Abstract
Natural time-varying images possess significant temporal correlations. It has been shown that the LGN improves efficiency of visual representation through temporal decorrelation of the retinal signal (Dong, Atick 1995). But under natural viewing conditions, the temporal correlations are changed significantly by the saccadic eye movements and hence the signal sent to the LGN has quite different correlation characteristics across and between saccades. Based on the measured statistical properties of visual input during free viewing of natural time-varying images, we predicted that the LGN changes its temporal filter dynamically according to saccade timing to maintain decorrelation (Truccolo, Dong 2000, 2001; Dong 2001). To verify the prediction quantitatively, we recently developed new experimental paradigms, in which alert subjects watch natural time-varying images and their eye movements are tracked. We have done two sets of experiments: extra- and quasi intra-cellular (S-potential) recordings of the LGN during free-viewing by awake cats which are allowed to move their eyes (both input and output spikes are recorded); fMRI recordings of the visual cortex during free-viewing and during fixed gaze viewing (of the same image sequences as free-viewing) by awake human subjects (the overall visual cortex activation by the LGN is recorded). The experimental results confirmed our theoretical predictions. Both right before/after and between saccades, the information transfer through LGN is optimized during natural viewing. Such optimization is achieved by dynamic temporal decorrelation through LGN: right before/after a saccade, temporal low-pass filtering; whereas between two saccades, temporal difference (band-pass) filtering. The predicted filter also agrees quantitatively with human psychophysical experiments (Diamond etc 2000). In conclusion, the theory of efficient coding gives a quantitative account for visual response differences across and between saccades.