Spar Restricted Boltzmann Machines as a model of the Mirror Neuron System
Abstract
I will defend a two-fold hypothesis. First (1), Restricted Boltzmann Machines (RBM) can successfully emulate the human Mirror Neuron System (MNS), by using association. This supports the association hypothesis, which states that Mirror Neurons are a by-product of associating perception with motor codes. Second (2), Spar Coding is an necessity for Mirror Neurons to emerge from association learning.
Methods: I simulated a datat with three actions and two goals. Each stimuli has five features; three to indicate the actions, two to indicate the goal. I trained Spar RBMs of various sparsities and sizes to model the MNS.
Conclusion 1: RBMs prove to be successful in emulating various aspects of MNS behavior. This includes action execution, obrvation, imitation, goal inference, dealing with missing values (i.e. in the dark) and handling multiple modalities (i.e. integrate vision and proprioception). The performance, strength and certainty of respons of the model in different circumstances is similar to data from experiments.
Conclusion 2: The Mirror Units that emerge are only dependent on the Spar Coding. They are robust to network size, although size tends to diminish strength of the respon. The units are capable of reprenting one or more caus (i.e. a single action, goal, or both). The optimal sparsity turns out to be , where is the number of distinctions the Unit needs to make.
While Mirror Units emerge, association learning was insufficient to create two distinct populations , as found in the brain. This might be solved with the addition of classification learning. Keywords: Restricted Boltzmann Machines, Mirror Neuron System, Spar Coding, Associative Memory.
Bachelor Thesis Artificial Intelligence Supervid by Sebo Uithol & Pim Halager Radboud University Nijme推优自我介绍
gen Mark Marijnisn
September 22, 2011 (revision)
Table of Contents
1. Introduction (3)
2. The Mirror Neuron System (5)
3. Restricted Boltzmann Machines (7)
4. Spar Coding (11)
5. Methods (15)
6. Results (21)
7. Conclusion & Discussion (30)
References (35)
Note for supervisors: (37)
1.Introduction
Mirror Neurons have caud m uch excitement since thei高山流水成语
r discovery in the early 90’s. Mirror Neurons are neurons that become active when an action is executed, as well as when that action is obrved being done by another. Not only do they link action with perception, they also link lf with others. This provides a foundation for theories about empathy, imitation and action understanding. (Heyes, 2010).
There are two competing theories that explain the existence of Mirror Neurons: the adaptation and the association hypothesis. (Heyes, 2010). The first theory argues that Mirror Neurons are the result of natural lection for either action understanding or imitation learning. The cond theory views Mirror Neurons as a by-product of association learning. It suggests Mirror Neur咳特灵片
ons emerge by associating the perception with the execution of an action.
This thesis will provide evidence to历史人物故事
support the association hypothesis. However, I will argue that associative learning is not enough for Mirror Neurons to emerge: we need Spar Coding too. Spar Coding limits the number of cas a neuron responds to, or limits the number of neurons that are active at any given time. Without Spar Coding, associations can be learned without creating neurons th关于除夕节的诗句
at过桥米线的传说
respond to a single goal or action.
In order to do this, I have modeled the human Mirror Neuron System (MNS) with a Restricted Boltzmann Machine (RBM). RBMs are undirected belief networks that learn a generative model of the obrved data. They can be ud as an associative model, becau given a partial obrvation, they can generate the remainder. I ud this to associate stimulus (perception) with respon (action). This way, RBMs can perform typical MNS tasks such as action execution, obrvation, imitation and goal inference.
There are more computer models of the MNS bad on association. For example, Chaminade et al ( 2008) have learned viomotor associations to a robotic hand from lf-obrvation. This could be ud to imitate a human hand. However, they did not focus on the occurrence of Mirror Units (which may not have emerged due to the lack of spar coding). There are a few other MNS models that u the association hyp粤语歌词
othesis, but none of them investigate sparsity. See Oztop et al (2006) for an excellent review on the different goals and methodologies ud to model the MNS. To be clear, this thesis investigates two hypothesis:
1.Restricted Boltzmann Machines can successfully emulate MNS behavior, by functioning as an
associative memory.
