This put up didn’t find yourself fairly the way in which I’d imagined. A fast follow-up on the current Time sequence prediction with FNN-LSTM, it was imagined to reveal how *noisy* time sequence (so frequent in follow) might revenue from a change in structure: As an alternative of FNN-LSTM, an LSTM autoencoder regularized by false nearest neighbors (FNN) loss, use FNN-VAE, a variational autoencoder constrained by the identical. Nevertheless, FNN-VAE didn’t appear to deal with noise higher than FNN-LSTM. No plot, no put up, then?

Then again – this isn’t a scientific examine, with speculation and experimental setup all preregistered; all that actually issues is that if there’s one thing helpful to report. And it seems like there’s.

Firstly, FNN-VAE, whereas on par performance-wise with FNN-LSTM, is way superior in that different that means of “efficiency”: Coaching goes a *lot* quicker for FNN-VAE.

Secondly, whereas we don’t see a lot distinction between FNN-LSTM and FNN-VAE, we *do* see a transparent influence of utilizing FNN loss. Including in FNN loss strongly reduces imply squared error with respect to the underlying (denoised) sequence – particularly within the case of VAE, however for LSTM as effectively. That is of explicit curiosity with VAE, because it comes with a regularizer out-of-the-box – specifically, Kullback-Leibler (KL) divergence.

In fact, we don’t declare that related outcomes will at all times be obtained on different noisy sequence; nor did we tune any of the fashions “to loss of life.” For what might be the intent of such a put up however to indicate our readers attention-grabbing (and promising) concepts to pursue in their very own experimentation?

## The context

This put up is the third in a mini-series.

In Deep attractors: The place deep studying meets chaos, we defined, with a considerable detour into chaos principle, the thought of FNN loss, launched in (Gilpin 2020). Please seek the advice of that first put up for theoretical background and intuitions behind the approach.

The next put up, Time sequence prediction with FNN-LSTM, confirmed learn how to use an LSTM autoencoder, constrained by FNN loss, for forecasting (versus reconstructing an attractor). The outcomes have been beautiful: In multi-step prediction (12-120 steps, with that quantity various by dataset), the short-term forecasts have been drastically improved by including in FNN regularization. See that second put up for experimental setup and outcomes on 4 very totally different, non-synthetic datasets.

At present, we present learn how to change the LSTM autoencoder by a – convolutional – VAE. In gentle of the experimentation outcomes, already hinted at above, it’s fully believable that the “variational” half just isn’t even so essential right here – {that a} convolutional autoencoder with simply MSE loss would have carried out simply as effectively on these knowledge. In actual fact, to search out out, it’s sufficient to take away the decision to `reparameterize()`

and multiply the KL part of the loss by 0. (We depart this to the reader, to maintain the put up at cheap size.)

One final piece of context, in case you haven’t learn the 2 earlier posts and want to bounce in right here instantly. We’re doing time sequence forecasting; so why this discuss of autoencoders? Shouldn’t we simply be evaluating an LSTM (or another sort of RNN, for that matter) to a convnet? In actual fact, the need of a latent illustration is because of the very thought of FNN: The latent code is meant to mirror the true attractor of a dynamical system. That’s, if the attractor of the underlying system is roughly two-dimensional, we hope to search out that simply two of the latent variables have appreciable variance. (This reasoning is defined in numerous element within the earlier posts.)

## FNN-VAE

So, let’s begin with the code for our new mannequin.

The encoder takes the time sequence, of format `batch_size x num_timesteps x num_features`

similar to within the LSTM case, and produces a flat, 10-dimensional output: the latent code, which FNN loss is computed on.

```
library(tensorflow)
library(keras)
library(tfdatasets)
library(tfautograph)
library(reticulate)
library(purrr)
vae_encoder_model <- perform(n_timesteps,
n_features,
n_latent,
identify = NULL) {
keras_model_custom(identify = identify, perform(self) {
self$conv1 <- layer_conv_1d(kernel_size = 3,
filters = 16,
strides = 2)
self$act1 <- layer_activation_leaky_relu()
self$batchnorm1 <- layer_batch_normalization()
self$conv2 <- layer_conv_1d(kernel_size = 7,
filters = 32,
strides = 2)
self$act2 <- layer_activation_leaky_relu()
self$batchnorm2 <- layer_batch_normalization()
self$conv3 <- layer_conv_1d(kernel_size = 9,
filters = 64,
strides = 2)
self$act3 <- layer_activation_leaky_relu()
self$batchnorm3 <- layer_batch_normalization()
self$conv4 <- layer_conv_1d(
kernel_size = 9,
filters = n_latent,
strides = 2,
activation = "linear"
)
self$batchnorm4 <- layer_batch_normalization()
self$flat <- layer_flatten()
perform (x, masks = NULL) {
x %>%
self$conv1() %>%
self$act1() %>%
self$batchnorm1() %>%
self$conv2() %>%
self$act2() %>%
self$batchnorm2() %>%
self$conv3() %>%
self$act3() %>%
self$batchnorm3() %>%
self$conv4() %>%
self$batchnorm4() %>%
self$flat()
}
})
}
```

