Generative AI in Wireless Networks:

The O-RAN Chatbot Use Case

June 18, 2024
Agenda

• Generative AI in Wireless Networks

• O-RAN Chatbot
• Future Trends: Generative AI in 6G
Networks
Generative AI in Wireless Networks
 What is Generative AI?
Exploring generative AI: key models and mechanisms for data generation

• Definition: Generative AI models are a class of machine learning architectures designed to generate new data
instances that resemble the training data. These models learn the distribution of data in a training set and can
generate new data points with similar characteristics. These are the common features of Generative AI algorithms:
• Learn Data Distribution
• Unsupervised/Semi-supervised Learning
• Use of Latent Space and Conditioning
• Sample Generation
• Types of Generative AI

Generative Adversarial Networks Variational Auto Encoders

Learn through a competitive process Use a probabilistic approach to
where the generator aims to produce encode input data into a
increasingly realistic data, and the compressed representation and
discriminator aims to distinguish then decode it back to
generated data from real data. generate new data.

Diffusion Models Transformer Models

Learn to reverse a process that adds Learn by using self-attention
noise to data, effectively learning to mechanisms to weigh the significance
generate clean data from random of different words in a sentence,
noise. enabling it to understand the context
and relationships within the text

[1] "At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless Intelligence"
 Applications of Generative AI in Wireless Networks
Empowering wireless networks with generative AI

GANs VAEs Diffusion Models Transformer Models

❑ Network Security: ❑ Data Compression: VAEs ❑ Noise and Distortion ❑ Channel Estimation: Self-
Generator makes synthetic compress data into a Removal: Models denoise attention captures long-
attack data; discriminator lower-dimensional space and reconstruct signals, in range dependencies in
evaluates. for efficient transmission. noisy environments. massive MIMO processing

❑ Spectrum Sensing: GANs ❑ Joint Source Channel ❑ Signal Processing at the ❑ Chatbots using Large
create data mimicking Coding: VAEs balance data Receiver: Models enhance Language Models (LLMs):
radio environments to compression and error system reliability by LLMs enhance Telco
optimize frequency use. correction, enhancing reconstructing original services, examples include:
robustness against signals from noisy inputs. ❑ Customer Support:
❑ Anomaly Detection: channel noise. Nois Responses to
Can you service
Generator creates outlier ❑ Waveform Shaping at the inquires
data signaling potential ❑ Signal Processing: VAEs Transmitter: Diffusion ❑ Network Fault
failures; discriminator filter noise by learning models pre-optimize Detection: Identify
checks plausibility. compact signal waveforms for channel Network Issues
representations. conditions. ❑ Configuration:
❑ Channel Modeling: GANs Recommends Network
simulate channel ❑ Network Resource ❑ Energy-Efficient Routing: Settings
conditions for Management: VAEs Algorithms mimic natural ❑ Standards Compliance:
communication system optimize networks by diffusion to optimize data Assists in implementing
testing.. optimizing resources distribution in networks. compliance

[1] "At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless Intelligence"
 Applying GANs to Channel Modeling
Exploring enhanced GAN architectures for direct prediction of multipath components in indoor environments

3D Object

Interaction
Segmentation

Surface
Surface
Classification MPC Parameters
• Delay
• AoA
• AoD
• PathLoss
Spatial Rendering • Doppler Shift
GAN
• Gain/Amplitude
TRX Location
Radiance Network
Synthetic Data

Ray Aggregation
RX Location

GAN is applied where a generator is used to generate predicted MPCs based on the input of an indoor snapshot and TX-RX position, a
discriminator makes binary decision on whether the MPC is real, or fake based on segmentation information and propagation
knowledge distillation

[2] "Physics-informed Generalizable Wireless Channel Modeling with Segmentation and Deep Learning: Fundamentals, Methodologies, and Challenges."
 Network Optimization with Conditional Variational Autoencoder
Using conditional variational autoencoders for spectrum and resource allocation in wireless networks

Training
cu3
1 0 .. 0.8 0

cu2 cu4

Z
cu1
cuj Assignment Encoder Decoding Assignment

0.34 0.84
D2D Pair
di Tx
D2D Pair
d2 Rx
Tx

Tx
Rx
D2D Pair
Cost Matrix Encoder
d3
Rx
Inference

0.34 0.84 0.8 0

Conditional Variational Autoencoders (CVAEs) are used

to optimize resource allocation by solving linear sum
assignment problems, significantly improving resource Z
allocation in device-to-device communication systems Cost Matrix Encoder Decoding Assignment

[4] "Generative neural network-based spectrum sharing using linear sum assignment problems
."
 Applying Diffusion Models for Channel Estimation
MIMO channel estimation using score-based models — a type of generative AI diffusion models

