Frontier Index · Place Techniques in the Map¶
Scope and evidence
Representative families checked on 2026-10-18, not a popularity ranking or exhaustive research inventory. Use courses, author code and documentation for scope; prioritize objectives, interfaces and limits over model names.
Architectures, representations and generation¶
| Family | Role | Entry | First-pass scope |
|---|---|---|---|
| Transformer, ViT, DiT | Similar components across language, vision and diffusion; architecture does not fix the objective | Architectures / generation | Attention and data flow |
| Efficient attention and long context | Structure, approximations, positions and implementations affect cost/context | Transformer / systems | Linear attention vs. FlashAttention |
| MoE | Partial expert activation with routing/balance/communication trade-offs | Architectures | Total vs. active parameters |
| SSM, Mamba and hybrids | State updates and selection for sequences | Mamba | Compare recurrence and attention conceptually |
| Self-supervision, contrast, CLIP, DINO, MAE | Different objectives from views, teachers and masks | Representation | Frozen vs. fine-tuned features |
| JEPA and predictive representations | Predict features rather than pixels | Representation / world models | Read V-JEPA 2 author code, not retraining |
| VAE, GAN and normalizing flows | Latent-variable, adversarial and invertible-transform approaches | Generation | Objectives and sampling |
| Diffusion, flow matching, rectified flow | Generation along noise/probability paths with different targets | Diffusion / flow matching | Same low-dimensional labs |
| Consistency/distillation and few-step sampling | Extra training/objective design to reduce sampling steps | Generation / systems | Purpose; acceleration has trade-offs |
| Discrete diffusion/non-autoregressive generation | Discrete-state generation beyond token-wise autoregression | Generation | Optional MIT lecture 5 |
Perception, reasoning, decisions and systems¶
| Family | Role | Entry | First-pass scope |
|---|---|---|---|
| Prompted segmentation, VLM and video | Conditional/multimodal processing beyond fixed-label tasks | Vision / multimodality | Tasks, metrics and spatial limits |
| NeRF, 3DGS, implicit geometry and points | Different shape/appearance representations | 3D vision | Appearance vs. geometry |
| Graph Transformers and equivariance | Explicit relations and symmetry | Graphs | Small GCN and edge-free baseline |
| SFT, LoRA/PEFT and preferences | Task and behavior adaptation | Training | One task with held-out checks |
| RLVR, reasoning budgets and test-time compute | Verifiable feedback or added computation for specific tasks | Training / decisions | Rewards, verification and costs; GRPO details later |
| RAG, tools, memory, agents and multi-agent systems | Compose models with material/interfaces | Applications | One traceable tool workflow |
| Imitation, offline/model-based RL and VLA | Learn actions from demonstrations/interaction | Decisions / robot policies | ACT and closed-loop evaluation |
| World models and latent planning | Predict environment changes and evaluate decision benefit | World models | Prediction vs. task success |
| Time-series foundation models, PINNs and operators | Temporal transfer, physics constraints and function maps | Scientific AI | Baselines and controlled experiments |
| Quantization, distillation, sparsity and runtimes | Algorithmic vs. hardware efficiency | Systems | Quality, memory and latency on one device |
| Continual/active/federated learning | Changing tasks/data, annotation budgets or distributed data constraints | Paradigms | Understand problems before adding training stacks |
| Calibration, OOD, explanations and contamination | Confidence, shifts, evidence and failure | Evaluation | Checks shared by all projects |
A few coverage references¶
- CS231n 2026: vision, representations, generation, 3D and vision-language in one curriculum.
- CS336 2026: data, training systems, alignment and reasoning RL; substantial Python/deep-learning/systems prerequisites make it a later specialization.
- MIT 6.S184 2026: continuous/discrete generation, sampling, latent spaces and guided labs.
These inform scope and classification, not new required courses. Place new models by task, architecture, objective, data, cost and evidence; add pages only for distinct learning problems.