Using stimulation to modulate activity in neuronal circuits has broad applicability in both clinical and basic scientific domains. A pressing need in these applications is the design of principled stimulation inputs that go beyond the ‘square pulse’ type waveforms commonly used in current practice. Such waveforms are likely inefficient and too blunt to achieve neuroscientific goals such as investigations of temporal neural coding. Control theory can help to obviate this issue by optimizing extrinsic inputs for maximizing activity-based objective functions. However, in this regard the complexity and diversity of neural dynamics presents a major analytical challenge for the deployment of classical control methods. Indeed, progress in neurostimulation design and optimization has usually required abstraction of dynamics to low-dimensional canonical models, or assumptions of heterogeneity across a neuronal population. This talk will explore approaches that do not require an explicit mathematical model for the circuit to be controlled. These approaches blend ideas from adaptive control theory, system identification and machine learning, wherein the controller builds a representation of the targeted circuit in an online fashion. The representation is not a biophysical model, but rather a recurrent network construct that approximates the dynamics of the circuit in question. Thus, control can be performed on large, irregularly connected populations of spiking neurons, at the expense of needing sufficient observations to build the representation. Interestingly, for some formulations, the learned control strategy itself produces spiking activity, providing interpretations of how control objectives may manifest intrinsically within neuronal networks.