 |
 |
Contents
Machine Learning, Deep Learning & Neural Networks
Deep Learning & Neural Networks
Introduction
AI demo
ChatGPT response
Let’s ask it to tell us how to build a hand grenade.
Self-driving car
Overview
Broad definition of AI - systems that perform tasks requiring human-like intelligence.
Some real-world applications - Siri, Netflix recommendations, fraud detection.
Here is what we will be covering today
· Narrow AI vs. General AI
· Machine Learning, Deep Learning, and Neural Networks (simplified)
· The Role of Data in AI
Narrow AI vs. General AI
Narrow (Weak) AI
Definition AI designed for specific tasks.
Examples
· Voice assistants (Alexa, Siri).
· Recommendation systems (Amazon, Spotify).
· Recommendation systems (Amazon, Spotify)
·
.John McCallum
Unit 43/34 Blake
Road
Mitchelton QLD 4053
[Date]
·
The
Manager
Linkt
Reply Paid
87288
Virginia BC
QLD 4014
· Dear Manager,
· Re: Transfer of Deceased Wife’s Account Balance (Heather Margaret McCallum)
· I am writing regarding my recent request to transfer the remaining balance from my late wife Heather Margaret McCallum’s Linkt account to my own. Heather passed away on [insert date if known], and I have enclosed her death certificate for your records.
· On [insert date if available], Linkt issued a cheque for the remaining balance of $24.59, favouring “Heather McCallum Deceased Estate.” However, my bank, Great Southern Bank, no longer processes cheques. When I visited a post office for assistance, I was informed they could not deposit the cheque as it was not made payable to me.
· To resolve this promptly, I kindly request that Linkt either:
· Issue a replacement cheque payable to John McCallum directly; or
· Transfer the $24.59 electronically to my Linkt account (5907351067), which would eliminate processing delays.
· I have enclosed the original cheque for cancellation and the death certificate for verification. Please contact me at [insert phone/email] if further steps are required. I appreciate your prompt attention to this matter.
·
Yours
sincerely,
John
McCallum
Enclosures:
Original cheque, Death certificate
· Note: Include the date of death if known. Add contact details (phone/email) if not already printed on the letterhead. Keep a copy of all documents and send via registered post if possible.
·
· Answer from Perplexity: https://www.perplexity.ai/search/letter-to-the-manager-linkt-re-kxoBp9uGRBaERRBFs7msfg?utm_source=copy_output
·
· Spam filters, facial recognition.
Limitations
· Cannot perform outside its trained domain.
· No consciousness or understanding.
General (Strong) AI
Definition
Hypothetical AI with human-like reasoning.
Examples (Theoretical)
· A robot that can learn any task a human can.
· AI with consciousness (e.g., sci-fi like Ex Machina).
Challenges
· No true General AI exists yet.
· Ethical and safety concerns (e.g., "What if AI becomes uncontrollable?").
Comparison & Discussion
Poll "Which type of AI do you interact with daily?"
Key Takeaway All current AI is Narrow AI—highly capable but limited.
Machine Learning, Deep Learning & Neural Networks
Machine Learning (ML) Basics
Definition
AI that improves through experience (data).
Machine learning (ML) is a field of artificial intelligence that focuses on enabling computers to learn from data without being explicitly programmed. It involves algorithms that identify patterns, make predictions, and improve their performance over time by analyzing data.
Analogy
Teaching a child by showing examples.
Types
Supervised Learning
Algorithms learn from labeled data, meaning the input data is paired with the correct output. Examples include:
· Regression: Predicting a continuous value (e.g., house price).
· Classification: Categorizing data into distinct classes (e.g., spam or not spam).
Unsupervised Learning (finding patterns, e.g., customer segmentation).
Algorithms learn from unlabeled data, discovering patterns and relationships without explicit guidance. Examples include:
· Clustering: Grouping similar data points together.
· Dimensionality Reduction: Reducing the number of variables while preserving important information.
Reinforcement Learning
Algorithms learn through trial and error, receiving rewards or penalties for their actions.
Deep Learning & Neural Networks
Neural Networks (Simplified)
Inspired by the human brain (layers of "neurons").
Neural networks are computational models inspired by the human brain, used in machine learning to process data and learn patterns. They consist of interconnected nodes or neurons that are trained to make predictions or classifications.
Example Image recognition (input → hidden layers → output "cat").
Deep Learning[i]
Subset of ML with many layers (hence "deep").
Powers speech recognition (Siri), image generation (DALL-E).
