Summary: The human-centered view

This is the closing lesson of Track 16. T16 built classifiers, detectors, segmenters, generative models, 3D recovery, vision-language systems, and world models, many of them deployed in the real world. The final question this track owes: what do these systems get right and wrong in deployment, and how to reason about it as engineering rather than policy. Scope: the engineering side (failure modes, bias as training-data property, measurement, mitigation as design choices). Out of scope: policy debates about what to permit, regulate, or restrict, which are real and important but belong with the right stakeholders (legal, ethics, regulatory). The engineering view sharpens those debates with measurable inputs; it does not replace them.

Core ideas

• Failure-mode catalog. Distribution shift (degrade on data unlike training; addressed by broader curation, domain adaptation, domain randomization, monitoring). Adversarial examples (tiny crafted perturbations flip predictions; addressed by adversarial training, certified robustness, input validation). Out-of-distribution (OOD) inputs (model confidently wrong on unfamiliar input; addressed by calibrated confidence, OOD detection heads, ensemble disagreement). Shortcut learning (model latches onto spurious correlation: wolf-vs-husky-snow; addressed by dataset curation, augmentation, careful eval splits). Calibration / overconfidence (confidence scores misaligned with accuracy; addressed by temperature scaling, isotonic regression, deep ensembles).
• Bias is an engineering property of training data + architecture + evaluation. Mechanical: model fits the data’s statistical structure; web-scraped data is uneven; model inherits skews. Gender Shades audit (Buolamwini and Gebru 2018) found face-detection accuracy several times worse for darker-skinned women than lighter-skinned men, traced to training-set skew. Different data, different bias profiles (the L14 framing generalized).
• Measurement first: disaggregated (sub-group) accuracy reporting reveals what aggregate hides. Mitigation second: (1) data-side (balanced curation, targeted collection, datasheets for datasets), (2) model-side (adversarial debiasing, loss reweighting, fairness-aware auxiliaries), (3) evaluation-side (disaggregated reports, stress-test sets, audits).
• Trustworthiness gap. Benchmark accuracy ≠ real-world reliability. Distribution shift, calibration issues, sub-group disparities, tail events all live in the gap. Closed by: monitoring in production (per-group accuracy over time, OOD-input rate, calibrated-confidence distribution); calibration to enable deferral; human-in-the-loop where automation cannot be fully trusted; explicit rollback / graceful-degradation / escalation plans.
• Engineering scope vs policy scope. Engineering view: failure modes and bias are measurable design problems with engineering responses. Policy view: what is acceptable to deploy, regulate, or restrict, with policy stakeholders. Both are needed; this lesson covers only the engineering view explicitly.

What changes for you

When you read about an AI deployment claim (“95 percent accuracy in clinical trials,” “fair face-detection for all users,” “human-level driving”), the engineering questions become automatic: What was the distribution of the training data? What sub-groups were measured? How calibrated is the confidence? What is the monitoring story? What is the failure-mode plan? A vision system in production is not a single trained network; it is the network plus the data pipeline plus the evaluation suite plus the monitoring dashboard plus the failure plan. The trustworthiness gap is closed by engineering at all these layers, not by a better single model.

Track 16 has built the mechanics; this last lesson is the bridge to deployment-grade thinking. Vision used to be a hand-engineered, brittle craft. Sixteen lessons ago we started with the simplest data-driven move that worked; the world ended up here.

Closing the track

T16 = 16 lessons across 3 phases. Phase 1 (Foundations): linear classifiers, loss and optimization, neural networks and backpropagation, building the general-purpose image classifier. Phase 2 (How machines see): convolution, landmark CNN architectures, sequence tools, detection / segmentation / visualization, video. Phase 3 (Generating and grounding vision): self-supervised learning, GANs and VAEs, diffusion, 3D vision, vision and language, world modeling. This lesson closed with the human-centered engineering view.

Cross-track ties: T11 (Neural Network Intuition) is the gentler companion; T5 (AI Foundations) covers attention and transformers in depth; T14 (planned, Practical Transformers) covers the transformer mechanics T16 relied on; T18 (planned, Reinforcement Learning) covers model-based RL depth (Dreamer, MuZero); T19 (planned, Generative Modeling) covers the ELBO and score-based derivations underlying lessons 11 and 12; T24 (planned, Image Generation and Multimodal) covers production-scale image and video generation systems.