Fig 1.
Recurrent neural network models.
Left. The Bump attractor Model consists of excitatory connections (green) between units that receive similar inputs and inhibitory connections (red) with units that receive different inputs (inputs are not shown in the figure, but adjacent units receive similar inputs). Connections are only shown for a unit at the top, but the other units have similar connectivity. Right. The Mixed model has a fraction of units connected with the bump attractor connectivity, while the rest are randomly connected (note that these 2 sub-networks are mutually connected).
Fig 2.
Task Description and electrode locations
(A) Task description. Both tasks are different versions of a visually-guided delayed saccade task. Trials were initiated by fixating for 500 ms on a central fixation spot, after which a red square (or target) was shown for 300 ms in 1 out of 8 locations (Monkeys P and J) or 1 out of 4 locations (Monkey W). After a 1000 ms Delay 1 period, a second stimulus was presented for 300 ms in a different location from the target. For Monkeys P and J, this second stimulus was always a green square, while for Monkey W, the second stimulus had a 50% chance of being a green square and a 50% chance of being a red square. The task rule for both monkeys was to report the location of the last red square seen. Thus, the green square, if shown, served as a distractor. After the second stimulus, a 1000 ms Delay 2 period was followed by a go-cue, which was the disappearance of the fixation spot. The monkeys had to saccade to the remembered location within 500 ms to receive a juice reward (B) Location of implanted electrode arrays. In the three monkeys, we chronically implanted electrode arrays in the pre-arcuate region, which includes the FEF (blue), and along the dorsal and ventral banks of the principal sulcus, denoted as DLPFC (red). For all arrays, electrodes along the sulcus were longer (5 – 5.5 mm), while further from the sulcus they were shorter (1 – 1.5 mm).
Fig 3.
Noise correlation analysis in PFC regions.
(A) The proportion of neurons with selective activity increases for FEF (left) and DLPFC (right). (B) The correlation coefficient for the neuron pairs (blue bars: significant values; gray bars: non-significant values). (C) Fano factor for the neuron pairs. The red squares highlight the properties of this data that were used to select the RNN models in subsequent analyses.
Fig 4.
Matching model properties to FEF (left) and DLPFC (right) constraints.
(A) Percentage of selective neurons for models trained using different percentages of bump attractor units (x-axis) for different noise levels (left to right: 0.07, 0.08, 0.09). (B) Same as (A) but for the percentage of correlated pairs. (C) Same as (A) but for the percentage of neurons in correlated pairs. (D) (left) Fano factor for models that consistently match the properties in A-C: yellow window highlights models that match the specific property, and orange window highlights models that match all properties in A-C (for noise levels without a match for all properties, we selected the bump percentages that matched at least 2 of them). (Right) The correlation coefficient for the neuron pairs (blue bars: significant values; gray bars: non-significant values).
Fig 5.
(A) Cross-temporal decoding of FEF, its matching model (20% bump), DLPFC, and its matching model (7.5% bump) for neurons/units that participate in at least one significantly correlated pair (top plots) and those that do not (bottom plots). (B) Decoding accuracy of the diagonal (0 to 2.6 s) of non-significantly-correlated units divided by significantly-correlated units. (C) Code stability in the first 500 ms of Delay 1 (0.3 to 0.8 s) for significant neurons/units. Code stability is quantified as the mean time-specific decoder performance (0.3 to 0.8 s) divided by the mean performance of decoders trained during the 0.3 to 0.8 second period and tested between 0.8 and 2.6 s.
Fig 6.
An illustration of the method used to determine the significant locations in individual neurons/units.
The same method was applied to neurons in the brain and units in the models.
Fig 7.
(A) The connectivity matrix of the bump attractor connectivity showing the input weights structure (left) and bump connectivity (right) where adjacent units excite each other (yellow color) and further away units inhibit each other (blue color). (B) Connectivity matrix of a mixed network with 20% bump attractor initialization (after training). The red asterisk shows the initialized units with the bump attractor connectivity. (C) Same as B but for the 7.5% bump attractor network.