ADDITIONAL PROJECT TOPICS

1. Adaptive Generative AI for Edge Devices

• Problem Statement: Generative AI models require significant computational
resources, which is impractical for edge devices like IoT sensors or smartphones.
• Research Goal: Explore techniques for adapting large generative models (e.g.,
transformers) to edge devices. This could include:
Developing lightweight, energy-efficient generative architectures.
Investigating novel model compression techniques (e.g., quantization, pruning)
tailored for generative tasks.
Proposing strategies for split inference between edge and cloud.

2. Generative AI for Ethical AI Systems

• Problem Statement: AI systems often lack transparency and can generate biased or
unethical outputs.
• Research Goal: Investigate methods for integrating ethical principles directly into
generative AI models, including:
Novel loss functions or priors enforcing ethical constraints.
A generative framework that self-assesses and rectifies biases during training or
inference.
Applications for fairness-aware text or image generation.

3. Hybrid Generative Models for Multi-Modal Scientific Research

• Problem Statement: Many scientific problems involve multi-modal data (e.g., text,
images, graphs, and numerical data), and existing generative AI models struggle with
effective integration.
• Research Goal: Propose a hybrid generative model capable of:
Seamlessly combining diverse data types (e.g., combining textual research papers
with experimental image data).
Generating new hypotheses or visualizations for unexplored scientific
phenomena.
Enhancing cross-domain research in fields like materials science, climate
modelling, or neuroscience.
4. Generative AI for Non-Euclidean Data
• Problem Statement: Most generative AI models work on Euclidean spaces (e.g.,
images, text) but struggle with non-Euclidean data like graphs or manifolds.
• Research Goal: Research generative methods for:
Graph-based data (e.g., social networks, molecular structures).
Data represented on manifolds (e.g., geospatial data, brain imaging).
Applications in social network analysis, urban planning, or computational
biology.
Detailed Study:
1. Adaptive Generative AI for Edge Devices
1. Problem Statement
Generative AI models, such as transformers and VAEs (Variational Autoencoders), are
computationally intensive. They demand high memory, energy, and processing power,
making them unsuitable for resource-constrained edge devices like IoT sensors,
smartphones, or embedded systems. This project addresses how to adapt these models for
edge computing without compromising their performance significantly.

2. Research Goals
Primary Goal:
• Enable generative AI models to operate efficiently on edge devices.
Sub-Goals:
1. Develop lightweight generative architectures:
Build models optimized for low-power and low-memory environments.
2. Investigate model compression techniques:
Explore advanced methods such as quantization and pruning to reduce model size
while retaining performance.
3. Implement split inference strategies:
Propose mechanisms to divide the computation between edge devices and the
cloud to balance efficiency and performance.
3. Proposed Process
Step 1: Understanding Requirements and Constraints
• Edge Device Limitations:
Limited memory (e.g., 1GB or less).
Restricted processing power (e.g., low-power ARM processors).
Energy efficiency (e.g., battery-powered devices).
• Generative Model Characteristics:
Size of pre-trained models (e.g., GPT-3 has billions of parameters).
Computational complexity (e.g., attention mechanisms in transformers).
Step 2: Selection of Generative Models
• Choose state-of-the-art generative AI models to adapt for edge devices. Examples:
Transformers: Models like GPT, BERT, or Vision Transformers.
VAEs: Useful for image or audio synthesis.
GANs (Generative Adversarial Networks): For tasks like image generation.
Step 3: Designing Lightweight Generative Architectures
• Explore compact architectures tailored for edge environments:
Tiny Transformers: Reduce the number of attention heads and hidden layers.
MobileNet-inspired Models: Use depth-wise separable convolutions for lighter
computation in vision-based generative tasks.
Sparse Transformers: Use sparse attention mechanisms to reduce computational
overhead.
Step 4: Applying Model Compression Techniques
1. Quantization:
Reduce precision of weights and activations (e.g., from 32-bit floats to 8-bit
integers).
Use quantization-aware training (QAT) to minimize performance degradation.
2. Pruning:
Remove less significant weights or entire neurons from the model.
Focus on structured pruning to retain compatibility with hardware accelerators.
3. Knowledge Distillation:
Train a smaller model (student) using the outputs of a larger pre-trained model
(teacher).
Useful for transferring knowledge without retraining from scratch.
Step 5: Developing Split Inference Strategies
• Split inference divides computation between edge devices and the cloud:
Feature Extraction on Edge: Perform initial lightweight computations locally and
send intermediate representations to the cloud.
Latency Optimization: Minimize communication delays by selecting optimal
split points.
• Example Workflow:
▪ Edge device runs the first few layers of a generative model.
▪ Intermediate outputs are sent to the cloud for more resource-intensive
processing.
▪ Final results are returned to the edge device for display or further use.
Step 6: Implementation and Testing
• Implement the proposed solutions in real-world edge devices using popular
frameworks:
Frameworks for Lightweight AI: TensorFlow Lite, PyTorch Mobile, ONNX
Runtime.
Hardware: Raspberry Pi, Nvidia Jetson Nano, or microcontrollers with AI
accelerators.
• Evaluate the performance:
Metrics: Latency, memory usage, energy consumption, and accuracy.
Comparison: Evaluate trade-offs between model compression and generative
quality.

