Beyond the
Last-Frame
Overview
Finding 03 / 05Methodology

A cheap test beforean expensive one

GPU time on a shared cluster is scarce, so I built a way to rank ideas before training any of them.

Training a temporal model is expensive, and the GPU grid at Idiap is shared — a wasteful experiment is a cost to the whole group. That constraint shaped the method: wherever possible, rank ideas before committing compute to them.

The Temporal Separability Benchmark reuses the same lightweight LDA probe from the phase work as a fast scoring instrument — so competing configurations can be compared in minutes, not in full training runs.

What the benchmark answered

Three questions, settled cheaply.

Is the signal there?

Yes — temporal structure exists

The probe confirmed measurable separability between phases, justifying a temporal model at all.
How fine should binning be?

Diminishing returns

Finer temporal binning stopped paying off quickly — a coarse, cheap binning was enough.
Which backbone?

Keep RETFound-Green

Efficient, robust, and continuous with the group's existing pipeline — no reason to switch.
An honest negative

The PCA study that was worth running anyway.

A separate controlled study asked whether compressing the frozen embeddings with PCA before the recurrent model would help. It did not — no improvement in macro-ROC-AUC or average precision.

Why a negative result still counts

A negative result is not a wasted one. It removed an attractive but non-beneficial branch from the design space, kept the implementation simpler, and is documented so the next researcher does not have to re-run it.