Supervised Spike-Timing-Dependent Plasticity: A Spatiotemporal Neuronal Learning Rule for Function Approximation and Decisions

How can an animal learn from experience? How can it train sensors, such as the auditory or tactile system, based on other sensory input such as the visual system? Supervised spike-timing-dependent plasticity (supervised STDP) is a possible answer. Supervised STDP trains one modality using input from another one as “supervisor.” Quite complex time-dependent relationships between the senses can be learned. Here we prove that under very general conditions, supervised STDP converges to a stable configuration of synaptic weights leading to a reconstruction of primary sensory input.

Human/animal communications, language, and evolution

The article compares the research programs of teaching symbolic language to chimpanzees, pointing on the dichotomy between artificial language vs. ASL, and the dichotomy between researchers who decided to establish emotional relationships between themselves and the apes, and those who have seen apes as instrumental devices. It is concluded that the experiments with the most interesting results have been both with artificial language and ASL, but with strong affiliation between researchers and animal involved in the experiments. The experiments on talking apes are not so much experiments in psycholinguistics (how far can animal learn human language) but wonderful experiments in the communities of communication between human beings and great apes.

Contrasting approaches to a theory of learning

The general process view of learning, which guided research into learning for the first half of this century, has come under attack in recent years from several quarters. One form of criticism has come from proponents of the so-called biological boundaries approach to learning. These theorists have presented a variety of data showing that supposedly general laws of learning may in fact be limited in their applicability to different species and learning tasks, and they argue that the limitations are drawn by the nature of each species' adaptation to the particular requirements of its natural environment. The biological boundaries approach has served an important critical function in the move away from general process learning theory, but it is limited in its ability to provide an alternative to the general process approach. In particular, the biological boundaries approach lacks generality, it is in some respects subservient to the general process tradition, and its ecological content is in too many cases limited to ex post facto adaptive explanations of learning skills. A contrasting, ecological approach to learning, which can provide a true alternative to general process theory, is presented. The ecological approach begins by providing an ecological task description for naturally occurring instances of learning; this step answers the question: What does this animal learn to do? The next step is an analysis of the means by which learning occurs in the course of development, answering the question: How does the animal learn to do this? On the basis of such analyses, local principles of adaptation are formulated to account for the learning abilities of individual species. More global principles are sought by generalization among these local principles and may form the basis for a general ecological theory of learning.

Visual Responses by Octopus to Crabs and other Figures Before and After Training

1. Octopuses were fed for some days with fishes in their homes at one end of the tank, and the effect of figures shown at the opposite end was tested. 2. Any small moving object may then be attacked, but the probability of attack is low and the delay long. 3. Crabs and a black horizontal rectangle were attacked more often than other figures. If a figure is withdrawn on the occasions when it is attacked, then the probability of attacks remains low and may slowly decrease. 4. If the octopus is allowed to eat crabs the probability of attacks on them quickly rises towards unity. If food is given when some other figure (say a black vertical rectangle) is attacked, then this figure will later be attacked. An octopus makes attacks more regularly on a figure that has recently been associated with food than on any other figure. 5. This learning to attack a given figure can take place after removal of most of the vertical lobe, but the learning is slower and responses less accurate than in the normal animal. 6. Octopuses from which the vertical lobe has been removed attack crabs more readily than other figures. The system that ensures attack therefore resides elsewhere, probably in the optic lobes. 7. If shocks are given when a previously neutral figure is attacked, then the probability of further attacks on this figure rapidly approaches zero. Attacks on other figures or crabs remain frequent. If the vertical lobe is then removed, no attacks are made on such an originally neutral figure. 8. If a figure that has been associated with food is then rewarded with shocks, the probability of attack falls more slowly than if the figure had been neutral. If the vertical lobe is now removed, the frequency of attacks rises. The memory preventing attack is therefore partly resident in the vertical lobe. 9. The greater the amount of previous training to attack, the more slowly does the animal learn not to attack. This applies also if the training to attack is given after vertical lobe removal. 10. After vertical lobe removal, shocks given when crabs are attacked rapidly reduce the probability of attack if the octopus had not previously eaten crabs while in the laboratory. If such feeding had taken place, the probability of attack falls only slowly without the vertical lobes. 11. The experiments show that information stores associating a given figure with either food or shock can be set up after removal of the vertical lobes but tend to dissipate more rapidly than in a normal octopus.