Adaptative
Beamforming

Signal to Noise Ratio

Spatial
Channel State Prediction Multiplexing
Channel Impulse • Attenuation
Response • Delay
• Dispersion Inference
Optimiation
Channel Coefficients DIFFUSION
Model
Power
Control

Diffusion models are applied for channel estimation in MIMO systems by leveraging their ability to learn from noisy and incomplete
data. This allows them to effectively reverse the noise addition process, thus aiding in reconstructing the original signal and improving
channel estimation accuracy in MIMO systems

[3] "Physics-informed Generalizable Wireless Channel Modeling with Segmentation and Deep Learning: Fundamentals, Methodologies, and Challenges."
 Chatbots with Telco Knowledge
Map from “all previous words” to “next word”

A few thousand
Predict the next word or
previous words for
group of words
context

Through hard work, he supported Transformer Architecture

Deep Neural Network “family”
himself and his •••

Because it crossed state lines, that

criminal behavior attracted the “FBI”
attention of the •••

Bender Rodríguez was the one of

the stars of the television show ••• “Futurama”

// loop over the string

int i; “strlen”
for (i = 0; i < •••

This restaurant was fabulous! “five”

My star rating is •••
O-RAN Chatbot
Large Language Models (LLMs) Are Powerful Tools but Not Accurate Enough for
Enterprises
Without a connection to enterprise data sources, LLMs cannot provide accurate information

Lacking sufficient domain-specific data

Response Prompt

User

Risk of outdated information

?
Hallucinations

Foundation Model
Retrieval Augmented Generation (RAG) Lets Enterprises Talk to Their Data
Enable LLMs to provide up to date and domain-specific answers

Response Prompt

User Domain-Specific
Data

Data Framework Text

for LLMs Embedding

Ranked Data

Foundation Model Vector

Database (DB)
 O-RAN Chatbot
Bot designed to answer Questions about O-RAN standards.

ORAN Embedding Milvus Reranking

Standards Database
How O-RAN enables Network
Slicing ?

NIM LLM
User Frontend Chain Server Deployment

O-RAN enables network slicing through several mechanism. First the O-RAN architecture
support slicing features as defined by 3GPP, including the ability to associate an instance of
an O-RAN network function with one of more slice. Additionally, O-RAN, utilizes the E2
inference and the assistance of Near RT RIC dynamic slice optimization with enhance O1……

[5] "GenerativeAIExamples/experimental/oran-chatbot-multimodal at main · NVIDIA/GenerativeAIExamples (github.com)"

 O-RAN Chatbot Optimization Process
Steps to improve the accuracy and optimize RAG systems

  
Evaluation Dataset Evaluation Framework Test Retrieval
Gather a set of questions and Created evaluation framework in Verify the right context is
answers along with context Python to compare Chatbot retrieved comparing with source
provided by O-RAN experts that results with “ground truth” of truth.
will be considered source of truth evaluation dataset. Retrieval Metrics: Recall,
for Bot evaluation. precision, F1, Relevancy

 
Human Evaluation Test Accuracy
Gather Domain experts and Verify the right context is
user's feedback. Improve your retrieved comparing with source
evaluation dataset with curated of truth.
data from feedback. Accuracy Metrics: Faithfulness,
Relevancy, LLM-as-Judge
Future Trends:
Generative AI in 6G Networks
 •Diffusion Models: These models excel in creating highly detailed, realistic visual content for dynamic and interactive 6G applications like VR and AR, enhancing the visual fidelity and realism necessary for immersive experiences.
•Transformer
•Diffusion Models:
Models: These
Transformers
models facilitate the generation
excel in creating of contextually
highly detailed, relevant,
realistic coherent
visual content fortextual content
dynamic for multifaceted
and interactive 6G communications,
6G applications like VR andenriching user interactions
AR, enhancing across and
the visual fidelity varied digitalnecessary
realism media. for immersive experiences.
•Transformer
•TransformerModels: Transformers
Models: facilitate
Transformers improve thereal-time
generation of processing
data contextuallyandrelevant, coherent textual
decision-making in edgecontent for multifaceted
computing 6Gmanaging
by efficiently communications, enriching
sequential user interactions
data streams, across
enabling faster andvaried
moredigital media.
accurate responses directly at the network edge.
•Transformer
•Transformer
Models: Transformers
Models: facilitate
Transformers improve thereal-time
generation of processing
data contextuallyandrelevant, coherent textual
decision-making content
in edge for multifaceted
computing by efficiently6Gmanaging
communications, enriching
sequential user interactions
data streams, across
enabling faster varied
and moredigital media.
accurate responses directly at the network edge.
•Variational Autoencoders (VAEs): VAEs aid in optimizing bandwidth and reducing latency in edge AI applications by compressing high-dimensional data into compact, meaningful representations, essential for efficient transmission and processing in constrained edge environments
•Variational Autoencoders (VAEs): VAEs aid in optimizing bandwidth and reducing latency in edge AI applications by compressing high-dimensional data into compact, meaningful representations, essential for efficient transmission and processing in constrained edge environments

Generative AI in 6G Networks
Generative AI models will enable new research frontiers in 6G Networks

SEMANTIC COMMUNICATIONS DIGITAL TWINS

❑ GANs will enhance semantic communications by ❑ GANs can simulate various network scenarios
creating synthetic datasets that mimic complex, and conditions, creating accurate and dynamic
real-world interactions. digital twins of wireless networks.

