We developed a few interesting memory metrics for the sherlock paper, which would be relatively easy to port into quail:
Precision: how precisely the recall events match the video events
the mean correlation between each recall event and the highest matching video event (can also be computed on a per event basis)
code: [np.max(1 - cdist(video_events, r, 'correlation'), 0).mean() for r in recall_events]
Distinctiveness: how distinctive the recall events are from other recall events
the mean of the upper triangle of the recall event correlation matrix (can also be computed on a per event basis)
-code: [1 - np.triu(np.corrcoef(r)).mean() for r in recall_events]
Recall autocorrelation: the rate at which the recall content changes over time
the autocorrelation of the recall matrix (not recall event matrix)
We developed a few interesting memory metrics for the sherlock paper, which would be relatively easy to port into quail:
Precision: how precisely the recall events match the video events
[np.max(1 - cdist(video_events, r, 'correlation'), 0).mean() for r in recall_events]
Distinctiveness: how distinctive the recall events are from other recall events
[1 - np.triu(np.corrcoef(r)).mean() for r in recall_events]
Recall autocorrelation: the rate at which the recall content changes over time