Wednesday, June 13, 2018

Robotics, Automation and the Cerebellum

(Drake et al. 2010. Gray’s Anatomy for Students 2nd edn)

The cerebellum can learn complex sensorimotor tasks using a simple technique called imitation. If you are a roboticist or an automation expert, you will find the powerful supervised learning technique I describe below of special interest because it could potentially simplify your work. What makes this technique so powerful is its sheer simplicity and its ability to a learn complex tasks very fast.

The First, the Last and Everything in Between

There are two kinds of sensors in the brain. One kind (poor) is used by the neocortex and the other (rich) is used by the cerebellum. Poor sensors come in complementary pairs, stimulus onset and offset. For example, we may have a sensor (A) that fires a single pulse when the amplitude of an audio frequency climbs above a particular level. The complementary sensor (B) would fire when the amplitude falls below the same level. It so happens that there is a train of pulses between A and B but the neocortex does not care about what happens between them. What matters to it is the precise timing of the first and last pulses. Of course, for every type of stimulus, the brain uses many sensors to handle multiple levels or amplitudes.

Unlike the neocortex, the cerebellum is a hungry beast because it wants it all: the first, the last and everything in between. Thus every sensory input going into the cerebellum is a train of pulses. This might seem like a total waste of pulses but it is actually essential to the learning method used by the cerebellum. Again, for emphasis, I differentiate between the two types of sensors by referring to cerebellum sensors as rich sensors. Single pulse sensors (first and last) are poor sensors.

Cerebellar Neurons

Cerebellar cortical neuronal circuits. Mossy fibers from pontine nuclei etc., send excitatory synaptic outputs to granule cells. A granule cell forms one or a few excitatory glutamatergic synapses on a Purkinje cell, where LTD occurs depending on the activity of the granule cell and a climbing fiber. Molecular layer interneurons (stellate and basket cells) receive excitatory synaptic inputs from granule cells and inhibit Purkinje cells. At inhibitory GABAergic synapses between a stellate cell and a Purkinje cell, rebound potentiation (RP) is induced by climbing fiber activity.
Tomoo Hirano and Shin-ya Kawaguchi
Regulation and functional roles of rebound potentiation at cerebellar stellate cell—Purkinje cell synapses

The main neuron in the cerebellum is the Purkinje cell (PC) which was named after its discoverer, Czech physiologist Jan Evangelista PurkynÄ›. There are approximately 15 million PCs in the human brain. Each PC emits pulses that are used to control a motor effector. They are arranged in tight formations like a forest with lots of parallel fibers running through the dendrites like telephone wires. Each PC can receive signals from as many as 200,000 parallel fibers. Each parallel fiber is a long bifurcated axon of a granule cell, an intermediary neuron that conducts sensory signals arriving on mossy fibers. However, not all of the input signals arriving on mossy fibers have sensory origins. Some are control signals that are used to inhibit the PCs when necessary. These fibers are likely used for task control. They do so via so-called Stellate and Basket cells which make inhibitory synaptic connections with the PCs.

Supervised Learning in the Cerebellum

The second most important entity in the cerebellum is the climbing fiber (CF). There is one CF for every PC. The CF carries training input signals to the PC. Those signals originate from the inferior olivary nucleus in the medulla oblongata which relays motor signals from motor effectors in the spinal cord to the cerebellum.

In order to understand how the cerebellum is trained to perform a sensorimotor task, it is important to know how motor effectors work. An effector is the opposite of a sensor. It, too, has a first (start) and last (stop) pulse and pulses in between. It is attached to a muscle and generates a train of pulses that contracts the muscle for as long as the pulses keep coming. The cerebellum accomplishes motor control via the use of a mix of excitatory neurons, inhibitory neurons and tonic neurons. The latter are neurons that continually generate pulses unless they are inhibited. The exact circuit details are not important and is implemented differently in various animals. What matters are the principles.

Learning in the cerebellum consists of finding parallel fiber inputs to Purkinje cells that activate and deactivate motor effectors at the correct time. The training occurs while the neocortex is going through a given sensorimotor task. The cerebellum learns to faithfully imitate the task. Remember that parallel fibers carry pulse trains from rich sensors. These fibers try to make synaptic connections with as many PCs as possible. To train a PC, the training mechanism only needs to send corrective signals to the PC via the climbing fiber whenever the associated motor effector stops firing. The CF signal will suppress and disconnect any parallel fiber connection that is still receiving sensory pulses. The end result is that only parallel fibers that cause the PC to fire and stop firing at the right time will remain connected.

Once the cerebellum has fully learned a task, the neocortex can just turn it on or off whenever it needs to in order to focus on other important matters.


This training system can be put to good use in all sorts of applications that require automation. Notice that there is no need for either pattern detectors or a conventional multi-layered neural network. Lots of simple rich sensors will do the trick. Sensors are essentially connected directly to motor effectors. Potential applications can range from self-driving trains, cars and buses to self-flying aircrafts and self-navigating ships. The learning system simply learns by imitating human operators.

Robots might be a little harder to train. It would require a human trainer to wear a harness fitted with special sensors that can record precise movements. These could then be used as training signals for the robotic cerebellum. I expect training to be extremely fast.

Coming Soon

In an upcoming article, I will describe how I got my understanding of the cerebellum. Stay tuned.

Thursday, June 7, 2018

I'm Rather Busy But Cerebellum Post Is Coming Soon

"Behold, I stand at the door and knock"

"If anyone hears my voice and opens the door, I will come in to him and will dine with him, and he with me." Believe it or not, this metaphor from the Book of Revelation is the essence of supervised sensorimotor learning in the cerebellum. It is as simple as it is powerful. If you are into robotics, you will not want to miss this. Stay tuned.