Adaptive signals are typically, any signal carried on an axon (axon level
) that is connected to the adaptive inputs
of one or more neurons
. More abstractly, it can refer to any propagated signal (AL
) that is being used to have an effect on changes to connection-strength
s in a neuron.
Adaptive Inputs (I-D)
in Netlab have two special input connection-points, where axons
can be connected so that their signals can be used to effect weight-changes. These are called Adaptive Inputs.
Adaptive inputs are also called Increase
inputs, or simply I-D
. . . . . . .
- The Increase or 'I' input is generally used to control how much connection-strengths (traditionally represented by weights) are enhanced, and is denoted by an inward pointing arrow-head ('v') on schematics of neurons. The increase input is shown on the top of the above neuron-schematic, but it can be placed anywhere on the neuron's body.
- The Decrease or 'D' input is generally used to control how much connection-strengths are reduced, and is denoted by a small 'o' on the inside edge of the neuron body. It can be anywhere on the neuron body, but is shown at the bottom of the neuron shown in the above diagram.
In traditional ANN
s, the terms "increase", and "decrease" have some confusion associated with them, because weight-values
often use sign to denote connection-type. Positive (+) weight values are used for excitatory connection strengths, and negative (-) weight values represent inhibitory connections. In both cases, the absolute value of the weight represents the connection strength
Because of this, in traditional ANNs, increasing (I) the weight is not necessarily making the connection-strength stronger. In a traditional ANN
, increasing an inhibitory (negative)
weight will reduce
the connection strength
represented by that weight. That is, it will make a negative weight-value more positive, thereby moving its absolute value closer to zero.
Normally, in Netlab's™ connection-strength-centric model, increasing
an inhibitory synapse
would be the same as making the signed weight-value
more negative (increasing its absolute value). Likewise, reducing
of an inhibitory connection in Netlab would be the equivalent, in a traditional ANN
, of making the weight-value more positive.
. . . . . . .
How Ambiguities Are Facilitated In Netlab
The ambiguity embodied in traditional neural-network practices is accommodated in Netlab's description language (Noodle™) by permitting the designer to specify whether a given learning method of a given weight-layer responds to adaptive cues in the traditional way. In this case, the weights within the weight-layer can be specified to be trained as:
- signed-value based (denoted with 'PN', meaning Positive/Negaitve) - the traditional practice, in which increase always means to make more positive, or
- connection-strength based (denoted with 'CS', meaning Connection-Strength) - in which increase will increase the absolute strength of the connection, regardless of whether it is inhibitory or excitatory. In Netlab™, this is the default behavior and does not need to be explicitly set.
This is discussed in more detail in the glossary-entry for Connection Strength
. . . . . . .
Adaptive Signaling in Netlab
Signals carried on axons
connected at adaptive inputs
are called adaptive signals
. Any given signal in Netlab
, which is produced by a neuron
or other signal source (such as a sensor) can serve as both reactive, and adaptive in a given neural network
. Adaptive signaling and feedback mechanisms are roughly analogous to the signaling mechanisms that are used to achieve reinforcement learning
in biological neural networks.
Occasionally the term reinforcement may also be used here to refer to adaptive signal-values, feedback, and mechanisms. Adaptive inputs are special inputs to a neuron, which carry signals that can be used by the neuron's learning
method to produce an error value. The error value is used by the weight
-adjustment algorithm to alter the value of the neuron's connection weights
In other words, adaptive signals are normal signal values, such as those produced at the outputs of neurons
, input nodes, or by sensors, that are used in a special way within the neural network
. When serving in an adaptive capacity, signal values are not connected to the neuron's normal inputs to affect its output, but are instead used to effect changes to the neuron's connection weights
. . . . . . .
Implicit Adaptive Signaling
In Netlab™ weight-changes can be affected, and effected through a variety of means beyond axon-level
signals present at adaptive inputs. One such alternative mechanism for altering the value of weights in Netlab™ is CI
s (Chemical Influences), which can both directly alter connection-strengths, or affect how other factors alter connection-strengths. For example, they can change the learning or forget rate at a given weight-layer