Fig 4. DPX task and agent design.

A. Design of the DPX agent. The agent was comprised of perception and memory ring attractors that represented the current stimuli and the context, respectively (due to their underlying EI balances). Both ring attractors were structurally identical to the one specified in Fig 2B. The perceptual network’s pyramidal outputs projected weakly to the memory ring, providing its representational inputs. As with the ring attractor in Fig 2B, certain areas of the ring represented the cue and probe stimuli. How well the cue and probes were represented in the networks was used to inform the agent’s decision making process. Pyramidal cell activity in the ring attractors projected to leaky left and right response accumulators that utilized a softmax decision making algorithm to determine the agent’s likely response to the probe. Connection weights to the left accumulator (i.e., A from the memory ring and X from the perception ring) are weaker than those to the right accumulator (i.e., B from the memory and Y from the perception ring), because the representations of B and Y needed to override the A and X representations during decision making. B. Temporal structure of an AY DPX trial and how the agent’s decision making process changed in response to task events. The gray regions indicate when the cue and probe were presented (not to temporal scale). As the agent represented information about the cue and probe, its likelihood of engaging in a left or right response changed. For example, representation of the A cue caused the agent to be much more likely to engage in a left action, but as the Y probe information was represented, the agent became unlikely to engage in a left action. A drift-diffusion-like decision process was used to determine the timing and choice of the agent’s actions. As soon as the agent’s perception network represented the probe (X or Y), it initiated its decision making process. When a collapsing left or right response boundary crossed the left action probability, the agent engaged in the associated action at that time. This decision making process did not formally specify drift-diffusion-model-like dynamics, with the exception of the decision boundaries (a). However, the decision making process’ created similar constructs. How the components of this process are akin to an evidence accumulation process are identified with the terms and parameters (indicated a, t, v, and z) that are commonly used in the DDM literature.