The next chapter will questions some fundamental aspects of the formulations so far -- namely the gradients -- and aim for an even tighter integration of physics and learning.
The approaches explained previously all integrate physical models into deep learning algorithms.
Either as a physics-informed (PI) loss function or via differentiable physics (DP) operators embedded into the network.
In the PI case, the simulator is only required at training time, while for DP approaches, it also employed at inference time, it actually enables an end-to-end training of NNs and numerical solvers. Both employ first order derivatives to drive optimizations and learning processes, and we haven't questioned at all whether this is the best choice so far.
A central insight the following chapter is that regular gradients are often a _sub-optimal choice_ for learning problems involving physical quantities.
Treating network and simulator as separate systems instead of a single black box, we'll derive a different update step that replaces the gradient of the simulator.
As this gradient is closely related to a regular gradient, but computed via physical model equations,
we refer to this update (proposed by Holl et al. {cite}`holl2021pg`) as the {\em physical gradient} (PG).
- finally, we'll show how to use inverse functions (and especially inverse PDE solvers) to compute a very accurate update that includes higher-order terms.
All NNs of the previous chapters were trained with gradient descent (GD) via backpropagation, GD and hence backpropagation was also employed for the PDE solver (_simulator_) $\mathcal P$, computing the composite gradient
In the field of classical optimization, techniques such as Newton's method or BFGS variants are commonly used to optimize numerical processes since they can offer better convergence speed and stability.
The PG which we'll derive below can take into account nonlinearities to produce better optimization updates when an (full or approximate) inverse simulator is available.
We'll start by revisiting the most commonly used optization methods -- gradient descent (GD) and quasi-Newton methods -- and describe their fundamental limits and drawbacks on a theoretical level.
Assume two parameters $x_1$ and $x_2$ have different physical units.
Then the GD parameter updates scale with the inverse of these units because the parameters appear in the denominator for the GD update above.
The learning rate $\eta$ could compensate for this discrepancy but since $x_1$ and $x_2$ have different units, there exists no single $\eta$ to produce the correct units for both parameters.
**Function sensitivity** 🔍
GD has inherent problems when functions are not normalized.
Consider the function $L(x) = c \cdot x$ for example where $c \ll 1$ or $c \gg 1$.
Then the parameter updates of GD scale with $c$, i.e. $\Delta x = \eta \cdot c$.
Such behavior can occur easily in complex functions such as deep neural networks if the layers are not normalized correctly.
%
For sensitive functions, i.e. _small changes_ in $x$ cause **large** changes in $L$, GD counter-intuitively produces large $\Delta x$, causing even larger steps in $L$ (exploding gradients).
For insensitive functions where _large changes_ in the input don't change the output $L$ much, GD produces **small** updates, which can lead to the optimization coming to a standstill (that's the classic _vanishing gradients_ problem).
%
While normalization in combination with correct setting of the learning rate $\eta$ can be used to counteract this behavior in neural networks, these tools are not available when optimizing simulations.
Applying normalization to a simulation anywhere but after the last solver step would destroy the simulation.
Setting the learning rate is also difficult when simulation parameters at different time steps are optimized simultaneously or when the magnitude of the simulation output varies w.r.t. the initial state.
**Convergence near optimum** 💎
The loss landscape of any differentiable function necessarily becomes flat close to an optimum
(the gradient approaches zero upon convergence).
Therefore $\Delta x \rightarrow 0$ as the optimum is approached, resulting in slow convergence.
This is an important point, and we will revisit it below. It's also somewhat surprising at first, but it can actually
stabilize the training. On the other hand, it also makes the learning process difficult to control.
### Quasi-Newton methods
Quasi-Newton methods, such as BFGS and its variants, evaluate the gradient $\frac{\partial L}{\partial x}$ and Hessian $\frac{\partial^2 L}{\partial x^2}$ to solve a system of linear equations. The resulting update can be written as
produces the correct units for all parameters to be optimized, $\eta$ can stay dimensionless.
**Convergence near optimum** 💎
Quasi-Newton methods also exhibit much faster convergence when the loss landscape is relatively flat.
Instead of slowing down, they instead take larger steps, even when $\eta$ is fixed.
This is because the eigenvalues of the inverse Hessian scale inversely with the eigenvalues of the Hessian, increasing with the flatness of the loss landscape.
**Consistency in function compositions**
So far, quasi-Newton methods address both shortcomings of GD.