The decoder begins from this – flat – illustration and decompresses it right into a time sequence. In each encoder and decoder (de-)conv layers, parameters are chosen to deal with a sequence size (`num_timesteps`

) of 120, which is what we’ll use for prediction beneath.

```
vae_decoder_model <- perform(n_timesteps,
n_features,
n_latent,
identify = NULL) {
keras_model_custom(identify = identify, perform(self) {
self$reshape <- layer_reshape(target_shape = c(1, n_latent))
self$conv1 <- layer_conv_1d_transpose(kernel_size = 15,
filters = 64,
strides = 3)
self$act1 <- layer_activation_leaky_relu()
self$batchnorm1 <- layer_batch_normalization()
self$conv2 <- layer_conv_1d_transpose(kernel_size = 11,
filters = 32,
strides = 3)
self$act2 <- layer_activation_leaky_relu()
self$batchnorm2 <- layer_batch_normalization()
self$conv3 <- layer_conv_1d_transpose(
kernel_size = 9,
filters = 16,
strides = 2,
output_padding = 1
)
self$act3 <- layer_activation_leaky_relu()
self$batchnorm3 <- layer_batch_normalization()
self$conv4 <- layer_conv_1d_transpose(
kernel_size = 7,
filters = 1,
strides = 1,
activation = "linear"
)
self$batchnorm4 <- layer_batch_normalization()
perform (x, masks = NULL) {
x %>%
self$reshape() %>%
self$conv1() %>%
self$act1() %>%
self$batchnorm1() %>%
self$conv2() %>%
self$act2() %>%
self$batchnorm2() %>%
self$conv3() %>%
self$act3() %>%
self$batchnorm3() %>%
self$conv4() %>%
self$batchnorm4()
}
})
}
```

Word that although we referred to as these constructors `vae_encoder_model()`

and `vae_decoder_model()`

, there’s nothing variational to those fashions per se; they’re actually simply an encoder and a decoder, respectively. Metamorphosis right into a VAE will occur within the coaching process; in truth, the one two issues that can make this a VAE are going to be the reparameterization of the latent layer and the added-in KL loss.

Talking of coaching, these are the routines we’ll name. The perform to compute FNN loss, `loss_false_nn()`

, could be present in each of the abovementioned predecessor posts; we kindly ask the reader to repeat it from considered one of these locations.

```
# to reparameterize encoder output earlier than calling decoder
reparameterize <- perform(imply, logvar = 0) {
eps <- k_random_normal(form = n_latent)
eps * k_exp(logvar * 0.5) + imply
}
# loss has 3 parts: NLL, KL, and FNN
# in any other case, that is simply regular TF2-style coaching
train_step_vae <- perform(batch) {
with (tf$GradientTape(persistent = TRUE) %as% tape, {
code <- encoder(batch[[1]])
z <- reparameterize(code)
prediction <- decoder(z)
l_mse <- mse_loss(batch[[2]], prediction)
# see loss_false_nn in 2 earlier posts
l_fnn <- loss_false_nn(code)
# KL divergence to an ordinary regular
l_kl <- -0.5 * k_mean(1 - k_square(z))
# general loss is a weighted sum of all 3 parts
loss <- l_mse + fnn_weight * l_fnn + kl_weight * l_kl
})
encoder_gradients <-
tape$gradient(loss, encoder$trainable_variables)
decoder_gradients <-
tape$gradient(loss, decoder$trainable_variables)
optimizer$apply_gradients(purrr::transpose(checklist(
encoder_gradients, encoder$trainable_variables
)))
optimizer$apply_gradients(purrr::transpose(checklist(
decoder_gradients, decoder$trainable_variables
)))
train_loss(loss)
train_mse(l_mse)
train_fnn(l_fnn)
train_kl(l_kl)
}
# wrap all of it in autograph
training_loop_vae <- tf_function(autograph(perform(ds_train) {
for (batch in ds_train) {
train_step_vae(batch)
}
tf$print("Loss: ", train_loss$consequence())
tf$print("MSE: ", train_mse$consequence())
tf$print("FNN loss: ", train_fnn$consequence())
tf$print("KL loss: ", train_kl$consequence())
train_loss$reset_states()
train_mse$reset_states()
train_fnn$reset_states()
train_kl$reset_states()
}))
```

