient, and capable of understanding and predicting human-like responses.

💡Machine Learning

Machine Learning is a subset of AI that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. It is mentioned as a part of AI where systems learn according to the knowledge base provided.

💡Deep Learning

Deep Learning is a subset of machine learning with neural networks having multiple hidden layers, allowing the model to learn and make decisions based on patterns in large amounts of data. It is discussed in the context of providing more precise and evaluative results for problem statements.

💡Generative AI

Generative AI refers to AI models that can generate new data based on the training data provided. It is highlighted as a type of machine learning model that can create new data, such as text or images, rather than just classifying or predicting existing data.

💡Explainable AI

Explainable AI is a concept that counters the 'black box' nature of machine learning models by providing understandable explanations for their decisions. It is discussed as a necessity in fields like medicine and finance where understanding the decision-making process is crucial.

💡Supervised Learning

Supervised Learning is a type of machine learning where the model is trained on a dataset that includes input and output pairs. It is mentioned as a model that has a set of labels to check the correctness of predictions.

💡Unsupervised Learning

Unsupervised Learning is a type of machine learning where the model works with input data without labeled responses. It is discussed as a model that predicts based on the patterns it discovers in the data without any pre-existing labels.

💡Discriminative Models

Discriminative Models are used to classify, predict, or cluster data and are trained on labeled datasets. They are contrasted with generative models in the script, where the latter generates new data examples.

💡Large Language Models

Large Language Models are AI models specifically designed for natural language processing tasks. They are characterized by their large size and ability to generate human-like text. Examples mentioned include GPT-3 and ChatGPT.

💡Transformers

Transformers are a deep learning architecture that relies on attention mechanisms to process data. They are mentioned as a model that can be applied to a wide range of problems, including natural language tasks, after the advent of large language models.

💡Summarization

Summarization refers to the process of shortening a text while retaining the main points. It is discussed in the context of natural language processing, where the script mentions single document, multi-document, and different types of summarization based on the target audience.

Highlights

Introduction to the keynote session on Research prospects in machine learning and natural language processing.

Exploration of generative AI and explainable AI as new aspects of AI in natural language processing.

Description of AI's goal to make machines as intelligent as humans, with examples like Chat GPT.

Explanation of machine learning as a subset of AI and its function to make systems learn from a knowledge base.

Deep learning defined as a subset of machine learning with deep neural networks for precise results.

Differentiation between supervised and unsupervised machine learning models.

Introduction to generative AI, capable of generating new data based on training.

Example of generative AI in predictive text input on mobile phones.

Generative AI's ability to work with unstructured data like text, images, or audio.

Foundation models and large language models in the context of generative AI.

Research focus on local languages in natural language processing tasks.

Popular large language models like GPT-3 and their capabilities in text generation.

The rise of Transformers in deep learning architecture and their applications.

Explainable AI as a counter to the Black Box tendency of machine learning.

Importance of explainable AI in high-stakes domains like medicine, defense, and law.

Four principles of explainable AI: explanation, accuracy, knowledge limit, and user assistance.

Challenges in explainable AI, including the trade-off between interpretability and accuracy.

Categorization of explainable AI into model-agnostic, model-specific, global, and local explanations.

Research opportunities in natural language processing, including text summarization and sentiment analysis.

Different types of text summarization: extractive, abstractive, monolingual, multilingual, and cross-lingual.

Spell checkers and grammatical correction systems for local languages.

Automatic classification of news articles into predefined categories using discriminative models.

Semantic analysis of social media reviews and sentiment analysis models.

