Diagnosing the Limitations of Artificial Working Memory, Shi YanZhe

Recurrent working memory models score perfectly on clean benchmark trials, and that number says nothing about how they fail. I took a pretrained LSTM-96 from the WorM benchmark (NeurIPS 2023), froze it, and went looking for the break.

The task

Three panels from one trial: a clean grey memory array with two green bars, a retention frame of full contrast random static, and a probe array matching the memory array. — The trial Memory array, delay, probe. Same or Changed? Here the delay is filled with static and the arrays are untouched.

Five change detection variants, 445 delayed trials per condition, a fixed seed, and every manipulation label preserving by construction.

1. Noise in the delay

Stage 1, accuracy by what fills the delay

Hover a bar for the Changed rate. Chance is 0.5.

Only the delay frames change. Every pattern drives accuracy to chance, and the model answers Changed 93 to 97% of the time. The memory is being overwritten while it is held.

2. Objects against background

A Changed trial with objects recoloured to yellow and cyan, a clean grey retention frame, and a probe array in the same recoloured palette. — 2a Objects recoloured. Still right 93.6% of the time.

A Same trial with a steel blue background and green bars, a clean grey retention frame, and an identical probe array. — 2b Background recoloured. Identical arrays, and the model answers Changed on almost every trial.

Stage 2, accuracy by change detection variant

Clean grey
Object recolour
Background recolour

Hover a column for its exact accuracy. Chance is 0.5 on these balanced trial sets.

Robustness here tracks how much of the image moved, not whether the change mattered to the task.

3. A dose response

Four retention frames side by side, from clean grey through increasingly visible coloured static to full contrast static. — 3 The same distractor at four contrasts.

Stage 3, accuracy against distractor contrast

Drag the contrast

contrast 0.25 accuracy 0.693 answers Changed 62% of the time

Only the six measured contrasts. Full contrast reproduces the Stage 1 random static result.

Low contrast leaves the model intact. By 0.5 it is at chance. Saliency breaks it, not the mere presence of a pattern.

What it means

Maintenance across the delay is fragile, and the fragility is graded.
The encoder leans on the background: the shortcut learning story again.
Report decision bias beside accuracy. The pair separates corrupted maintenance from corrupted encoding; accuracy alone cannot.

Limits

One model, zero shot, behavioural only. 445 trials per condition is enough for the large effects here and no smaller ones.

Credits

Supervised by Asst Prof Zhang Mengmi, College of Computing and Data Science, NTU, under the URECA programme. Model and stimuli from WorM: Sikarwar and Zhang, “Decoding the Enigma”, NeurIPS 2023 Datasets and Benchmarks.