## 28 June 2020

### Interpreting Neural Networks by Reducing Nonlinearities during Training

Machine learning and neural networks are used more and more widely, despite the fact that they mostly function as black boxes, and we have little insight into how they generate their output. This makes it difficult to scrutinize and account for decisions made using these systems.

In this note I describe a simple solution that makes it possible to extract rules from a neural network that employs Parametric Rectified Linear Units (PReLUs). This is done by introducing a force, applied in parallel to backpropagation, that aims to reduce PReLUs into the identity function, which then causes the neural network to collapse into a smaller system of linear functions and inequalities suitable for review or use by human decision makers.

As this force reduces the capacity of neural networks, it is expected to help avoid overfitting as well.

Introduction

Machine learning solutions and, more specifically, neural networks are used in more and more areas of our lives. At the same time, since it is difficult to grasp at any level how they function and so they can be seen as black boxes, there is a reluctance in their adoption in the case of companies who need to remain accountable for business and/or legal reasons. This fact can also impact the trust the public places in such systems.

Because of this, providing tools that can help us understand how these models arrive at a certain decision is an area of active research. The tools range from visualization tools, tools that identify the most significant input, tools that allow experimenting with the input data to understand the relative importance the model attributes to different signals, and, the focus of this note, significantly reducing the complexity of a neural network to the point where “rules” for decisions could be extracted from it, which can then be reviewed or even used by human decision-makers.

One form these rules can take is a closed, relatively simple algebraic expression (for example, a linear function) which has a behavior more predictable for humans than the full function implemented by a neural network involving parameters and nonlinear functions in the range of millions.

In this note I describe a method of reducing the complexity of a neural network during its training, which then allows describing its output as a simple system of linear functions limited to regions defined by linear inequalities. In particular, feed-forward neural networks are considered where all nonlinearities are Parametric Rectified Linear Units (PReLUs) defined as

f(x) = max(0, x) + a min(0, x),

where 0 ≤ a ≤ 1.

Let the input of the network be the vector (or tensor, which we flatten) I = (i1 , i2 , . . . , in) and the output be the vector (or tensor) O = (o1 , o2, . . . , om).

The Neural Network as a Combination of Linear Functions

It is easy to see that the output O of such a network can be expressed as a set of linear combinations of ix, each restricted to a particular region of input values by a set of linear inequalities on ix (the conditions).

Intuitively, this is true because the output is a linear function of the input as long as (that is, inside regions of the possible input vectors in which) none of the inputs of PReLUs cross from negative to positive or vice versa; and because the borders of these regions are linear as well.

More formally, we can show that this statement is true by constructing a suitable system of linear combinations and inequalities for any given neural network. For all PReLUs in the network, let us decide whether their inputs are negative, or nonnegative. If the number of PReLUs is δ, this means 2δ possible scenarios. In each scenario, the neural network becomes completely linear, with the output being a linear function of the input, and, in fact, with all partial results inside the network being linear functions of the input as well. This entails that the inputs of PReLUs are also linear functions of the input I, yielding the δ linear inequalities (the conditions) that mark out the region of the input space where this scenario applies.

Naturally, this system of 2δ linear combinations yielding the output, each controlled by δ conditions is likely to be redundant, as the conditions may not be independent, or may be contradictory, in which case some of the 2δ regions of the input space will be empty. But this system will be an accurate description of the output of the neural network nevertheless.

While a linear function may seem easy to interpret and to predict for us humans, clearly 2δ functions will not be so if δ is in the range of thousands or millions as in most modern neural networks. We therefore need to construct an approximation of the output of our neural network which is linear in much larger regions of the input space, and has far less different regions as well. This is equivalent to trying to find another neural network that approximates our network well, but contains a small number of PReLUs only.

Reducing PReLUs into Linearities

This can be achieved this by continuing training the original neural network while applying a force on each of its PReLUs that moves its parameter a towards 1.

Note that if a = 1, then the PReLU degenerates into the identity function f(x) = x, and ceases to be a nonlinearity. In a sense, it disappears, and the neural network “collapses” around it, inasmuch as the linear mapping that precedes the degenerate PReLU and the one after it can now be combined into a single linear map.

This force therefore finds a balance between approximating the training data and reducing the number of PReLUs, thereby yielding a neural network that is feasible to express as a set of linear functions and inequalities for human consumption. By removing nonlinearities, this force also reduces the capacity of the neural network, and therefore it can be expected to help avoid overfitting.