However, similar to GD, the update of an intermediate space still depends on all functions before that.
This behavior stems from the fact that the Hessian of a function composition carries non-linear terms of the gradient.
and as a consequence, the update of any latent space is unknown during the computation of the gradients.
% chain of function evaluations: Hessian of an outer function is influenced by inner ones; inversion corrects and yields quantity similar to IG, but nonetheless influenced by "later" derivatives
The first obvious drawback is the _computational cost_.
While evaluating the exact Hessian only adds one extra pass to every optimization step, this pass involves higher-dimensional tensors than the computation of the gradient.
As $\frac{\partial^2 L}{\partial x^2}$ grows with the square of the parameter count, both its evaluation and its inversion become very expensive for large systems.
Many algorithms therefore avoid computing the exact Hessian and instead approximate it by accumulating the gradient over multiple update steps.
The memory requirements also grow quadratically.
The quasi-Newton update above additionally requires the _inverse_ Hessian matrix. Thus, a Hessian that is close to being non-invertible typically causes numerical stability problems, while inherently non-invertible Hessians require a fallback to a first order GD update.
Another related limitation of quasi-Newton methods is that the objective function needs to be _twice-differentiable_.
While this may not seem like a big restriction, note that many common neural network architectures use ReLU activation functions of which the second-order derivative is zero.
%
Related to this is the problem that higher-order derivatives tend to change more quickly when traversing the parameter space, making them more prone to high-frequency noise in the loss landscape.
are still a very active research topic, and hence many extensions have been proposed that can alleviate some of these problems in certain settings. E.g., the memory requirement problem can be sidestepped by storing only lower-dimensional vectors that can be used to approximate the Hessian. However, these difficulties illustrate the problems that often arise when applying methods like BFGS.
%\nt{In contrast to these classic algorithms, we will show how to leverage invertible physical models to efficiently compute physical update steps. In certain scenarios, such as simple loss functions, computing the inverse gradient via the inverse Hessian will also provide a useful building block for our final algorithm.}
%, and how to they can be used to improve the training of neural networks.
Here, the Jacobian $\frac{\partial x}{\partial z}$, which is similar to the inverse of the GD update above, encodes how the inputs must change in order to obtain a small change $\Delta z$ in the output.
The crucial step is the inversion, which of course requires the Jacobian matrix to be invertible (a drawback we'll get back to below). However, if we can invert it, this has some very nice properties.
Note that instead of using a learning rate, here the step size is determined by the desired increase or decrease of the value of the output, $\Delta z$.
IGs scale with the inverse derivative. Hence the updates are automatically of the same units as the parameters without requiring an arbitrary learning rate: $\frac{\partial x}{\partial z}$ times $\Delta z$ has the units of $x$.
IGs show the opposite behavior of GD close to an optimum: they typically produce very accurate updates, which don't vanish near an optimum. This leads to fast convergence, as we will demonstrate in more detail below.
Additionally, IGs are consistent in function composition.
The change in $x$ is $\Delta x = \Delta L \cdot \frac{\partial x}{\partial z} \frac{\partial z}{\partial L}$ and the approximate change in $z$ is $\Delta z = \Delta L \cdot \frac{\partial z}{\partial x} \frac{\partial x}{\partial z} \frac{\partial z}{\partial L} = \Delta L \frac{\partial z}{\partial L}$.
% In the example in table~\ref{tab:function-composition-example}, the change $\Delta z$ is the same no matter what space is used as optimization target.
The change in intermediate spaces is independent of their respective dependencies, at least up to first order.
Consequently, the change to these spaces can be estimated during backpropagation, before all gradients have been computed.
Note that even Newton's method with its inverse Hessian didn't fully get this right. The key here is that if the Jacobian is invertible, we'll directly get the correctly scaled direction at a given layer, without "helpers" such as the inverse Hessian.
The above properties make the advantages of IGs clear, but we're not done, unfortunately. There are strong limitations to their applicability.
%
The IG $\frac{\partial x}{\partial z}$ is only well-defined for square Jacobians, i.e. for functions $z$ with the same inputs and output dimensions.
In optimization, however, the input is typically high-dimensional while the output is a scalar objective function.
%
And, somewhat similar to the Hessians of quasi-Newton methods,
even when the $\frac{\partial z}{\partial x}$ is square, it may not be invertible.
