ARTIFICIAL INTELLIGENCE (CLASS – IX)

Revision Learning Pack

WALT:
 Students will be able to define Generative Al & classify different kinds.
 Applying Generative Al tools to create content.
 Understanding the ethical considerations of using Generative Al.
Supervised Learning and Discriminative Modelling:
 Supervised Learning:
Supervised learning is a machine learning where a model is trained on a labelled dataset,
implying that each input data point is associated with a corresponding output label. The goal of
supervised learning is to learn the mapping between input data and output labels, enabling the
model to make predictions on new, unseen data.
For example: In the given image, first is the input image and characteristics of this image are
marked as boy and ball which can be seen in centre image. Now according to supervised learning, it
must learn the mapping between input labels and output labels, which is shown in last image and
highlights "ball" as red, "boy" as purple and "boy playing with a ball" in a rectangle.

 Discriminative Modelling:
Discriminative modelling is an approach in machine learning where the focus is on learning the
boundary or decision boundary that separates different classes or categories directly from the
data. So, if an image contains a combination of Dogs and Cats, the model can tell which is a Dog
and which is a Cat.

In supervised learning, discriminative modelling contrasts with generative modelling, where the
goal is to model the joint probability distribution of both the input features and the output labels.
Generative models can be used to generate new data points that resemble the training data, whereas
discriminative models are primarily focused on classification or regression tasks.
Unsupervised Learning and Generative Modelling:
 Unsupervised Learning:
Unsupervised learning is a type of machine learning where the model is trained on input data
without any corresponding output labels. The goal of unsupervised learning is to find patterns,
structure, or representations in the data without human intervention. An unsupervised learning
approach works on an unlabelled dataset. This means that the data which is fed to the machine is
random and there is no know-how available about it to the trainer.
 Generative Modelling:
Generative Modelling do not necessarily require labelled datasets. It can work with unlabelled
data to learn the underlying distribution of the data and can generate structured data from the
Random Noise dataset. So, if random images are fed as training data for the model it can create
relevant output based on the features of the input data. For example, if there are random images
which depict streets, cars, buildings, sky etc. Given a dataset of street images, a Generative
Modelling can learn to generate new street scenes that look like the ones in the dataset. In
another example, if given a dataset of news articles, a generative model can learn to generate
new articles that resemble the style and content of the training data.

Random Noise Dataset:

What is Generative AI?
• Generative artificial intelligence (AI) refers to the algorithms that generate new data that
resembles human-generated content, such as audio, code, images, text, simulations, and videos.
• This technology is trained with existing data and content, creating the potential for applications
such as natural language processing, computer vision, the metaverse, and speech synthesis.
Timeline of Generative AI:
Generative AI vs Conventional AI
In contrast to other forms of AI, Generative AI is specially made to produce new and unique content
rather than merely processing or categorizing already-existing data. Here are some significant
variations:

1. Generative Adversarial Networks (GANs):

GANs are neural networks that work to produce fresh data. It is made up of two neural networks
which work together in a unique adversarial process to create realistic synthetic data. These two
neural networks are as follows:
• Generator Network: It produces the data that is as close as possible to real data.
• Discriminator Network: It analyses the data and provides feedback, i.e. it takes real data and the
data generated by the Generator as input and attempts to distinguish between the two.
Some of the examples of GANs are as follows:
• It can create portraits of non-existing people.
• It can convert images from day to night.
• It can generate images based on textual description, for example, if we give description of a bird
then it will create an image that is similar to the description.
• It can generate realistic video which can be used in film production, video games, and generating
synthetic data for training other AI models, etc.
2. Variational Autoencoders (VAEs):
This is another class of generative models. To produce fresh data, VAES learn the distribution of
the data and then sample from it.
Some of the examples of VAES are as follows:
• It can generate new images like the given training set. For instance, a VAE trained on images of
faces can generate new, realistic-looking faces.
• VAE's can produce new text that follows the same style and structure as the training data, assisting
writers with drafts and ideas.
• It can be used for composing new music pieces or creating sound effects, music composition etc.
3. Recurrent Networks (RNNs):
RNNs are a special class of neural networks that excel at handling sequential data, like music or
text. They excel at tasks where the order of the data points is important, as they can remember
previous inputs and use this information to influence current outputs.
Some of the examples of RNNs are as follows:
• It can generate novel text in the style of a specific author or genre, like creating new sentences that
mimic the style of Shakespeare or generating dialogue for a chatbot.
• It can predict the next character or word in a sequence, like autocomplete features in text editors
and predictive text input on smartphones.
• It can be used to predict future values in a time series, such as stock prices or weather data, by
learning patterns from historical data.
4. Autoencoders (AEs):
These are Neural networks that have been trained to learn a compressed representation of data. They
work by compressing the data into a lower-dimensional form (encoding) and then decompressing it
back to its original form (decoding). This process helps the network learn the most important
features of the data.
Some of the examples of AEs are as follows:
• It can help in cleaning up noisy images to produce clear and highly realistic samples.
• It can help in compressing high-resolution images for efficient storage and transmission.
• It can create artistic images based on learned features from famous paintings.
• It can help in drug discovery by learning and generating molecular structures that have desirable
properties.
Example of Generative AI:
1) Art:
Generative AI can create new artworks by learning styles from famous painters and generating
novel pieces in similar styles. For example:
• AI artists like "AI Portraits" and "DeepArt" have gained popularity for their ability to create
visually stunning images.
• The Next Rembrandt project used data analysis and 3D printing to create a new painting in the
style of Rembrandt.
2) Music:
Generative AI is transforming the music industry by enabling the creation of new music, either
through composing original pieces or remixing existing ones.
One prominent example of this innovation is AIVA, an AI composer capable of creating original
music in various genres.
3) Language:
Generative AI is being used to generate new languages, such as chatbots that can hold conversations
with users or natural language generation systems that can produce written content.
GenAI in Language: Language Translation, Text to speech, Data creation.

