Key figures · one line each#

Visual recall10 figuresSay the mechanism

The diagrams that carry most of the exam vocabulary. For each one, the caption is what you should be able to say out loud in the exam — the picture is just the anchor.

Attention: queries, keys, values#

Attention pooling: keys and a query produce weights that mix values

Self-attention: \(q\) from one position is compared with all keys; the softmax weights mix the corresponding values. Self = Q, K, V from the same sequence; cross = Q from one sequence, K/V from another. Trainable projections, unlike the picture's fixed \(v_i\).

DINO: two views, one teacher#

Two augmented views pass through teacher and student; the student matches the teacher's distribution

Two augmented views of the same image; the teacher (EMA of the student) produces targets, the student predicts them. The mismatch shown across positions is the cross-view / patch-level matching problem. Slow EMA teacher → stable targets, no collapse.

CLIP: contrastive image–text#

CLIP trains image and text encoders so matching pairs have higher similarity

Maximise similarity of the \(N\) matching pairs, minimise the \(N^2-N\) mismatched pairs. Zero-shot classification = compare the image with text prompts; the softmax is over the candidate set, not a calibrated disease probability.

How an image enters a language model#

Vision encoder output goes through a projection into the LM embedding space

Matching the length of the vector is not enough: the projection \(W\) and alignment/instruction tuning are what make the LM able to read the visual tokens. Without training, the CLIP vector is out of distribution.

Workflow pattern: routing#

A router LLM sends each request to one specialist path

One path is selected. Contrast with parallelization, where all paths run and an aggregator combines them. This is the most common MCQ trap.

Workflow pattern: prompt chaining#

Sequential LLM calls with a gate between stages

Fixed stages, each consuming the previous output; a gate decides whether to continue. Extract → draft → translate is chaining, not routing.

Workflow pattern: parallelization#

Independent LLM calls merged by an aggregator

Subtasks are defined in advance and independent; results are aggregated. Use when the decomposition is known before execution.

Workflow pattern: orchestrator–workers#

An orchestrator assigns subtasks to workers, a synthesizer combines them

The orchestrator decides at runtime which subtasks are needed. Extra coordination cost: more calls/latency, error propagation, non-determinism, harder evaluation — justify it only when the paths depend on the case.

Workflow pattern: evaluator–optimizer#

Generator drafts, evaluator accepts or returns feedback

Generator drafts, evaluator checks against criteria and either accepts or returns feedback. Acceptance must be distinguishable from "revision limit reached"; validate the evaluator itself against human ratings.

The agent loop#

Human, LLM call, environment and stop condition interact in a loop

Reason → act → observe → repeat until a stop condition. The environment executes tools; observations come from the tool, not from the model's prediction. Use an agent only when the next step depends on observations.