4. Technologies Used
AI and Machine Learning Frameworks:
1. TensorFlow Lite: For deploying optimized models on edge devices.
2. PyTorch Mobile: Provides support for deploying PyTorch models on mobile and edge
platforms.
3. ONNX Runtime: Useful for running pre-trained models with cross-platform support.
Model Compression Techniques:
• TensorRT: Nvidia’s platform for accelerating deep learning inference with quantization
and pruning.
• Distiller: A library for compression techniques like pruning and quantization.
• OpenVINO: Intel’s toolkit for optimizing AI models on edge devices.
Hardware Platforms:
• Edge Devices:
Raspberry Pi (ARM Cortex processors).
Nvidia Jetson Nano (AI-specific accelerators).
Smartphones (Qualcomm Snapdragon Neural Processing Units).
Generative Model Libraries:
• Hugging Face Transformers: Pre-trained models for text-based generation tasks.
• FastGAN or StyleGAN: Generative models for image synthesis.
• OpenAI Codex or GPT models: For text-based applications like content creation or
summarization.
Cloud Platforms for Split Inference:
• AWS IoT Greengrass: Allows edge devices to interact with cloud services.
• Google Cloud IoT Core: Supports real-time communication between devices and cloud.
• Microsoft Azure IoT Edge: Helps split workloads between cloud and edge
environments.
5. Expected Outcomes
• A lightweight, energy-efficient generative AI model suitable for edge devices.
• Demonstrated improvement in latency, memory usage, and energy consumption
compared to existing solutions.
• A framework for split inference that can adapt to various edge-cloud use cases.
• Practical applications in IoT, healthcare (e.g., diagnostics), and mobile devices (e.g.,
personalized AI models).
2. Generative AI for Ethical AI Systems:

1. Problem Statement
AI systems, including generative models, sometimes produce biased, unethical, or harmful
outputs. These issues arise because:
• Training data often contains biases reflecting societal inequalities.
• Generative models, such as transformers and GANs, lack mechanisms to enforce
fairness or ethical guidelines.
• AI systems operate as "black boxes," offering little transparency about their decision-
making processes.
Addressing these shortcomings is essential to build trustworthy AI systems for real-
world applications.

2. Research Goals
Primary Goal:
• Develop generative AI systems that incorporate ethical principles into their design to
ensure fairness, accountability, and transparency.
Sub-Goals:
1. Novel Loss Functions:
Design loss functions or priors that penalize unethical or biased outputs during
training.
2. Bias Assessment Mechanism:
Develop a framework enabling the generative model to self-assess and mitigate
biases during training or inference.
3. Fairness-aware Applications:
Create generative models for text or image tasks that prioritize diversity,
inclusivity, and fairness.

