MACHINE LEARNING (ML)

Policymakers have been struggling with the increasing complexity of Machine Learning (ML) models, which use deep neural networks and Large Language Models (LLMs) to churn through enormous volumes of data. Data fiduciaries have found it challenging to "correct, complete, update, and erase" sensitive data from computer systems due to its complexity. At the same time, we are seeing a rise in AI (Artificial Intelligence) bias, misinformation, and privacy violations, which gets heightened during events such as elections.

Machine learning:

Machine learning is a technique that enables computers to learn from data without explicit programming. It uses algorithms to analyze data and make predictions or decisions based on insights.

Types of Machine Learning:

• Supervised Learning:

• Involves training models using labeled data, where the output is known.

• Used to make predictions on new data. For example, identifying whether an image is of a cat or a dog.

• Unsupervised Learning:

• Works with unlabeled data, where the computer must identify patterns and structures on its own.

• Applications:

• Grouping similar data points together.

• Finding a lower-dimensional representation of data.

• Reinforcement Learning:

• Trains models to make decisions or take actions within an environment to achieve a specific goal.

• Often used in robotics for tasks like navigation and interaction with environments.

These categories cover the broad spectrum of how machine learning is applied to analyze data and enable computers to perform tasks autonomously.

Applications of Machine Learning

• Healthcare:

• Diagnostics: Machine learning algorithms can analyze medical images to accurately diagnose conditions like cancer.

• Finance:

• Data Analysis: AI technologies analyze large volumes of financial data to predict market trends and stock performance, aiding traders in making informed decisions.

• Speech Recognition:

• Utilizes natural language processing to transform human speech into written text, facilitating voice searches and enhancing accessibility features.

• Examples: Siri, Alexa, Google Assistant.

• Customer Service:

• Chatbots are revolutionizing customer engagement by providing instant responses to frequently asked questions (FAQs) and offering personalized advice.

• Examples: Slush, Maya Chatbots, Zendesk Chat.

• Computer Vision:

• Machine learning enables computers to derive meaningful information from visual inputs, with applications spanning from medical imaging to photo analysis.

• Examples: Google Translate's visual translation, Facebook 3D Photo, Faceapp.

• Recommendation Engines:

• Analyzes past consumer behavior to enhance cross-selling strategies, often utilized by online retailers to suggest products or services.

• Examples: Spotify for song recommendations, Netflix for movie suggestions, Amazon product recommendations.

• Natural Language Processing (NLP):

• Language Understanding: NLP algorithms process and generate human language, enabling applications like chatbots and virtual assistants to understand and respond to human speech and text.

• Text Analysis: NLP is used to analyze large volumes of text data, such as customer reviews and social media posts, to derive insights and improve decision-making.

• Fraud Detection:

• Utilizes machine learning to identify suspicious transactions through techniques like supervised learning and anomaly detection.

• Examples: Credit card fraud detection, Point-of-Sale (POS) fraud detection, PayPal's fraud prevention systems.

The antithesis of ML:

Concept Origin:

• Initiators: Cao and Yang in their work "Towards Making Systems Forget with Machine Unlearning."

• Purpose: To make AI systems "forget" specific data from trained models, counteracting the traditional aims of Machine Learning (ML).

• Objective: Add an algorithm to AI models to identify and delete false, incorrect, discriminatory, outdated, or sensitive information.

• Challenge: Due to the constant updating and usage of data by AI models, tracking and managing data quality becomes complex, creating issues with data lineage and making it hard to remove sensitive information.

Issues with Current Methods:

• Data Manipulation: Difficulty in managing and removing data due to complex algorithmic interactions, which can lead to manipulation and adversarial outputs.

• Data Poisoning: Risk of hackers inserting manipulated data into models, resulting in biased outcomes.

Alternative Solutions:

• Data Pruning: Involves deleting the entire dataset and retraining the model, which is computationally expensive, time-consuming, and may result in reduced accuracy.

Advantages of MUL:

• Cost Efficiency: Offers reduced costs and time efficiency compared to data pruning.

• Accuracy: Aims to maintain or enhance model accuracy while improving the ability to unlearn specific data.

Current Adoption:

• IBM: Testing MUL models for improved accuracy, intelligibility, and cost efficiency.

• Innovation: Encourages companies to develop and refine MUL solutions tailored to their needs.

Approaches to Implementing Machine Unlearning (MUL)

1) Private Approach: Data fiduciaries (companies) are responsible for testing and implementing MUL algorithms within their own training models.

Advantages:

• Flexibility: Allows companies to enhance their AI models and protect user rights without government intervention.

• Innovation: Encourages companies to develop and refine MUL solutions tailored to their needs.

Challenges:

• Expertise and Affordability: Smaller companies might struggle with the technical expertise and financial resources required for MUL implementation.

Current Status: This model is in a preliminary stage, with limited adoption.

2) Public Approach: Government prepares statutory frameworks to mandate MUL implementation through legal mechanisms.

Variations:

• Soft-Law Approach: Government issues guidelines or recommendations (e.g., EU’s AI Act) to encourage MUL adoption.

• Hard-Law Approach: Enacts specific laws requiring data fiduciaries to implement MUL models.

Advantages:

• Regulatory Pressure: Provides a structured approach to ensure compliance and uniformity.

• Contextual Relevance: Aligns with rising legislative focus on AI, with potential for future regulatory requirements.

Challenges:

• Implementation: Government mandates may be slow to adapt and may not address all industry-specific needs.

3) International Approach: Countries collaborate to develop and adopt a global framework for MUL, aiming for uniform standards across borders.

Advantages:

• Global Standards: Promotes consistency in AI governance and MUL practices across jurisdictions.

• Collaboration: Encourages international cooperation in setting standards.

Challenges:

• Geopolitical Frictions: Variations in national interests and regulatory environments can complicate uniform adoption.

• Standard-Setting Bodies: Relies on international organizations (e.g., International Electrotechnical Commission) to develop and enforce standards.

These approaches represent a formal blueprint for one of the solutions that can be utilised to subdue the menace of Generative AI and preserve the user’s right to be forgotten. The MUL is still in the preliminary stages. Therefore, stakeholders must address technical and regulatory considerations to ensure its effective implementation in this evolving landscape of AI.