Prototypography

…  what is stored is not a static representation but the means to dynamically reproduce it.

… unlike the conventional link between a symbol and what the symbol stands for, distributed representations are connected to the world in a non-arbitrary way because the process through which they emerge is a direct accommodation or adaptation to the demands of an external reality.

This is from Philosophy and Simulation: The Emergence of Synthetic Reason by Manuel De Landa (2011):

… The explosive divergence that multicellular organisms underwent six hundred million years ago had many consequences. The diversity of the survival problems that these organisms had to solve, for example, increased enormously as their environments filled with other living creatures capable of affecting and being affected by them. To confront this diversity organisms began to develop internal models to guide their behavior. Even humble motile bacteria, as they swim up a gradient of nutrients toward the point of maximum concentration, can be said to have an internal model of their environment. This internal model is not, of course, based on representations of any kind and it is not a model of the world at large but only the concrete opportunities and risks afforded to bacteria by their immediate surroundings. In other words, from the beginning of life the internal models mediating the interaction between a primitive sensory system and a motor apparatus evolved in relation to what was directly relevant or significant to living beings. With the advent of multicellular organisms and the progressive differentiation of their cells into multiple kinds the scope for internal models was greatly enlarged. In particular, a new kind of biological material, neuronal material, began to grow and interpenetrate the living jelly accumulating at the bottom of the ocean. And with the availability of neurons the capacity to distinguish the relevant from the irrelevant, the ability to foreground only the opportunities and risks pushing everything else into an undifferentiated background, was vastly increased.

… These assemblies of neurons, in turn, sustain the emergent capacity to modulate inherited behavioral responses during the lifetime of an organism. The simplest of these capacities is called habituation: it operates on sensory-motor activities that are innate but it allows organisms to control their intensity. An organism like a hydra, for instance, can gradually decrease its response to a stimulus — can become habituated to it — as experience finds it harmless or avoidable. In other words, habituation transforms a significant stimulus into an insignificant one. The opposite capacity, sensitization, allows these organisms to behave toward a previously irrelevant stimulus as a potential source of opportunity or a risk, that is, it makes the stimulus behaviorally relevant. In a sense these two ancient forms of learning mark the emergence of subjective gradients: before a hydra becomes habituated to a novel stimulus, for example, it may be thought of as being “surprised” by it, a proto-subjective state that slowly diminishes until it disappears as it reaches equilibrium. Behind these subjective gradients, on the other hand, there are objective ones: concentration gradients of electrically charged substances constantly produced and maintained in the fluids inside and outside neurons.

… The capacity to become habituated or sensitized implies only the possession of innate internal models. But with the advent of more complex assemblies of neurons new forms of learning emerged that made possible the creation of internal models during the lifetime of an organism. In particular, ancient insects were capable of the kind of learning referred to as classical conditioning in which an inherited association between a stimulus and a response provides the basis to form a novel association with a previously insignificant stimulus.

… a complete simulation of classical conditioning needs neural nets that display the abilities of a simple brain. To do this two improvements must be made to the basic design: increasing its computational power and giving it the capacity to generalize. The first change is needed because animals may be conditioned to predict an electric shock from two different stimuli (a light and a sound) when presented individually but not when presented together. [ … ] The second change is necessary because when insects learn through classical conditioning it is very unlikely that they will find the exact same conditioned stimulus on every occasion. This implies that, in however simple a form, insects must be able to organize their experience through the use of general “categories” of which similar stimuli are particular instances.

Both improvements can be achieved by supplying a perceptron with one or more intermediate layers between its input and output layers. Because these extra layers are not accessible from the outside they are referred to as hidden layers. A neural net with hidden layers is called a “multilayer perceptron.” Although this change may seem simple it took years to be properly implemented because no one knew how to train a neural net with multiple layers. The core of the training process is the learning rule, a procedure to adjust the connection weights to ensure a progressive convergence to the stable configuration that can produce a target output pattern. An effective learning rule for multilayer perceptrons is based on a simple but powerful idea: instead of only allowing activation to flow forward from the input to the output layers, information about the degree of mismatch between the currently produced output pattern and the desired final pattern is allowed to flow backward. In other words, the rule allows learning through the back-propagation of error ensuring the convergence on the target pattern through the progressive minimization of this error. A training process guided by back-propagation constitutes a search in the space of possible weight configurations, a search process known as “gradient descent” for its resemblance to the dynamic through which a physical gradient cancels itself.

As in the case of simple perceptrons training a multilayer one involves supplying it with a sensory stimulus and a desired motor response. As several different but similar input activation patterns are presented to it, and as information about the correctness of its current output pattern flows backwards, the units in the hidden layer develop their own pattern of activation. Unlike the input and output patterns the pattern in the hidden layer is emergent and its composition is influenced by similarities in the different sensory patterns included in the training set: as the input layer is stimulated by different patterns and as this activation travels to the hidden units those parts of the patterns that resemble each other have a larger effect on the shape of the emergent pattern than those that are different. At the end of the training the connections between the input and the hidden layer have acquired a configuration of strengths that can produce the emergent activation pattern whenever the neural net is presented with a similar stimulus even if it does not belong to the training set. The hidden units, in turn, will cause the right motor response using the stored weight configuration in the connections between the hidden and output layers. This implies that a trained multilayered perceptron has the capacity to generalize from the sensory stimuli it received during training to many other similar stimuli. And this, in turn, suggests that the emergent activation pattern behaves like a non-symbolic representation of sensory stimuli, a representation that can be used to recognize related stimuli. Or to put this differently, during training the hidden layer slowly extracts a prototype from the different patterns contained in the training set, a prototype that it can use to produce the right motor response after training. These emergent non-symbolic representations and the way they capture similarities in what they represent are exactly what we need to understand in what sense insect brains can build internal models of their environment.

… The concept of a prototype extracted from experience and stored as a non-symbolic representation will play such a crucial role in the explanation of animal behavior in the following two chapters that its nature should be made very clear. First of all, an emergent representation is not explicitly stored as such, the product of the learning process being a configuration of connection weights that can recreate it when presented with the right input. In other words, what is stored is not a static representation but the means to dynamically reproduce it. Second, unlike a photograph these representations are dispersed or distributed in all the hidden units and are thus closer to a hologram. This means that they can be superimposed on one another so that the same configuration of weights can serve to reproduce several representations depending on its input, simulating the ability of insects to associate several colors or odors with the presence of food. The dispersed way in which extracted prototypes are represented is so important for the capacity of neural nets to generalize that these emergent representations are usually referred to as distributed representations. Finally, unlike the conventional link between a symbol and what the symbol stands for, distributed representations are connected to the world in a non-arbitrary way because the process through which they emerge is a direct accommodation or adaptation to the demands of an external reality. Thus, multilayer perceptrons offer a plausible account of the intentionality of mental states. When neural nets are studied in a disembodied way, that is, when their input is preselected and prestructured by the experimenter and their output is simply an arbitrary pattern of activation that has no effect on an external environment, this emergent intentionality is not displayed. But the moment we embody a neural net and situate the simulated body in a space that can affect it and be affected by it, the creatures behave in a way that one feels compelled to characterize as intentional, that is, as oriented toward external opportunities and risks.

My most recent previous post from De Landa’s excellent book is here.

-Julie

http://www.unrealnature.com/