3. Proposed Process
Step 1: Problem Analysis
• Identify common ethical challenges in generative AI:
Gender, racial, and cultural biases in text/image outputs.
Propagation of stereotypes or harmful misinformation.
 Lack of explainability in generated content.
• Analyze existing frameworks and their shortcomings in handling ethical constraints.
Step 2: Dataset Analysis and Preprocessing
• Bias Identification:
Analyze training data for inherent biases (e.g., imbalanced representation of
groups or contexts).
Use statistical methods or explainable AI (XAI) tools to detect biased patterns.
• Data Augmentation:
Introduce diverse and inclusive examples into the dataset.
Balance underrepresented categories to minimize bias.
Step 3: Model Design
1. Incorporating Ethical Priors:
Add priors (constraints) into the generative model that promote ethical outputs.
For example:
▪ Diversity priors to ensure balanced representation.
▪ Content filters to prevent offensive or harmful outputs.
2. Custom Loss Functions:
Define loss functions that penalize biased outputs, e.g.:
▪ Fairness-aware loss: Penalizes disproportionate representation of any
group.
▪ Content moderation loss: Ensures outputs comply with ethical guidelines.
3. Self-assessment Mechanism:
Add modules for real-time bias detection in model outputs using fairness
metrics.
Enable feedback loops for the model to correct biases during inference.
Step 4: Training and Optimization
• Ethical Constraints during Training:
Implement adversarial training to challenge the model with biased data and
improve robustness.
Use reinforcement learning with ethical feedback to optimize the model.
• Explainability Integration:
 Incorporate explainable AI techniques to make the generative process
transparent.
Generate logs or explanations for each output, highlighting how ethical
constraints were enforced.
Step 5: Validation and Testing
• Metrics for Fairness and Bias:
Use quantitative metrics like demographic parity, equalized odds, or statistical
parity for bias evaluation.
Conduct qualitative tests with diverse user groups to assess inclusivity and
ethical compliance.
• Case Studies for Real-world Scenarios:
Test the model in fairness-sensitive tasks such as job description generation,
educational content creation, or AI-assisted art.

4. Technologies Used
AI Frameworks:
1. Transformers (Hugging Face):
Useful for fairness-aware text generation tasks.
2. GANs (Generative Adversarial Networks):
Applied for fairness-aware image generation.
3. Fairlearn or AIF360:
Libraries to measure and mitigate bias in AI systems.
Bias Mitigation Techniques:
• Reinforcement Learning from Human Feedback (RLHF):
Incorporate ethical feedback into the generative process.
• Adversarial Debiasing:
Use adversarial networks to reduce unwanted biases in generated outputs.
Explainability and Transparency Tools:
• SHAP (SHapley Additive exPlanations):
Explain individual predictions made by the model.
• LIME (Local Interpretable Model-Agnostic Explanations):
Provide interpretable explanations for text/image outputs.
Hardware and Software Platforms:
• Cloud-based Training:
Leverage AWS, Google Cloud, or Azure for large-scale training with ethical
constraints.
• Edge Deployment:
Implement lightweight versions of fairness-aware models for real-time
applications.

5. Applications
1. Fairness-aware Text Generation:
Generate inclusive content for education, entertainment, and policy documents.
Example: AI-generated job descriptions that avoid gender-specific terms.
2. Bias-free Image Generation:
Design AI systems for creating diverse and inclusive imagery.
Example: AI-generated advertisements or educational materials representing all
demographics fairly.
3. Content Moderation Tools:
Develop AI systems that identify and filter biased, harmful, or unethical content
automatically.