Transcripts

good morning everyone

hope you all are fine

let me introduce myself I'm Annie Thomas

from the line Street of Technology dirt

from sattisgarh State and from the

country of India

today's keynote session is on Research

prospects in the field of machine

learning and natural language processing

the two new aspects of AI in the field

of natural language processing I want to

introduce over here that is generative

Ai and explainable AI

so as we all know about AI

artificial intelligence everyone knows

nowadays is to make

the machine as intelligence intelligent

as human beings and with the Advent of

chat GPT and all all are familiar with

the term AI nowadays to make the machine

as efficient as humans

so machine learning that is a part of AI

there is a subset of AI and in machine

learning we make the system learn

according to the knowledge base that is

provided again going deeper into that we

have deep learning that is a subset of

machine learning where we have deep

neural networks to

um

have more precise and evaluative results

for the problem statements it has more

than one hidden layers which makes it

possible to go deep into the network

frames so depending on all this we'll go

further into generative AI the machine

learning model can be divided into

supervised and unsupervised models

supervised model has a set of labels

that are fixed and they are

always existing like if we get an output

also we will have a label attached to

the output from which we can we can

always

check whether our predictions correct or

not the unsupervised models they are not

having a class of labels we have to

predict based on the model that is

generated and maybe that may be

unstructured data or unstructured data

so unsupervised learning will be on the

basis of the input data that is given a

model will be working on that input data

and it will be having a generated

example

going further into deep learning we

again have two types discriminative and

generative

discriminative is used to classify

predict cluster and it is trained on a

data set of labeled data we already have

the labeled data then

um

the training is easy as we have to put

it into one particular class and it

learns the relationship between the data

points and the labels but generative is

different that it generates new data

based on the training that is provided

to it so it has to understand the

distribution of data and How likely a

given example is going to be

categorized into that generative example

but in discriminative we don't have to

generate any new data the best example

of generative AI when it came into

existence is everyone is having a mobile

and we type the text so when we are

typing the text it will give you like if

I type Delight will give the next

predicted word as Institute and the next

predicted word or as of Billa Institute

of Technology I'll get an option for

that because it learns from the patterns

I have been doing all these these days

so the it is predicting the next word in

a sentence that's the best example of a

generative AI now if we come to the

concept of generative AI it is also

called chain AI it is artificial

intelligence capable of generating text

images or other media using the

generative models and

with the right side if whatever I have

written if you can see not j a i when y

input is a number or a discrete or a

class or a probability but it is when it

is a natural language text or an image

or an audio that shows that it works on

unstructured data so generative AI works

on unstructured data to generate new

patterns we have Foundation models large

language models which we'll be

discussing in the further slides

so now we understood that generative AI

where we generate new content content

and that content may be a text or image

or audio

so if we see the category of the

supervised semi-supervised unsupervised

chain AI you can have input as training

code or label data or unlabeled data and

the foundation model works on all this

and generates new content and the output

can be a text or image or code or a

video audio anything

so

the main concept here of generative AI

is to generate the new content based on

the training that is provided so here

since I am interested in text so I have

taken only the text input you can also

take image or audio so in the

uh some research which we are pursuing

here we are working on text and that

also we are working on the local

languages because in English already the

work uh text to text if you see the post

column translation summarization

question answering grammar correction

all these my research Scholars are doing

in the for the local languages here

because for English we are having

already uh or much work is being done on

this uh in these areas I'm not working

in image and audio and musicians but you

can because image generation text to

image text to video text to speech all

these are again natural language

processing generative way I can be used

in all these areas

we see we know about the Google's

Foundation models that is Palm API for

text is a very popular model word we use

vit for vision we use

vitgp2 blip vqa all these models we are

using now uh when we work talk about

natural language processing we talked

about large language models a large

language model is specifically designed

and trained for natural language

processing tasks what is the

characterization of large language

models is its large size so it can be

working for worst amounts of Text data

or it can be working for other areas

also and it is capable of generating

human-like text understanding context

answering questions all these things the

large language models are doing nowadays

the notable examples are open AIS GPT

GPT 3 GPT 4 and chat PT and birth all

these examples are very popular they

train

um

Text data or image or video and are

capable of generating the new text based

on the training data the large language

models are able to handle the large

volume of data that is taken from

internet or from social media or any

other place this shows some other models

which are also popular

then after the large language models

came into existence

there was an era of Transformers

a Transformer is a deep learning

architecture that relies on the parallel

multi-head attention mechanism the

modern Transformer it was proposed in

the year 2017 I think and the paper from

which it was introduced was attention is

all you need so attention was given to

the important data which has to be

picked up from the given set of huge

databases and the less attention has to

be provided to the data which is not of

much importance on the basis of that the

Transformers were broken so so a general

pre-trained Transformer is a more

broader term for models based on

transform architecture with these models

they can be applied to wide range of

problems not only for natural language

tasks computer vision speech recognition

reinforcement learning all the things it

is being used then these models are

pre-trained on large data sets and can

be fine-tuned for specific tasks some

more examples I'll be showing you in the

next slide that is Vision Transformer

the detr conformer swim Transformer

perceiver and perceiver IO this is what

we have talked about generative Ai and

the concepts which are used in

generative AI moving on ahead to

explainable AI that is also a new term

that is being used in the industry

nowadays and

when we talk about explainable AI it

counters the Black Box tendency of

machine learning where even the AI

designers cannot explain why it arrived

at a specific decision now what happens

is we see the black box design of

machine learning where the we have

reached to a conclusion but there needs

to be an explanation about how we are

getting the results why it is not like

this and why it is like this and why

this explainable AI was needed when we

are getting the results why is this

explainable AI was needed that was

because the domains like medicine

Defense Finance and law where it is

crucial to understand the decisions and

build trust in the algorithms they have

