Machine Learning Assignment 8: Ethical AI, Causal Inference &
TinyML
1. Causal Inference and Directed Acyclic Graphs (DAGs)
Objective: Move beyond correlation to understand cause-and-effect.
Structural Causal Models (SCM): Define the difference between
$P(Y | X)$ (observational) and $P(Y | do(X))$ (interventional).
The Backdoor Criterion: Identify which variables must be
controlled for to block all spurious paths between a treatment
and an outcome.
Implementation: Use the DoWhy library to estimate the causal
effect of a feature on a target variable in a dataset with known
confounders.
2. Algorithmic Fairness and Bias Mitigation
Objective: Design systems that do not perpetuate societal prejudice.
Fairness Metrics: Define and calculate "Equalized Odds,"
"Demographic Parity," and "Counterfactual Fairness" for a loan
approval model.
Mitigation Strategies: Implement a Pre-processing (Re-
weighing), In-processing (Adversarial Debiasing), and Post-
processing (Reject Option Classification) technique.
Case Study: Analyze the "Gender Shades" study and explain
why intersectional analysis is critical for facial recognition
systems.
3. TinyML and On-Device Intelligence
Objective: Deploy ML models on ultra-low-power microcontrollers.
Pruning and Sparsity: Explain how magnitude-based pruning
removes "unimportant" neurons to reduce model size without
significant accuracy loss.
Knowledge Distillation: Train a "Teacher" model (Complex CNN)
and use it to guide the training of a "Student" model (TinyNet).
Hardware Constraints: Profile a model’s RAM and Flash usage
for a target device like an Arduino Nano 33 BLE or ESP32.
4. Robustness against Adversarial Attacks
Objective: Secure models against malicious manipulation.
Fast Gradient Sign Method (FGSM): Generate adversarial
perturbations to fool a pre-trained ImageNet model.
Defensive Distillation: Explain how this technique can increase
the "smoothness" of the model's decision surface to resist
attacks.
Certified Robustness: Discuss the concept of "Randomized
Smoothing" as a way to provide a mathematical guarantee on
model stability.
5. Federated Learning with Differential Privacy
Objective: Secure collaborative learning across edge devices.
Secure Aggregation: Describe the protocol that allows a
central server to sum local model updates without ever seeing
individual device data.
The Privacy Budget ($\epsilon$): Track the cumulative privacy
loss over multiple communication rounds of Federated
Averaging (FedAvg).
Implementation: Use PySyft or Flower to simulate a federated
training environment across five virtual clients.
6. Symbolic AI and Neuro-Symbolic Integration
Objective: Combine deep learning with logic-based reasoning.
Logic Tensor Networks (LTN): Explain how First-Order Logic
formulas can be grounded into neural network operations.
Program Synthesis: Discuss how models like DeepCoder use ML
to predict which logical primitives are needed to solve a
programming task.
Benefits: Why does adding a symbolic layer improve the
interpretability and "out-of-distribution" generalization of a
neural network?
7. Explainable AI for Computer Vision (XAI-CV)
Objective: Visualize the "Look" of a model's decision.
Grad-CAM: Implement Gradient-weighted Class Activation
Mapping to produce heatmaps showing which pixels
contributed most to a specific classification.
Influence Functions: Use influence functions to find the "Most
Influential" training images that led to a specific test-time
prediction.
Human-Centered Evaluation: Conduct a small study to see if
these explanations actually help a non-expert trust the model's
decisions.
Submission Requirements
Technical Report: 10 pages total. Must include DAG diagrams,
fairness metric tables, and Grad-CAM heatmaps.
Code Repository: Python scripts using DoWhy, AIF360,
TensorFlow Lite, and Art-Toolbox.
Reflection: A 1-page essay on the "Right to Explanation" in
automated decision-making as outlined in the GDPR.