I chose a force that is independent of a and is applied after each backpropagation step, and moves the parameter by an adjustment rate ηp but never above 1:

at+1 = max(0, min(1, at + ηp sgn(1 − at)))

In my research I considered any PReLU fully linear if

a > 0.9995

ηp was also made dependent on the current error exhibited by the network, to allow the network initially to train without interference, and then pull the PReLU parameters to 1 more and more aggressively. In particular,

ηp = η0 0.01 max(0, − log10(err) − 2)

where η0 = 0.01 is the learning rate for the whole model, and err is the training error. This means that ηp will start becoming non-zero when the training error falls below 0.01, and can grow indefinitely as the training error falls (although the a parameters are clipped anyway).

Example

Sample code demonstrating the above is available at https://github.com/csirmaz/trained-linearization. The code implements training a small neural net with the force on PReLU parameters, and contains logic to extract a system of linear inequalities and combinations from the weights and other parameters of the model.

The neural network itself has 4 input nodes, 3 output nodes, and 4 hidden layers in between with 5 nodes on each. These layers are linked by trainable fully-connected layers with PReLU activation following them, except for the last fully connected layer, which simply provides the output.

The network is expected to learn the following relationships:

out1 = in1 xor in2
out2 = in3 xor in4
out3 = in1 · in2

where all the inputs and outputs are 0 or 1 with some extra noise added to the inputs. The code generates the training data according to the expressions above, for all 24 possible combinations of 0 and 1 for the input. New noise is generated for each batch of training data.

The figure below visualizes a trained model. Two nonlinearities are enough to solve this problem, and indeed the training settled on a solution where only two PReLUs were nonlinear.

(There are 4 input nodes (top) and 3 output nodes (bottom). Red lines denote positive, blue lines negative weights. Lines leading from the right denote bias values. Green boxes mark PReLUs that are considered linear, while red ones mark nonlinear PReLUs.)

Although in the visualization the relationships between the inputs and the outputs are far from clear, once the rest of the PReLUs are considered linear, all the weights collapse into very simple expressions. See below for the output of the code listing these expressions for the 22 cases arising from the two nonlinear PReLUs. These correspond completely to the relationships in the training data.

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 < 0
(PReLU #3 on level 1 is neg. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.00 < 0
(PReLU #1 on level 4 is neg. ln(1-weight)=-2.60)
THEN
out1 = +1.01*in1 +1.01*in2 -0.00*in3 -0.00*in4 +0.00
out2 = -0.00*in1 -0.00*in2 -1.03*in3 -1.03*in4 +2.04
out3 = -0.00*in1 -0.00*in2 +0.00*in3 +0.00*in4 -0.00

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 < 0
(PReLU #3 on level 1 is neg. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.00 > 0
(PReLU #1 on level 4 is pos. ln(1-weight)=-2.60)
THEN
out1 = -1.01*in1 -1.01*in2 +0.00*in3 +0.00*in4 +2.02
out2 = -0.00*in1 -0.00*in2 -1.03*in3 -1.03*in4 +2.04
out3 = +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.01

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 > 0
(PReLU #3 on level 1 is pos. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.00 < 0
(PReLU #1 on level 4 is neg. ln(1-weight)=-2.60)
THEN
out1 = +1.01*in1 +1.01*in2 +0.00*in3 +0.00*in4 +0.00
out2 = -0.00*in1 -0.00*in2 +1.00*in3 +1.00*in4 +0.01
out3 = -0.00*in1 -0.00*in2 -0.00*in3 -0.00*in4 +0.00

IF +0.00*in1 +0.00*in2 -1.00*in3 -1.00*in4 +1.00 > 0
(PReLU #3 on level 1 is pos. ln(1-weight)=-2.58)
IF +1.00*in1 +1.00*in2 +0.00*in3 +0.00*in4 -1.00 > 0
(PReLU #1 on level 4 is pos. ln(1-weight)=-2.60)
THEN
out1 = -1.01*in1 -1.01*in2 -0.00*in3 -0.00*in4 +2.02
out2 = -0.00*in1 -0.00*in2 +1.00*in3 +1.00*in4 +0.01
out3 = +1.00*in1 +1.00*in2 -0.00*in3 -0.00*in4 -1.01


References

[1] Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” Microsoft Research, 6 February 2015. http://arxiv.org/abs/1502.01852.