made their algorithms we are employing

algorithms we are getting the results

unless and until we have trust in that

those models who will is going to use

those models you know chargpt and Google

search are so popular because of the

trust we have in those systems that we

are getting such important informations

according to the need of the query we

are putting there so explainable AI came

into existence to attract find out how

we are getting the results

so four principles of explainable AI we

have done this four principles as

explanation meaningful explanation

accuracy we we are giving an explanation

but that explanation should have the

satisfactory level of accuracy not that

if it is 15 accurate we cannot say we

are having a good explanation so

explanation is there that is Meaningful

that is interpretable but that

explanation should have an accuracy and

the knowledge limit should be so high

that explainable a I will be able to

provide the system to operate under the

conditions for which it was designed and

when it should reach the sufficient

confidence in its output

so

the assisting its users and determining

appropriate trust that suppose part

trust we develop in the system in the

model which we are generating and the

second part is we have an

interpretability and explainability

mechanism which can explain to the users

how it is working so the next part is

that all that is having so many

advantages there are so many

disadvantages or issues challenges which

are existing with this one is

contrasting the interpretability and

accuracy we know we have to reach the

human accuracy levels sometimes it may

the explainable AI may not be able to

give the correct explanations it may be

contrasting with the human explanations

and those cases which has to be dealt in

the model as we improved like that so

still it exists these issues exist

describing the

and there should be the use of

abstractions to clarify the explanations

now based on this explainable AI xaia it

is also called we see there is

categorization based on agnosticity that

is model agnostic or it is model

specific if it is applied to all the

model types then it is called Model

agnostic and if it can be applied to

only particular specific model types for

which particular task it's been made the

latest model specific that is the

categorization of economic agnosticity

now depending on the scope we have

Global explanation or local explanation

if you want some part or the prediction

of some particular area only to be

displayed it may say it is local

explanation and we we want the

explanation of the whole model how it is

looking then we go for Global

explanation but local expression nations

are also important because individual

predictions are also taken into account

in X AI because Global explanation can

be given on a larger scale but local

explanation need to go deeper into the

problem so that is again an area where

the research can be done and the

different problems can have different

perspectives different aspects to

include either model agnostic model spec

or Global or local explanation

as I told you we are working on local

languages and on the basis of this local

languages we are doing text

summarization

so again now for the persons who are

hearing me those who want to do

um

and research in the field of natural

language processing machine learning you

can go for single document summarization

you can go for multi-document

summarization you can take any of the

languages on which work has not been

done you can go for indicative

informative evaluative summarization any

of the tasks can be taken upon and then

on the basis of this you can generate

models generative models or

discriminative models which will help

you to give a proper text summarization

and on the basis of target audience

there is a generic summarization and

there is a query focused summarization

on the basis of type of summarizer you

have monolingual multilingual

cross-lingual all types of summarization

so these I'm I have only included this

slide so that if you have some have

interest in generic summarization some

in multilingual so you can pursue your

area of Interest

then the two main Concepts that come

over here is extractive and abstracted

in extractive summarization the parts of

the original text are taken to form the

summary like if we are

taking a document and from the document

we want to summarize it the important

sentences will be taken out and given as

the extracted summarization but in the

abstractive summarization you need to

generate the new text depending on the

extracted important parts of the

documents so all this has to be done

which shows that abstractive

summarization needs more

fields which have to be covered so that

the abstracted summarization finds out

the meaning and again generates the new

test text based on the training data

some comparative studies of peer-wise

Publications has been done

the New Concept which has been included

for the new language has been done in

using these steps from the input

pre-processing has been done on the

pre-processing part text cleaning was

done stockpot removal based on the local

language was done then feature

extraction and then optimization was

applied to find the text summary on the

basis of some of the models bioelestem

was applied word for supplied CPT was

applied and we could find that the

results are

we could compare the models on the basis

of the results and the new model

Innovative models are being generated by

the research scholars in this area

then next topic was spelling and

correction

so if we are can go for two types of

grammatically text can be corrected or

spelling can be corrected that again

leads to two different we are working on

spell Checkers only for that particular

language maybe in the future we'll be

working on grammatical checking of the

sentences as well so we know we have in

English we have so good systems when we

type of

word and if it is wrong it gives the

suggestions and also many systems are

already existing in binding to all the

particular softwares but for the local

languages still we wrote half the

systems like which finds the text errors

may be the typographic maybe the

syntactic maybe the discourse there are

non-word errors real word errors

phonological errors

transformationals deletion errors

insertion errors of substitution all

these are the type of Errors which exist

in the text errors and you can work on

any of this type to remove all such type

of Errors being on the domain or outside

that only

then another system was to classify the

news via articles into the predefined

classes of the newspapers so that

whenever we get an article it will

automatically be sent to the national

International Business Sports

entertainment health or weather and this

has been done on our

national language that is Hindi for this

we have made the different

categorizations labels and the

discriminative models are being prepared

to classify the news where articles into

proper particular

sex label sections so that the newspaper

can automatically be aligned to the

pages according to what they are

so for challenging task was because of

the every language is a challenging for

processing for every language is a

challenging task because it has

different different types of consonants

vowels combinations and maybe patterns

or different Center structures are

different

so from different uh data set

huge amount of data has been collected

and work has been done on this I am not

able to show you the particular

results of these models but I'd like to

tell you that the news via articles they

will be classified into predefined

labels another work that is going on is

semantic analysis of the

reviews which are collected from the

social media and based on those reviews

the sentiment analysis where the

sentiment analysis models have still not

been applied on and based on that

suggestion in mining techniques are

being proposed so that they give more

accuracy than the existing semantic

analysis models so that's all from my

side thank you for the patient here

thanks

foreign