6. Expected Outcomes
• A generative AI system that produces ethical, fair, and transparent outputs.
• Reduced biases in AI-generated content, promoting inclusivity and trust.
• A scalable framework for integrating ethical principles into various generative AI
applications.
3. Hybrid Generative Models for Multi-Modal Scientific Research

1. Problem Statement
Scientific research often involves multi-modal data, which includes combinations of text
(e.g., research papers), images (e.g., microscopy scans), graphs (e.g., time series plots), and
numerical datasets (e.g., experimental measurements).
Challenges with Existing Models:
• Most generative models are specialized for a single modality (e.g., text or images),
making them inefficient at combining different types of data.
• Integration of multi-modal data requires handling distinct characteristics, such as
temporal dependencies in numerical data or semantic structures in text.
• Without proper multi-modal integration, scientific discoveries in areas like materials
science, climate modelling, and neuroscience are hindered.

2. Research Goals
Primary Goal:
• Develop a hybrid generative model capable of seamlessly integrating diverse data
types to advance scientific research.
Sub-Goals:
1. Multi-modal Data Fusion:
Design a model architecture that efficiently combines text, images, graphs, and
numerical data.
2. Generating New Hypotheses:
Enable the model to produce novel hypotheses or visualizations for unexplored
scientific phenomena.
3. Cross-Domain Research Applications:
Demonstrate how the model can facilitate interdisciplinary studies, e.g.,
connecting neuroscience with computational biology or climate modeling with
materials science.
3. Proposed Process
Step 1: Problem Analysis and Scope Definition
• Identify fields where multi-modal integration is critical, such as:
Materials Science: Combining textual research papers with microscopy images.
Climate Modelling: Integrating historical climate records (graphs) with satellite
imagery.
Neuroscience: Merging experimental data (numerical/graphs) with brain
imaging scans.
• Define metrics to evaluate the success of the hybrid model:
Accuracy in generating valid multi-modal outputs.
Utility in generating novel, scientifically relevant insights.
Step 2: Dataset Preparation
• Curate Multi-Modal Datasets:
Use existing repositories or create datasets combining different modalities.
Examples:
▪ Text and Images: Combine arXiv research papers with experimental
images.
▪ Graphs and Numerical Data: Use datasets with time-series and correlated
numerical outputs, e.g., climate records.
• Preprocessing for Uniform Representation:
Normalize numerical data and standardize image resolutions.
Use embeddings (e.g., Word2Vec for text, node embeddings for graphs) to
create a common representation space.
Step 3: Hybrid Model Design
1. Model Architecture:
Backbone Framework: Use a hybrid architecture combining transformers,
CNNs, and graph neural networks (GNNs):
▪ Text Encoding: Transformers (e.g., BERT) to process text.
▪ Image Processing: Convolutional Neural Networks (e.g., ResNet) for
image features.
▪ Graph Integration: GNNs for structured graph data.
▪ Numerical Data: Dense neural networks or recurrent layers for time-
series and numerical data.
 Multi-modal Fusion Layer:
▪ Design a layer to combine embeddings from text, images, graphs, and
numerical data into a unified latent space.
2. Generative Capability:
Use Variational Autoencoders (VAEs) or Generative Adversarial Networks
(GANs) to generate new multi-modal outputs:
▪ Example: Generate a research hypothesis (text) alongside a predicted
experimental image.
3. Attention Mechanisms:
Apply cross-attention mechanisms to allow the model to focus on relevant parts
of each modality during generation.
Step 4: Training and Optimization
• Loss Functions:
Multi-modal reconstruction loss: Measure accuracy of regenerating each
modality.
Cross-modal consistency loss: Ensure outputs across modalities align logically
(e.g., textual hypothesis matches numerical predictions).
• Training Strategy:
Train the model on pairs of modalities first (e.g., text and images) before
integrating all data types together.
Use transfer learning for pre-trained models (e.g., BERT, ResNet) to reduce
training time.
Step 5: Validation and Testing
• Test the model’s performance using multi-modal benchmarks:
Generate hypotheses or visualizations from mixed data inputs.
Compare with ground truth or assess novelty using domain expert evaluations.
• Conduct case studies to evaluate real-world applications:
Example: Use the model to predict properties of a new material based on textual
and imaging data.
4. Technologies Used
AI Frameworks and Libraries:
• Transformers: Hugging Face library for processing textual data.
• PyTorch or TensorFlow: For implementing custom hybrid architectures.
• DGL (Deep Graph Library): To process and integrate graph-based data.
• OpenCV: For image preprocessing and augmentation.
Multi-Modal Fusion:
• CLIP (Contrastive Language-Image Pretraining): A pre-trained model that aligns text
and image embeddings.
• ALIGN: A multi-modal framework for aligning text and images in a shared latent
space.
Generative Models:
• Variational Autoencoders (VAEs): For multi-modal data generation.
• GANs (Generative Adversarial Networks): For generating images or numerical
visualizations aligned with textual inputs.
Visualization and Analysis Tools:
• Matplotlib, Seaborn, Plotly: For visualizing generated numerical data or graphs.
• Tableau: For cross-modal analysis and insights.