To complete up the mannequin part, right here is the precise coaching code. That is almost an identical to what we did for FNN-LSTM earlier than.

```
n_latent <- 10L
n_features <- 1
encoder <- vae_encoder_model(n_timesteps,
n_features,
n_latent)
decoder <- vae_decoder_model(n_timesteps,
n_features,
n_latent)
mse_loss <-
tf$keras$losses$MeanSquaredError(discount = tf$keras$losses$Discount$SUM)
train_loss <- tf$keras$metrics$Imply(identify = 'train_loss')
train_fnn <- tf$keras$metrics$Imply(identify = 'train_fnn')
train_mse <- tf$keras$metrics$Imply(identify = 'train_mse')
train_kl <- tf$keras$metrics$Imply(identify = 'train_kl')
fnn_multiplier <- 1 # default worth utilized in almost all instances (see textual content)
fnn_weight <- fnn_multiplier * nrow(x_train)/batch_size
kl_weight <- 1
optimizer <- optimizer_adam(lr = 1e-3)
for (epoch in 1:100) {
cat("Epoch: ", epoch, " -----------n")
training_loop_vae(ds_train)
test_batch <- as_iterator(ds_test) %>% iter_next()
encoded <- encoder(test_batch[[1]][1:1000])
test_var <- tf$math$reduce_variance(encoded, axis = 0L)
print(test_var %>% as.numeric() %>% spherical(5))
}
```

## Experimental setup and knowledge

The concept was so as to add white noise to a deterministic sequence. This time, the Roessler system was chosen, primarily for the prettiness of its attractor, obvious even in its two-dimensional projections:

Like we did for the Lorenz system within the first a part of this sequence, we use `deSolve`

to generate knowledge from the Roessler equations.

```
library(deSolve)
parameters <- c(a = .2,
b = .2,
c = 5.7)
initial_state <-
c(x = 1,
y = 1,
z = 1.05)
roessler <- perform(t, state, parameters) {
with(as.checklist(c(state, parameters)), {
dx <- -y - z
dy <- x + a * y
dz = b + z * (x - c)
checklist(c(dx, dy, dz))
})
}
instances <- seq(0, 2500, size.out = 20000)
roessler_ts <-
ode(
y = initial_state,
instances = instances,
func = roessler,
parms = parameters,
technique = "lsoda"
) %>% unclass() %>% as_tibble()
n <- 10000
roessler <- roessler_ts$x[1:n]
roessler <- scale(roessler)
```

Then, noise is added, to the specified diploma, by drawing from a standard distribution, centered at zero, with customary deviations various between 1 and a couple of.5.

```
# add noise
noise <- 1 # additionally used 1.5, 2, 2.5
roessler <- roessler + rnorm(10000, imply = 0, sd = noise)
```

Right here you may examine results of not including any noise (left), customary deviation-1 (center), and customary deviation-2.5 Gaussian noise:

In any other case, preprocessing proceeds as within the earlier posts. Within the upcoming outcomes part, we’ll examine forecasts not simply to the “actual,” after noise addition, take a look at break up of the info, but additionally to the underlying Roessler system – that’s, the factor we’re actually excited by. (Simply that in the true world, we will’t do this test.) This second take a look at set is ready for forecasting similar to the opposite one; to keep away from duplication we don’t reproduce the code.

```
n_timesteps <- 120
batch_size <- 32
gen_timesteps <- perform(x, n_timesteps) {
do.name(rbind,
purrr::map(seq_along(x),
perform(i) {
begin <- i
finish <- i + n_timesteps - 1
out <- x[start:end]
out
})
) %>%
na.omit()
}
practice <- gen_timesteps(roessler[1:(n/2)], 2 * n_timesteps)
take a look at <- gen_timesteps(roessler[(n/2):n], 2 * n_timesteps)
dim(practice) <- c(dim(practice), 1)
dim(take a look at) <- c(dim(take a look at), 1)
x_train <- practice[ , 1:n_timesteps, , drop = FALSE]
y_train <- practice[ , (n_timesteps + 1):(2*n_timesteps), , drop = FALSE]
ds_train <- tensor_slices_dataset(checklist(x_train, y_train)) %>%
dataset_shuffle(nrow(x_train)) %>%
dataset_batch(batch_size)
x_test <- take a look at[ , 1:n_timesteps, , drop = FALSE]
y_test <- take a look at[ , (n_timesteps + 1):(2*n_timesteps), , drop = FALSE]
ds_test <- tensor_slices_dataset(checklist(x_test, y_test)) %>%
dataset_batch(nrow(x_test))
```

## Outcomes

The LSTM used for comparability with the VAE described above is an identical to the structure employed within the earlier put up. Whereas with the VAE, an `fnn_multiplier`

of 1 yielded adequate regularization for all noise ranges, some extra experimentation was wanted for the LSTM: At noise ranges 2 and a couple of.5, that multiplier was set to five.