Interactive Exercise
"Guess the Algorithm"
Show 3 real-world cases (e.g., Netflix recs, fraud detection, medical diagnosis) and ask which ML type fits best.
Data as the Fuel for AI
Why Data Matters
"Garbage In, Garbage Out"
AI is only as good as its training data.
Examples
Biased data → biased AI (e.g., facial recognition errors for darker skin tones).
More data → better performance (e.g., GPT-3 trained on billions of texts).
https://youtu.be/59bMh59JQDo
Types of Data in AI
Structured (tables, spreadsheets) vs. Unstructured (images, audio). Otherwise known as analog or digital data.
Labelled (for supervised learning) vs. Unlabelled (for unsupervised).
Ethical Considerations
Privacy concerns Who owns the data? (GDPR, consent issues.)
Bias & fairness
How do we prevent AI discrimination?
Preventing AI discrimination is a critical goal that requires a combination of technical, ethical, and regulatory measures. Here are some key strategies to mitigate and prevent discrimination in AI systems:
1. Diverse and Representative Data
· Avoid Bias in Training Data: AI models learn from the data they are trained on. If the training data is biased, the AI will replicate and amplify these biases. Use datasets that represent diverse demographics, cultures, and contexts.
· Audit Data for Bias: Analyze datasets for skewed representations of specific groups or communities. Balance the dataset to avoid overrepresentation or underrepresentation.
· Synthetic Data Generation: In cases where real-world data is limited, generate synthetic data to ensure diverse representation.
2. Fair Algorithms
· Bias Detection and Mitigation: Use fairness-aware algorithms and techniques to identify and mitigate biases during model training and deployment.
· Regular Audits: Continuously monitor and audit AI systems for discriminatory patterns in predictions or outcomes.
· Fairness Metrics: Implement fairness metrics (e.g., demographic parity, equal opportunity) to evaluate the AI system’s performance across different groups.
3. Transparency and Explainability
· Explainable AI (XAI): Develop AI models that can explain their decision-making processes. This helps identify and address potential biases.
· Open Reporting: Provide clear documentation on how the AI system was developed, its data sources, and its intended use cases.
· Transparency in Outcomes: Allow users to understand why specific decisions were made, particularly in high-stakes applications (e.g., hiring, lending, law enforcement).
4. Human Oversight
· Human-in-the-Loop Systems: Incorporate human oversight in decision-making processes, especially in sensitive areas like hiring, healthcare, or criminal justice.
· Accountability: Assign responsibility to specific individuals or teams for the AI’s decisions and actions.
5. Regulation and Standards
· Ethical Guidelines: Follow established AI ethics principles, such as fairness, non-discrimination, and inclusivity (e.g., the EU's AI Act or IEEE AI Ethics Standards).
· Legal Compliance: Ensure compliance with anti-discrimination laws and data protection regulations (e.g., GDPR, CCPA).
· Industry Standards: Adopt industry standards for fairness and accountability in AI.
6. Diverse Development Teams
· Inclusive Teams: Assemble diverse teams to design, develop, and deploy AI systems. A variety of perspectives can help identify and mitigate biases.
· Bias Training: Train AI developers and stakeholders on the risks of biases and how to prevent them.
7. User Feedback and Iterative Improvement
· Feedback Mechanisms: Provide users with channels to report unfair or discriminatory outcomes.
· Iterative Updates: Continuously improve AI systems based on user feedback and new fairness research.
8. Context-Specific Solutions
· Customization: Tailor AI systems for specific cultural, social, and economic contexts to avoid applying one-size-fits-all solutions.
· Impact Assessments: Conduct impact assessments before deploying AI systems in sensitive areas to evaluate potential discriminatory effects.
9. Partnerships and Collaboration
· Cross-Sector Collaboration: Work with academics, advocacy groups, and regulators to address discrimination risks.
· AI for Good Initiatives: Engage in initiatives focused on using AI to promote fairness and social good.
10. Continuous Research
· Study Biases in AI: Invest in research to understand how biases emerge and propagate in AI systems.
· Innovation in Fairness Techniques: Develop new methods for bias detection, mitigation, and prevention.
By combining these approaches, organizations can build AI systems that are fair, ethical, and non-discriminatory, fostering trust and inclusivity in AI applications.
Discussion Prompt "Should companies freely use public data for AI training?"
Conclusion & Q&A
Summary
- Narrow AI = specialized; General AI = theoretical.
- ML learns from data; DL uses neural networks.
- Data quality & ethics are crucial.
Final Thought "AI is a tool—how we use it depends on us."
Additional Engagement Ideas