5. Applications
1. Materials Science:
Generate hypotheses for new material properties by combining microscopy
images with textual literature.
Visualize atomic structures or predict synthesis methods for new compounds.
2. Climate Modelling:
Combine satellite imagery with historical climate data to generate future
predictions.
Generate visualizations of potential climate phenomena like hurricanes or ice
melt.
3. Neuroscience:
Create predictive models combining MRI scans with experimental datasets.
Generate hypotheses for brain region functions based on multi-modal datasets.
6. Expected Outcomes
• A robust hybrid generative model capable of handling diverse data types.
• Enhanced scientific research efficiency by automating hypothesis generation and
visualization.
• Breakthrough cross-domain applications, enabling interdisciplinary discoveries in
materials science, neuroscience, and climate modelling.
 4. Generative AI for Non-Euclidean Data :

1. Problem Statement
Generative AI models like GANs or VAEs are traditionally designed for data in Euclidean
spaces, such as images (pixel grids) or text (linear sequences). However, many real-world
data types are inherently non-Euclidean, including:
1. Graph-based Data:
Examples: Social networks, molecular structures, transportation systems.
Challenges: The irregular structure, node relationships, and edge dependencies
are not well-suited for conventional models.
2. Manifold-based Data:
Examples: Geospatial data, brain imaging, curved surfaces like protein
structures.
Challenges: Data is constrained to non-linear spaces (manifolds) that standard
generative approaches cannot model effectively.
2. Research Goals
Primary Goal:
• Develop generative models that can handle non-Euclidean data such as graphs and
manifolds, enabling applications in social network analysis, urban planning, and
computational biology.
Sub-Goals:
1. Graph Data:
Design methods to generate realistic graphs with properties like scalability,
topology preservation, and community structure.
Example: Generate new molecular structures for drug discovery.
2. Manifold Data:
Create models capable of generating and interpolating data constrained to
manifolds, such as Earth’s surface or brain cortical structures.
Example: Generate geospatial data for urban planning simulations.
3. Applications:
Social Network Analysis: Predict future social interactions or detect anomalies.
Computational Biology: Model protein folding or simulate molecular behavior.
Urban Planning: Generate traffic flow patterns or infrastructure layouts.
3. Proposed Process
Step 1: Problem Analysis
1. Identify key use cases:
For graphs: Molecular design, social network analysis, knowledge graphs.
For manifolds: Climate models, 3D medical imaging, geographical studies.
2. Define desired properties of the generated data:
Graphs: Structural validity (degree distribution, connectivity), scalability.
Manifolds: Smoothness, adherence to the manifold’s curvature.