Because of this, in all instances, there was one latent variable with excessive variance and a second considered one of minor significance. For all others, variance was near 0.

*In all instances* right here means: In all instances the place FNN regularization was used. As already hinted at within the introduction, the primary regularizing issue offering robustness to noise right here appears to be FNN loss, not KL divergence. So for all noise ranges, apart from FNN-regularized LSTM and VAE fashions we additionally examined their non-constrained counterparts.

#### Low noise

Seeing how all fashions did fantastically on the unique deterministic sequence, a noise degree of 1 can virtually be handled as a baseline. Right here you see sixteen 120-timestep predictions from each regularized fashions, FNN-VAE (darkish blue), and FNN-LSTM (orange). The noisy take a look at knowledge, each enter (`x`

, 120 steps) and output (`y`

, 120 steps) are displayed in (blue-ish) gray. In inexperienced, additionally spanning the entire sequence, now we have the unique Roessler knowledge, the way in which they might look had no noise been added.

Regardless of the noise, forecasts from each fashions look glorious. Is that this because of the FNN regularizer?

Taking a look at forecasts from their unregularized counterparts, now we have to confess these don’t look any worse. (For higher comparability, the sixteen sequences to forecast have been initiallly picked at random, however used to check all fashions and situations.)

What occurs once we begin to add noise?

#### Substantial noise

Between noise ranges 1.5 and a couple of, one thing modified, or turned noticeable from visible inspection. Let’s bounce on to the highest-used degree although: 2.5.

Right here first are predictions obtained from the unregularized fashions.

Each LSTM and VAE get “distracted” a bit an excessive amount of by the noise, the latter to a good larger diploma. This results in instances the place predictions strongly “overshoot” the underlying non-noisy rhythm. This isn’t stunning, after all: They have been *skilled* on the noisy model; predict fluctuations is what they discovered.

Will we see the identical with the FNN fashions?

Curiously, we see a a lot better match to the underlying Roessler system now! Particularly the VAE mannequin, FNN-VAE, surprises with a complete new smoothness of predictions; however FNN-LSTM turns up a lot smoother forecasts as effectively.

“Clean, becoming the system…” – by now it’s possible you’ll be questioning, when are we going to give you extra quantitative assertions? If quantitative implies “imply squared error” (MSE), and if MSE is taken to be some divergence between forecasts and the true goal from the take a look at set, the reply is that this MSE doesn’t differ a lot between any of the 4 architectures. Put in a different way, it’s largely a perform of noise degree.

Nevertheless, we might argue that what we’re actually excited by is how effectively a mannequin forecasts the underlying course of. And there, we see variations.

Within the following plot, we distinction MSEs obtained for the 4 mannequin varieties (gray: VAE; orange: LSTM; darkish blue: FNN-VAE; inexperienced: FNN-LSTM). The rows mirror noise ranges (1, 1.5, 2, 2.5); the columns characterize MSE in relation to the noisy(“actual”) goal (left) on the one hand, and in relation to the underlying system on the opposite (proper). For higher visibility of the impact, *MSEs have been normalized as fractions of the utmost MSE in a class*.

So, if we wish to predict *sign plus noise* (left), it’s not extraordinarily important whether or not we use FNN or not. But when we wish to predict the sign solely (proper), with growing noise within the knowledge FNN loss turns into more and more efficient. This impact is way stronger for VAE vs. FNN-VAE than for LSTM vs. FNN-LSTM: The gap between the gray line (VAE) and the darkish blue one (FNN-VAE) turns into bigger and bigger as we add extra noise.

## Summing up

Our experiments present that when noise is more likely to obscure measurements from an underlying deterministic system, FNN regularization can strongly enhance forecasts. That is the case particularly for convolutional VAEs, and doubtless convolutional autoencoders normally. And if an FNN-constrained VAE performs as effectively, for time sequence prediction, as an LSTM, there’s a robust incentive to make use of the convolutional mannequin: It trains considerably quicker.

With that, we conclude our mini-series on FNN-regularized fashions. As at all times, we’d love to listen to from you should you have been capable of make use of this in your personal work!

Thanks for studying!