2.Spar Coding is a necessity for Mirror Units to emerge from learning associations. In fact,
there is an optimal sparsity for Mirror Units.
I will focus on the accuracy of the model to verify the first hypothesis. So, unlike other models, I will focus not only on the performance of MNS. Instead, I also focus on various other aspects, such as the strength of the respon under different circumstances. For example: Do Mirror-Units respond wi
th less strength if there is no goal-context, just like real Mirror-Neurons? Note that I u “units” when referring to the computer model, a nd “neurons” when referring to actual biological neurons.
In order to investigate the cond hypothesis, I will experiment with different network sizes and sparsities.
The thesis is structured as follows: First, I will briefly describe relevant properties of Mirror Neurons in chapter 2. Then, I will explain what RBMs are and how they learn in chapter 3. The RBMs will u Spar Coding, which I describe in chapter 4. The MNS, RBM and Spar Coding provide the background knowledge required to understand the experiment as described in chapter 5. I will prent the results in chap有月的成语
ter 6, and draw conclusions in chapter 7.
2.The Mirror Neuron System
In this ction, I will explain briefly various properties of Mirror Neurons that I will attempt to model.
Mirror Neurons are found using single cell-recordings in the F5 area of the macaque monkey, which is a part of the premotor cortex. Only 92 of the 532 neurons (17%) that are recorded by Galle et al (1996) have mirror properties. The other neurons only respond to action execution.
While single-cell recordings are not possible in humans due to the destructive surgery that is required, various indirect evidence suggest the existence of a human Mirror Neuron System (MNS). (Rizzolatti & Craighero, 2004)
Mirror Neurons only respond to goal-directed movements. They do not respond to random movements or movements without a goal (i.e. grasping without an object that can be grasped). Note that in thi 英语专业实习报告
s ca, “goal” refers to a successful execution of the grasping action: You have an object in your grasp. Experiments were conducted in which the object is hidden behind a screen. When the obrving monkeys knows there is an object, half of the Mirror Neurons still fire. A quarter of them even fire at the same strength as with full vision. (Umilta, et al., 2001)
In fact, the actual movement can be completely opposite, as long as the goal is the same. In an experiment, monkeys grasped an object with either pliers or rever pliers. (Umilta, 2008) The former required the monkey to clo the hand to grasp, while the latter required the monkey to open its hand. In spite of this difference, Mirror Neurons were found that respond to both cas.
So Mirror Neurons do not respond exclusively to the visual perception of movement, but they are able to take contextual information into account, such as the object grasped or tool ud.
In both action execution and obrvation, monkeys could e the action being executed. To exclude visual perception as explanation for action execution, monkeys executed actions in a dark room. The Mirror Neurons still fired (Galle, Fadiga, Fogassi, & Rizzolatti, 1996). In other experiments, the monkeys only heard action execution, and 15% still responded. Rizzolatti and Craighero (2004) conclude that mirror neurons can respond to multiple modalities and fire as soon as the perception contains enough clues to infer an action.
The clues can be quite general, becau Mirror Neurons also respond to perception of other species. This means Mirror Neurons do not respond to direct perceptions, but to more general features extracted from them. It is thought that Mirror Neurons respond if the action is in the action repertoire of the obrver. For example, human mirror neurons respond to monkeys that are “talking”, but not to a dog that is barking. (Buccino, et al., 2004)
This is congruent with what we know about the brain regions involved in the human MNS. Kilner et al (2007) describe this as follows: The Superior Temporal Sulcus (STS) provides the input for the human MNS. The STS is known to encode body positions and specific actions. The STS is reciprocally connected to area PF of the inferior parietal cortex. The area PF, in turn, is reciprocally connected to the F5 area of the premotor cortex. While the STS is often considered a part of the MN
S, the do not fire on action execution. Mirror Neurons are only found in area PF and F5.
