Teach machines
to see carefully.

OUTCOME

Frame a vision task, learn and reuse visual representations, measure clean and shifted performance, and turn model confidence into a bounded enterprise decision.

Follow one five-class quality-inspection scenario from image contracts and convolution through scratch CNN training, frozen ResNet/ConvNeXt embeddings, partial fine-tuning, failure analysis, abstention, and monitoring.

OUTCOME

Explain modern residual and efficient ConvNet design, measure representation quality and real execution separately, and choose a backbone against a written deployment contract.

Follow one source-aware inspection benchmark across ResNet-18, ResNet-50, MobileNetV3-Large, EfficientNet-B0, and ConvNeXt-Tiny. Compare frozen probes, resolution, partial fine-tuning, representation drift, robustness, and Pareto efficiency.

OUTCOME

Explain how patches, positions, and attention form visual representations; compare flat and hierarchical transformers with CNNs; and measure the quality and systems consequences of tokenization.

Follow one source-aware inspection scenario from patch projection and hand-computed attention through a minimal ViT, four official pretrained backbones, resolution interpolation, attention distance, and a provisional enterprise decision.

OUTCOME

Explain how visual representations can be learned without manual pretraining labels, implement the core objectives, diagnose collapse and shortcuts, and evaluate whether reusable features reduce downstream label demand.

Follow one industrial archive from augmentation contracts and manual NT-Xent through a tiny SimCLR encoder, EMA teacher and masked-patch mechanics, random/supervised/SSL feature comparison, probing, retrieval, source separation, and a governed representation decision.

OUTCOME

Turn images into localized, confidence-ranked decisions; explain assignment and post-processing; compare dense and set prediction; and choose a detector against source-aware quality, systems, and governance evidence.

Follow one multi-object factory inspection scenario from coordinate contracts, IoU, matching, and AP through anchors, feature pyramids, a trained anchor-free detector, NMS, Hungarian assignment, YOLO/DETR design philosophies, and an open-vocabulary extension.

OUTCOME

Turn exact pixels into a governed prediction contract; compare semantic, instance, panoptic, and prompted masks; and choose an operating model using source, boundary, prompt, and review evidence.

Move from label-mask integrity, U-Net skips, loss priorities, and boundary metrics through Factory C failure slices, point and box prompts, detector-box perturbation, human correction, and a current but isolated SAM 3.1 adapter.

OUTCOME

Define what similarity means for a product decision, learn and evaluate a normalized embedding space, and operate exact or approximate retrieval without confusing speed, scale, and semantic quality.

Follow one multi-factory archive through frozen ResNet baselines, triplet learning, P×K batches, online mining, simulated oracle adjudication, group-isolated duplicate evaluation, exact-search parity, HNSW tuning, metadata filters, drift, and blue/green embedding migration.

OUTCOME

Maintain measurable identity through time, attach landmarks and geometry to each track, and diagnose whether a temporal-state failure began in detection, association, pose, or smoothing.

Follow one procedural workcell from timestamped video and detector observations through IoU/Hungarian matching, lifecycle, motion, appearance, Byte-style recovery, identity metrics, heatmaps, PCK, pose geometry, Camera C shift, and versioned evidence.

OUTCOME

Reuse visual representations and promptable interfaces across tasks while preserving capability-specific evaluation, provenance, and bounded operating contracts.

Synthesize the Beginner track through normalized image–text alignment, prompt and vocabulary sensitivity, global versus patch features, open-vocabulary grounding, detector→segmenter error propagation, the adaptation ladder, and current foundation-model governance.

OUTCOME

Trace visual features into a language generator and evaluate whether its answers remain tied to the supplied evidence.

Move from CLIP-style alignment through visual tokens, projection, resampling, cross-attention, autoregressive generation, instruction tuning, resolution budgets, grounding, hallucination diagnostics, capability-specific evaluation, and enterprise inference contracts.

OUTCOME

Turn a bounded visual question into a conclusion whose required facts, evidence, tool calls, checks, and uncertainty can be independently reviewed.

Move from recognition to task DAGs, observed and derived facts, capability-specific evidence, exact tools, four-valued states, contradictions, relevant and irrelevant counterfactuals, multi-image attribution, and governed decisions—without treating hidden reasoning traces as an audit artifact.

OUTCOME

Convert mixed, multi-page documents into structured fields whose page, region, span or cell, and transformation provenance can be independently replayed.