❑ Transformer Models will boost semantic ❑ VAEs:. model and compress real-world network
communications through contextual data, enabling the creation of digital twins that
understanding, enhancing data transmission reflect network conditions.
relevance and accuracy.

AI-GENERATED CONTENT IN WIRELESS NETWORKS MOBILE EDGE COMPUTING AND EDGE AI

❑ Diffusion Models excel in creating highly ❑ Transformer Models improve real-time data
detailed, realistic visual content for dynamic and processing efficiently managing sequential data
interactive applications like VR and AR streams, enabling faster and accurate responses
at the network edge.
❑ Transformer Models facilitate the generation of
contextually relevant, coherent textual content ❑ VAEs reduce latency in edge AI applications by
enriching user interactions across varied digital compressing high-dimensional data, essential
media. for efficient transmission.

[1] "At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless Intelligence"
References and Documentation
 References

• [1] Smith, J., & Doe, J. (2021). At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless
Intelligence. arXiv. https://arxiv.org/abs/2106.12345
• [2] Zhu, E., Sun, H., & Ji, M. (2023). Physics-informed generalizable wireless channel modeling with segmentation and
deep learning: Fundamentals, methodologies, and challenges. arXiv. https://arxiv.org/abs/2401.01288v1
• [3] Shlezinger, N., Eldar, Y. C., & Rodrigues, M. R. D. (2022). MIMO Channel Estimation using Score-Based Generative
Models. arXiv. https://arxiv.org/pdf/2204.07122
• [4] Zaky, A. B., Huang, J. Z., Wu, K., & ElHalawany, B. M. (2020). Generative neural network based spectrum sharing
using linear sum assignment problems. China Communications, 17(2), 14-29. https://doi.org/10.23919/JCC.2020.02.002
• [5] GenerativeAIExamples/experimental/oran-chatbot-multimodal at main · NVIDIA/GenerativeAIExamples
(github.com)
Backup Slides
 Applications of Generative AI in Wireless Networks Option #2

Empowering wireless networks with generative AI

GANs VAEs Diffusion Models Transformer Models

Network Security Data Compression: VAEs learn to Noise and Distortion Removal: Channel Estimation: Use self-
encode data into a lower- These models learn to denoise and attention to process data, capturing
Spectrum Sensing: GANs are used dimensional latent space and then reconstruct cleaner signals, which is long-range dependencies in input
to generate synthetic data that decode it back to its original form. particularly beneficial in data effectively is used in massive
mimics different radio environmental This compressed representation can environments with low signal-to- MIMO processing.
conditions. This data helps in then be transmitted over the noise ratios
training predictive models to network, reducing the amount of Chatbots using Large Language
optimize frequency allocation and data that needs to be sent. Signal Processing at the Receiver: Models: Use of LLMs language
usage without field testing. Models operate by reconstructing capabilities to provide answers and
Joint Source Channel Coding: VAE the original information signals from recommendations based on Telco
Anomaly Detection: Generator can compact representations of data distorted and noisy received signals, data. Some examples are:
creates outlier data points that that are robust to channel noise, thereby enhancing the accuracy and
could indicate potential network thus optimizing the trade-off reliability of the communications • Automated Customer Support:
failures, while the discriminator between data compression and error system. Nois Can youabout network
Automate inquiries
learns to identify whether the data correction. issues, billing, and service plans.
points are plausible or anomalies. Waveform Shaping at the • Network Fault Detection: Identify
Signal Processing: VAEs can be Transmitter: Diffusion models can and diagnose network problems
Channel Modeling: GANs can trained on raw signal data to learn a generate complex waveforms that by analyzing error messages, logs,
generate synthetic channel compact representation in a latent are pre-optimized for expected and user complaints
conditions that mimic real-world space. This representation can then channel conditions and interference. • Network Configuration: LLMs can
environments, allowing for the be used to filter out noise. recommend network settings
robust testing and optimization of Energy-Efficient Routing: Diffusion configuration changes based on
communication systems under Network Resource Management: algorithms in WSNs can distribute performance logs
various simulated scenarios. VAE can significantly enhance data or tasks across the network in a • Standards Compliance: Help
network optimization by applying way that mimics this natural process network operators understand
their capabilities to solve complex of diffusion. and implement compliance
assignment problems. requirements.

[1] "At the Dawn of Generative AI Era: A Tutorial-cum-Survey on New Frontiers in 6G Wireless Intelligence"