Fig. 3

Object position axes across background variations are aligned with each other. Since object position decoding is tolerant to background variations, we tested whether the linear decoding axes for each background configuration were aligned by visualizing the decoders in the first two principal components of the neural response space. These dimensions were computed for the full set of neural responses obtained in each session. (a) As with Fig. 2a, each panel represents a unique configuration of the background rotation and depth with rows representing variations in rotation and columns representing variations in depth. Each dim point represents a single image presentation, and bright points represent trial-averaged responses. Gray-to-yellow gradient represents monotonic variation in object position. A gradient line was fit to the responses for each background condition, shown here in two dimensions for illustration. The lines shown have been normalized to have the same length in the projected space shown. The text label at the top left represents the relative angle between each decoder and the central decoder (the middle background condition plot, marked with ∗) calculated in the full dimensional space of responses used for decoding. (b) Distribution of angles in A as a histogram. Arrow at the top represents the median angle for this session (7.78°). (c) Distribution of relative decoder angles across all object position decoders (like those in A) across sessions for both monkeys (blue and red distributions). Blue and red arrows represent the median of angles across sessions (16.4° for monkey 1, 12.07° for monkey 2). Dark gray distribution represents the angles of object position decoders after shuffling the position values for each trial (median 88.13°, shown as dark gray arrow). Light gray distribution represents the angles between randomly chosen vectors of the same dimensionality as the neural population space (median 89.99°, shown as light gray arrow).