Notes for week 7

Barbara

Shimon

The computational nature of motor control

The "plant"

Motor control tasks in Tetrapoda

• Posture maintenance (how not to fall over)
• sensorimotor coordination
• "Simple" movement (e.g., reaching, vocalization)
• sensorimotor coordination
• Inverse problems:
  forward internal models can predict sensory consequences from efference copies of issued motor commands;
  inverse internal models can calculate necessary feedforward motor commands from desired trajectory information.
• inverse kinematics
• inverse dynamics
• key computational issues:
  • ill-posedness
  • complexity
• generally applicable optimization strategies conceivable?
• Sequenced movements (e.g., gait, speech)
• astronomical number of degrees of freedom
• complex, opaque relationship between reward and motor program

Doya

• According to Doya (Fig.1), unsupervised learning = open-loop control [a bit misleading; also there is such a thing as self-supervision]
• ["Simple" vs. "complex" spikes: see info on cerebellar Purkinje cells]
• "The use of internal models of the body and the environment can improve the performance of motor control. Such internal models could be acquired by supervised learning with the motor command as the input and the sensory outcome as the teacher signal." [cf. Hod Lipson's robots]
• "Furthermore, for supervised or reinforcement leaning, it is often helpful to use unsupervised learning algorithms to extract the essential information in the raw sensory input."

Atallah et al.

• One of the computational trade-offs: "Active maintenance (often referred to by the more general term of working memory) requires relatively isolated representations so that information does not spread out and get lost over time. In contrast, the overlapping distributed representations of posterior cortex support spreading associations and inference by allowing one representation to activate aspects of other related representations."
• Traditional verbal descriptions of and distinctions (e.g., familiarity and recall; procedural and declarative) between memory functions are inadequate: they do not capture the true computational distinctions.