Step 2: Dataset Preparation

1. Graph Datasets:
Use publicly available datasets like:
▪ Molecular Graphs: ZINC (molecular structures), QM9.
▪ Social Networks: Reddit or Twitter datasets.
▪ Knowledge Graphs: Freebase, DBpedia.
2. Manifold Datasets:
Geospatial Data: Use satellite imagery combined with geographical coordinates.
Brain Imaging Data: Use datasets like Human Connectome Project for cortical
surface representations.
3. Preprocessing:
Normalize data representations (e.g., node embeddings for graphs, coordinate
mapping for manifolds).
Apply dimensionality reduction (e.g., t-SNE, PCA) for initial manifold
visualizations.

Step 3: Model Design

For Graph-based Data:
1. Graph Generative Models:
Use models like GraphGAN or GraphVAE to learn and generate new graphs.
 Explore graph neural networks (GNNs) to capture relationships between nodes
and edges.
2. Structural Features:
Include loss functions that preserve graph metrics like clustering coefficient,
node centrality, and community structure.
For Manifold-based Data:
1. Manifold Learning:
Use techniques like Riemannian Neural Networks or Geometric Deep Learning
to map manifold data.
Adapt models to respect the curvature and constraints of the manifold.
2. Generative Frameworks:
Extend VAEs or GANs to operate in curved spaces:
▪ Riemannian VAEs: Model data directly on manifolds.
▪ ManifoldGANs: Generate synthetic manifold-constrained data.
Multi-modal Generative Models:
• If the data involves multiple modalities (e.g., graph + manifold), integrate embeddings
from both spaces using:
Cross-modal attention mechanisms for generating data consistent across both
modalities.

Step 4: Training and Optimization

1. Graph Data:
Use unsupervised learning to train models on existing graph datasets.
Employ graph-specific losses to evaluate the realism of generated graphs.
2. Manifold Data:
Train models with manifold-aware metrics like geodesic distances (shortest path
on curved surfaces) to preserve the structure.
Use data augmentation techniques, such as manifold interpolation, to expand
training datasets.
3. Evaluation Metrics:
For graphs: Degree distribution, modularity, and graph edit distance.
For manifolds: Smoothness, alignment with ground truth.
Step 5: Validation and Testing
• Graphs:
Test generated graphs on real-world tasks, such as predicting molecular
properties or simulating social interactions.
• Manifolds:
Compare generated data against ground truth manifold structures, such as the
Earth’s surface or brain cortical maps.

4. Technologies Used
AI Frameworks:
• PyTorch Geometric or DGL: For implementing graph-based generative models.
• Geometric Deep Learning Libraries: Libraries like PyTorch-Manifold for working
with curved spaces.
Generative Techniques:
• Graph-based Models:
GraphGAN: For generating graphs with realistic topology.
GraphVAE: Variational Autoencoders adapted for graphs.
• Manifold Generative Models:
Riemannian VAEs: To learn latent representations on curved surfaces.
ManifoldGANs: Extending GANs for manifold-constrained data generation.
Tools for Data Analysis and Visualization:
• NetworkX: For analysing graph metrics like centrality or clustering.
• 3D Visualization: Tools like Paraview or Matplotlib for visualizing manifolds and
curved surfaces.

5. Applications
1. Social Network Analysis:
Predict future links or simulate interactions in social networks.
Generate synthetic social graphs for testing privacy-preserving algorithms.
 2. Urban Planning:
Simulate geospatial layouts for urban infrastructure like roads, utilities, and
traffic.
Generate maps or layouts for new cities using geospatial data.
3. Computational Biology:
Generate new molecular structures for drug discovery or protein folding studies.
Simulate neural networks for brain connectivity research.

6. Expected Outcomes
• A scalable framework for generating non-Euclidean data, including realistic graphs
and manifold-constrained datasets.
• Enhanced understanding of complex systems, such as social networks, urban
geospatial layouts, or biological pathways.
• Applications in real-world domains, enabling advances in urban planning,
neuroscience, and molecular biology.