Route digital and scanned inputs; preserve coordinate frames; separate OCR, layout, reading order, tables, forms, and normalization; then evaluate source shift, perturbations, unsupported values, and human-review policy.

OUTCOME

Retrieve the minimum sufficient authorized multimodal evidence and produce claims whose granular citations can be independently checked.

Preserve document, page, cell, figure, image, region, canonical, version, and access lineage; compare lexical, semantic, visual, structured, and multi-vector retrieval; then separate retrieval, assembly, generation, citation, and freshness failures.

OUTCOME

Turn timestamped streams into event intervals and multi-event answers whose temporal evidence, identities, permissions, and citations can be independently checked.

Preserve presentation time through sampling and representation; separate candidate retrieval from boundary grounding; use deterministic interval tools; then test complete evidence, counterfactuals, long-video hierarchy, streaming delay, frame drops, and audio–visual conflict.

OUTCOME

Turn verified image, document, and video evidence into an observable, permissioned, bounded tool loop that stops or escalates safely.

Define typed state and tool contracts; separate planning from execution; enforce capability and resource policy before invocation; validate inputs and outputs; promote only source-bound evidence; and evaluate selection, authorization, recovery, and termination independently.

OUTCOME

Recover frame-aware 3D structure from calibrated observations, quantify geometric uncertainty, and turn reconstruction into auditable metric spatial evidence.

Move from coordinate frames, pinhole projection, back-projection, distortion, and calibration through robust epipolar matching, stereo depth, triangulation, pose, structure from motion, point clouds, representation choice, occlusion, and source-held-out spatial decisions.

OUTCOME

Build and evaluate view-synthesis systems without confusing a convincing image with correct geometry or deployable scene evidence.

Move from calibrated rays, bounds, positional encoding, transmittance, and hierarchical sampling through camera-held-out NeRF evaluation, appearance–geometry disagreement, explicit Gaussian projection and splatting, density-control lineage, compression, editability, dynamics, and semantic fields.

OUTCOME

Represent persistent world state, predict action-conditioned futures, and evaluate whether model-based plans remain supported and physically consistent.

Move from observations, memory, structured actions, and explicit transitions through 4D representations, recursive and stochastic futures, object permanence, counterfactual tests, planning exploitation, source shift, and governed evidence.

OUTCOME

Convert grounded multimodal observations and goals into embodiment-compatible action proposals, then validate, execute in bounded simulation, verify, and recover without granting a model physical authority.

Move from proprioception, frames, referent ambiguity, affordances, and action spaces through behavioral cloning, tokenized/continuous/diffusion VLA families, action chunks, independent feasibility checks, freshness, postconditions, and held-out embodiment evidence.

OUTCOME

Turn partial, timestamped observations into persistent and queryable world knowledge without confusing memory with current truth.

Move from pose, drift, occupancy evidence, unknown space, object association, permanence, and moved-object history through relation provenance, typed queries, semantic candidate verification, hierarchical navigation, source-held-out evaluation, and dependency-selective plan invalidation.

OUTCOME

Acquire useful capability on a new domain while making legacy interference, alignment drift, forgetting, lineage, and rollback evidence explicit.

Diagnose shift before training; compare frozen reuse, projectors, adapters, prompts, LoRA, partial, and full fine-tuning; then move through replay, regularization, distillation, parameter isolation, capability regression, and trusted release control.

OUTCOME

Design a vision system that measures degradation, recognizes incomplete evidence, abstains under a frozen risk policy, and verifies every attempted recovery.

Keep robustness, softmax confidence, calibration, uncertainty, OOD evidence, error detection, and conformal coverage distinct; then connect them through fail-closed policy and bounded recovery.

OUTCOME

Optimize a spatial or multimodal system without silently deleting evidence, reliability, isolation, or rollback capability.

Profile the complete path; compare input, token, precision, structure, distillation, export, batching, caching, and temporal-reuse choices; then freeze a capability-aware gate before shifted evaluation.

OUTCOME

Operate a complete spatial or multimodal AI configuration while preserving calibration, coordinate consistency, lineage, capability evidence, rollback, and safe recovery.

Move from immutable manifests and typed registries through structural, temporal, and operational readiness; time-valid frames; persisted calibration state; capability dependencies; drift and delayed outcomes; progressive delivery; fleet convergence; stateful rollback; incident recovery; and audit.