Types of Generative AI
Generative AI comes in a variety of forms, each with unique advantages and uses. Some of the most
typical varieties are listed below:
1. Text Generation:
Language Models: These generate coherent text based on prompts. Examples include OpenAI's
GPT-3 and GPT-4, Google's BERT, and T5.
Chatbots: AI designed to simulate conversation with users. Examples include OpenAI's ChatGPT
and Google's Meena.
2. Image Generation:
GANs (Generative Adversarial Networks): These generate realistic images from random noise.
Examples include StyleGAN and BigGAN.
VAEs (Variational Autoencoders): These generate images by learning the distribution of input data.
Diffusion Models: These iteratively refine noise to generate images, such as DALL-E 2 and Stable
Diffusion.
3. Audio Generation:
Music Generation: AI that composes music. Examples include OpenAI's MuseNet and Jukedeck.
Speech Synthesis: Converts text to natural-sounding speech. Examples include Google WaveNet
and Amazon Polly.
4. Video Generation: DeepFakes: AI-generated videos that can superimpose faces. Examples
include DeepFaceLab and FaceApp.
Generative Video Models: AI that creates new video content. Examples include VQ-VAE and
MoCoGAN.
5. 3D Object Generation:
3D GANs: Generate three-dimensional objects. Examples include 3D-GAN and PointNet.
6. Code Generation:
AI Code Assistants: These generate or autocomplete code. Examples include GitHub Copilot and
OpenAI Codex.
Benefits of using Generative AI

Limitations of Using Generative AI

• Ethical Concerns: The creation and spread of fake content, such as deep fake videos, fake news
articles, and forged documents, by generative AI can deceive and manipulate people, leading to
misinformation, identity theft, and privacy breaches. Robust regulations and ethical guidelines are
needed to address these issues.
• Bias and Fairness: Generative AI models can inherit biases from their training data, resulting in
discriminatory outcomes and perpetuating biases against certain groups. Developers need to actively
work on mitigating bias to ensure fairness and equity in AI systems.
• Data Dependency: These models require vast amounts of data to generate content effectively.
They struggle in situations with scarce or unrepresentative data, limiting their usefulness in niche
domains.
• Energy Consumption: Training large generative AI models demands significant energy and
computing resources, raising concerns about the environmental impact and sustainability of AI,
particularly in data centers. Researchers are striving to develop more energy-efficient AI models,
but this remains a significant limitation.
• Cost of Development: The development and training of generative AI models can be extremely
expensive, requiring access to high-quality data, powerful hardware, and expertise in machine
learning. This creates a digital divide, limiting the democratization of AI technology.
Ethical considerations of using Generative AI

The Potential Negative Impact on Society

Generative AI can be used to create fake news or deep fakes that can spread misinformation and
manipulate public opinion.
Lead to job displacement for humans who previously performed these tasks.
Generative AI has the potential to generate sensitive personal information, such as social security
numbers or medical records, which could be used for malicious purposes.
Responsible Use of Generative AI
 Ensuring that the training data used are diverse and representative.
 The outputs are scrutinized for bias and misinformation.
 Prioritizing user privacy and consent,
 Having clear guidelines around ownership and attribution of generative content.
 Engaging in public discussions around the social and ethical implications of this technology to
ensure that it is developed and used in ways that are beneficial to society.