2 Siding ReLU via Forward Projections

2 Siding ReLU via Forward Projections

The ReLU acivation function in neural networks has special properties that allow it to be considered in a different way to other activation functions.

The Switching Viewpoint

You can view ReLU as being non-conducting when x<=0 and fully conducting when x>0. 
It is a switch which is automatically turned on when x>0 and automatically turned off when x<=0. Which is an ideal rectifier in electrical engineering terms. 
Hence the term Rectified Linear Unit.
A switch isn't there when it is on from the point of view of the flowing electricity (or analogous thing.)  Electricity flows through pushed together switch contacts the same as through the wires to the switch. 
All the ReLUs in a neural network that are conducting, wire together various sub-components of the network.
The wiring being complete the ReLUs become essentially invisible, until one or more ReLUs change state.
Since neural networks are computed in a feedforward fashion the wiring together of weighted sums happens in a feedforward way. In the layers already computed, wired together structures of weighted sums exist. In the current layer swirching decisions are made based on those preexisting structures of weighted sums acting on the input vector.
Wired together structures of weighted sums can be simplifed to a single weighted sum using linear algebra, if you wanted to examine the network for a particular input.

2 Siding ReLU


Five forward connected weights from neuron N to the next layer.

Consider the ReLU neuron N in the diagram. The output of N is forward connected to the next layer by 5 weights, one for each neuron in the next layer.
When the value x (the input to the ReLU function of neuron N) is greater than zero the pattern of the forward connected weights is projected into the next layer with intensity x.
When x<=0 nothing is projected into next layer.  Then the neuron N is in a non-conducting state and it entirely blocks any information trying to flow through it.

Why not have an alternative set of forward connected weights for when x<=0 and project that different pattern onto the next layer with intensity x (x being negative this time)?
Alternative set of 5 forward connected weights when x<=0.

This prevents the information loss caused by ReLUs impeding the flow of information in the network, including early loss of information before it can be processed. 
It also causes the cost landscape during optimization to be far less rough than when single sided ReLUs are used. This can greatly help many neural network training/optimisation algorithms

Here is an example of 2 Siding ReLU using the processing P5 library. You get a local copy of the code which you can interact with.

https://editor.p5js.org/congchuatocmaydangyeu7/sketches/iZwTnULQ2



The end.















 




Comments

  1. AI462April 6, 2023 at 5:22 AM

    This paper has been mentioned:
    https://paperswithcode.com/method/crelu#:~:text=CReLU%2C%20or%20Concatenated%20Rectified%20Linear,non%2Dsaturated%20non%2Dlinearity.

    ReplyDelete

Post a Comment

Popular posts from this blog

Switch Net

Image
 Switch Net Switch Net. Switch Net is a neural network based on using the fast Walsh Hadamard tranform (WHT) as a fixed set of layer weights with a parametric switching based activation function. The cost of the WHT is nlog2(n) add subtracts as opposed to n squared fused multiply adds for a conventional dense neural layer's weight calculations. A fixed random pattern of sign flips is applied to the input data so that the first WHT fairly distributes information with a Gaussian distribution across its outputs. Then a switching decision is applied to each output of the WHT and the output is multiplied by one weight or another.  The section between the red lines is basically a layer and can be repeated a number of times. A final WHT is done to combine the last lot of switched weight outputs. Some example code is here: https://editor.p5js.org/congchuatocmaydangyeu7/sketches/7ekZTwQMF One slight issue is if the switching weights for a particular output have opposite signs then the resul
Read more

Switch Net 4 (Switch Net N) Combine multiple low width neural layers with a fast transform.

Image
 Switch Net 4  Switch Net 4. Random sign flipping before a fast Walsh Hadamard transform results in a Random Projection. For almost any input the result is a Gaussian distributed output vector where each output element contains knowledge of all the input elements. With Switch Net 4 the output elements are 2-way switched using (x<0?) as a switching predicate. Where x is the input to the switch.  The switches are grouped together in units of 4 which together form a small width 4 neural. When a particualr x>=0, the pattern in the selected pattern of weights is forward projected with intensity x. When x<0 a different pattern of forward selected weights is again projected with intensity x (x being negative this time.) If nothing was projected when x<0 then the situation would be identical to using a ReLU. You could view the situation in Switch Net 4 as using 2 ReLUs one with input x and one with input -x.  The reason to 2-way switch (or +- ReLU) is to avoid early information los
Read more

Artificial Neural Networks

Image
Artificial neural networks An artificial neuron. The basic behavior of an artificial neural network is determined by the dot product (weighted sum) operator in each neuron. This also has a geometric interpretation. The dot product of A and B is the magnitude (vector length) of vector A by the magnitude of vector B by the cosine of the angle between them.  Dot product video. To understand information storage in a weighted sum it helps to know the central limit theorem. Statistical properties of the dot product. The central limit theorem. If you store 1 <vector,scalar> association in a weighed sum the weight vector will point in the same direction as the input vector. Store 2 <vector,scalar> associations and both input vectors will point some angle away from the weight vector. As a result the scalar output will be more sensitive to small changes in the input vectors. A reduction in an initial mild error correction capacity. If yo
Read more