"""
A family of function objects for changing a set of weights over time.
Learning functions come in two varieties: LearningFunction, and
CFPLearningFunction. A LearningFunction (e.g. Hebbian) applies to one
set of weights, typically from one ConnectionField. To apply learning
to an entire CFProjection, a LearningFunction can be plugged in to
CFPLF_Plugin. CFPLF_Plugin is one example of a CFPLearningFunction,
which is a function that works with the entire Projection at once.
Some optimizations and algorithms can only be applied at the full
CFPLearningFn level, so there are other CFPLearningFns beyond
CFPLF_Plugin.
Any new learning functions added to this directory will automatically
become available for any model.
"""
import param
from topo.base.functionfamily import LearningFn
# Imported here so that all learning functions will be in the same package
from topo.base.functionfamily import Hebbian,IdentityLF # pyflakes:ignore (API import)
[docs]class AntiHebbian(LearningFn):
"""
Opposite of the basic Hebbian rule.
Same as Dayan and Abbott, 2001, equation 8.3, except that the
weight change is negative. I.e., each weight decreases in
proportion to the product of this neuron's activity and the input
activity.
Requires some form of output_fn normalization for stability.
"""
def __call__(self,input_activity, unit_activity, weights, single_connection_learning_rate):
weights -= single_connection_learning_rate * unit_activity * input_activity
[docs]class Oja(LearningFn):
"""
Oja's rule (Oja, 1982; Dayan and Abbott, 2001, equation 8.16.)
Hebbian rule with soft multiplicative normalization, tending the
weights toward a constant sum-squared value over time. Thus this
function does not normally need a separate output_fn for normalization.
"""
alpha=param.Number(default=0.1,bounds=(0,None))
def __call__(self,input_activity, unit_activity, weights, single_connection_learning_rate):
weights += single_connection_learning_rate * (unit_activity * input_activity -
self.alpha * (unit_activity**2) * weights)
[docs]class Covariance(LearningFn):
"""
Covariance learning rule supporting either input or unit thresholds.
As presented by Dayan and Abbott (2001), covariance rules allow
either potentiation or depression of the same synapse, depending
on an activity level. By default, this implementation follows
Dayan and Abbott equation 8.8, with the unit_threshold determining
the level of postsynaptic activity (activity of the target unit),
below which LTD (depression) will occur.
If you wish to use an input threshold as in Dayan and Abbott
equation 8.9 instead, set unit_threshold to zero and change
input_threshold to some positive value instead. When both
thresholds are zero this rule degenerates to the standard Hebbian
rule.
Requires some form of output_fn normalization for stability.
"""
unit_threshold =param.Number(default=0.5,bounds=(0,None),
doc="Threshold between LTD and LTP, applied to the activity of this unit.")
input_threshold=param.Number(default=0.0,bounds=(0,None),
doc="Threshold between LTD and LTP, applied to the input activity.")
def __call__(self,input_activity, unit_activity, weights, single_connection_learning_rate):
weights += single_connection_learning_rate * (unit_activity - self.unit_threshold) * (input_activity - self.input_threshold)
[docs]class CPCA(LearningFn):
"""
CPCA (Conditional Principal Component Analysis) rule.
(See O'Reilly and Munakata, Computational Explorations in
Cognitive Neuroscience, 2000, equation 4.12.)
Increases each weight in proportion to the product of this
neuron's activity, input activity, and connection weights.
Has built-in normalization, and so does not require output_fn
normalization for stability. Intended to be a more biologically
plausible version of the Oja rule.
Submitted by Veldri Kurniawan and Lewis Ng.
"""
def __call__(self,input_activity, unit_activity, weights, single_connection_learning_rate):
"""
Update the value of the given weights matrix based on the
input_activity matrix (of the same size as the weights matrix),
the response of this unit (the unit_activity), and the previous weights
matrix governed by a per-connection learning rate.
"""
weights += single_connection_learning_rate * unit_activity * (input_activity - weights);
[docs]class BCMFixed(LearningFn):
"""
Bienenstock, Cooper, and Munro (1982) learning rule with a fixed threshold.
(See Dayan and Abbott, 2001, equation 8.12) In the BCM rule,
activities change only when there is both pre- and post-synaptic
activity. The full BCM rule requires a sliding threshold (see
CFPBCM), but this version is simpler and easier to analyze.
Requires some form of output_fn normalization for stability.
"""
unit_threshold=param.Number(default=0.5,bounds=(0,None),doc="Threshold between LTD and LTP.")
def __call__(self,input_activity, unit_activity, weights, single_connection_learning_rate):
weights += single_connection_learning_rate * unit_activity * input_activity * (unit_activity-self.unit_threshold)
_public = list(set([_k for _k,_v in locals().items()
if isinstance(_v,type) and issubclass(_v,LearningFn)]))
# Automatically discover all .py files in this directory.
import os,fnmatch
__all__ = _public + [f.split('.py')[0] for f in os.listdir(__path__[0]) if fnmatch.fnmatch(f,'[!._]*.py')]
del f,os,fnmatch