Thus, we now consider the fact that inverse gradients are linearizations of inverse functions and show that using inverse functions provides additional advantages while retaining the same benefits.
### Inverse simulators
Physical processes can be described as a trajectory in state space where each point represents one possible configuration of the system.
A simulator typically takes one such state space vector and computes a new one at another time.
The Jacobian of the simulator is, therefore, necessarily square.
%
As long as the physical process does _not destroy_ information, the Jacobian is non-singular.
In fact, it is believed that information in our universe cannot be destroyed so any physical process could in theory be inverted as long as we have perfect knowledge of the state.
While evaluating the IGs directly can be done through matrix inversion or taking the derivative of an inverse simulator, we now consider what happens if we use the inverse simulator directly in backpropagation.
Let $z = \mathcal P(x)$ be a forward simulation, and $\mathcal P(z)^{-1}=x$ its inverse (we assume it exists for now, but below we'll relax that assumption).
Equipped with the inverse we now define an update that we'll call the **physical gradient** (PG) in the following as
$ \Delta x / \Delta z = \frac{\partial x}{\partial z} + \mathcal O(\Delta z^2) $.
%
The accuracy of the update also depends on the fidelity of the inverse function $\mathcal P^{-1}$.
We can define an upper limit to the error of the local inverse using the local gradient $\frac{\partial x}{\partial z}$.
In the worst case, we can therefore fall back to the regular gradient.
% We now show that these terms can help produce more stable updates than the IG alone, provided that $\mathcal P_{(x_0,z_0)}^{-1}$ is a sufficiently good approximation of the true inverse.
% Let $\mathcal P^{-1}(z)$ be the true inverse function to $\mathcal P(x)$, assuming that $\mathcal P$ is fully invertible.
The intuition for why the PG update is a good one is that when
applying the update $\Delta x = \mathcal P^{-1}(z_0 + \Delta z) - x_0$ it will produce $\mathcal P(x_0 + \Delta x) = z_0 + \Delta z$ exactly, despite $\mathcal P$ being a potentially highly nonlinear function.
When rewriting this update in the typical gradient format, $\frac{\Delta x}{\Delta z}$ replaces the gradient from the IG update above, and gives $\Delta x$.
**Fundamental theorem of calculus**
To more clearly illustrate the advantages in non-linear settings, we
apply the fundamental theorem of calculus to rewrite the ratio $\Delta x / \Delta z$ from above. This gives,
Here the expressions inside the integral is the local gradient, and we assume it exists at all points between $z_0$ and $z_0+\Delta z_0$.
The local gradients are averaged along the path connecting the state before the update with the state after the update.
The whole expression is therefore equal to the average gradient of $\mathcal P$ between the current $x$ and the estimate for the next optimization step $x_0 + \Delta x$.
This effectively amounts to _smoothing the objective landscape_ of an optimization by computing updates that can take nonlinearities of $\mathcal P$ into account.
The equations naturally generalize to higher dimensions by replacing the integral with a path integral along any differentiable path connecting $x_0$ and $x_0 + \Delta x$ and replacing the local gradient by the local gradient in the direction of the path.
### Global and local inverse functions
Let $\mathcal P$ be a function with a square Jacobian and $z = \mathcal P(x)$.
A global inverse function $\mathcal P^{-1}$ is defined only for bijective $\mathcal P$.
If the inverse exists, it can find $x$ for any $z$ such that $z = \mathcal P(x)$.
Instead of using this "perfect" inverse $\mathcal P^{-1}$ directly, we'll in practice often use a local inverse
$\mathcal P_{(x_0,z_0)}^{-1}(z)$, defined at the point $(x_0, z_0)$. This local inverse can be
easier to obtain, as it only needs to exist near a given $z_0$, and not for all $z$.
For $\mathcal P^{-1}$ to exist $\mathcal P$ would need to be globally invertible.
By contrast, a \emph{local inverse}, defined at point $(x_0, z_0)$, only needs to be accurate in the vicinity of that point.
If a global inverse $\mathcal P^{-1}(z)$ exists, the local inverse approximates it and matches it exactly as $z \rightarrow z_0$.
The update obtained with a regular gradient descent method has surprising shortcomings.
The physical gradient instead allows us to more accurately backpropagate through nonlinear functions, provided that we have access to good inverse functions.
Before moving on to including PGs in NN training processes, the next example will illustrate the differences between these approaches with a practical example.
