Autoencoders vs. PCA: I Rigged the Test and PCA Still Won

Autoencoders vs. PCA: I Rigged the Test and PCA Still Won A theoretical advantage that didn't survive contact with a real benchmark. The theoretical case for autoencoders Autoencoders — neural networks trained to reconstruct their own input through a compressed bottleneck — are a standard recommendation for anomaly detection. The theoretical argument is clean: train the network only on normal data, and it learns to reconstruct normal patterns well. Feed it an anomaly, and reconstruction error spikes, because the network never learned to compress that kind of pattern. Unlike PCA, which can only capture linear relationships between features, an autoencoder can in principle learn nonlinear ones — so it should catch anomalies that violate a nonlinear structure in the data, which a linear method structurally cannot. That's a specific, testable claim, not a vague one: autoencoders should have a real, measurable edge over PCA specifically on anomalies that violate nonlinear relationships. So I built two experiments — one easy case, and one deliberately designed to be the autoencoder's best shot — and measured whether the theoretical advantage actually shows up. Setup, both experiments: synthetic data (clearly labeled as such — this is not a real sensor or fraud dataset), trained on normal samples only (the realistic anomaly-detection setup — you rarely have labeled anomalies to train on), evaluated on a held-out mix of normal and anomalous samples. Three methods compared: an autoencoder (MLPRegressor trained to reconstruct its own input, with a 3-dimensional bottleneck), PCA reconstruction error (also reduced to 3 components — same bottleneck size, for a fair comparison), and Isolation Forest as a non-reconstruction-based reference point. Experiment 1: the easy case Normal data drawn from a mixture of Gaussian clusters (representing, say, a few normal operating regimes of a machine). Anomalies drawn from a distribution with a shifted mean and higher variance — a straightforward, linearly-separable kind of outlier. Autoencoder and PCA tied exactly — 0.885 F1, both catching every single anomaly (recall = 1.0), differing only slightly on precision. Isolation Forest came in just behind at 0.870. This result alone isn't surprising once you think about why: a mean shift is a linear phenomenon, so a linear method has no structural disadvantage detecting it. The autoencoder's extra representational capacity was simply unnecessary here. Here's why it was so easy — the actual reconstruction error distribution the autoencoder produced on the test set: In Experiment 1 (left panel), normal and anomalous reconstruction errors barely overlap at all — normal samples cluster tightly under 1.0, anomalies sit almost entirely above 4.0. Any reasonable threshold in that gap catches everything. That's what "easy" looks like in reconstruction-error terms, and it explains why a linear method does just as well as a nonlinear one: the separation is large enough that neither method's precision matters much. Experiment 2: rigging the test in the autoencoder's favor This is the part that actually tests the theoretical claim. I built anomalies specifically designed to be invisible to a linear method: normal data where two features follow a nonlinear relationship (y = sin(3x) + noise ), and anomalies that keep the same individual range for each feature but violate the relationship between them — x and y each look perfectly normal in isolation; only their joint, nonlinear relationship is wrong. This is close to a best-case scenario for an autoencoder's theoretical advantage: a pattern a linear projection genuinely cannot represent, by construction. The autoencoder still didn't win. PCA scored 0.318 F1, the autoencoder scored 0.302 — PCA very slightly ahead, both far weaker than Experiment 1 (which makes sense — this is a genuinely harder detection problem for any method) but with no autoencoder advantage anywhere in sight. Isolation Forest fell apart entirely on this task, at 0.091. The right panel of the histogram above shows why this one was hard for everyone: normal and anomalous reconstruction errors overlap heavily, with no clean gap to threshold on. Some anomalies produced lower reconstruction error than plenty of normal samples — meaning no fixed threshold, on either method's error signal, could have separated them cleanly. That's a materially different failure mode than "the wrong method was used" — it's "the detection signal itself didn't separate the classes well," which is a data and modeling-choice problem, not simply a which-algorithm problem. Why the theoretical advantage didn't show up This isn't evidence that autoencoders can't outperform PCA — it's evidence that an untuned, default-architecture autoencoder doesn't automatically realize its theoretical advantage, and that gap between theory and default practice is the actual finding worth taking seriously. A few concrete reasons this likely happened: 700 training samples is not much data for a neural network to learn a nonlinear manifold from scratch. PCA's linear solution has a closed-form optimum computable from a handful of samples; the autoencoder has to find a good nonlinear solution by gradient descent, which needs meaningfully more data to do reliably. A single default architecture (8-3-8, max_iter=2000 ) is a starting guess, not a tuned model. Capturing a specific nonlinear relationship well often requires deliberately shaping the architecture around the kind of nonlinearity expected — different depth, width, activation function, or training duration — none of which I searched over here.Reconstruction-error anomaly detection has a structural limitation that hits both methods: when the anomaly signal is concentrated in a subset of features and diluted by averaging across all of them, both linear and nonlinear reconstruction error can miss it. I saw this directly in an earlier version of this experiment with more noise dimensions, where both methods collapsed to near-random performance — a separate, useful lesson about reconstruction-based detection in high-dimensional settings. What would actually be needed to unlock the autoencoder's advantage here? A few concrete, testable next steps, in rough order of how cheap they are to try: increase training data volume substantially (the nonlinear relationship needs enough examples to be learnable, not just theoretically learnable); widen or deepen the architecture specifically around the 2 features carrying the signal rather than a generic 8-3-8 shape; inspect the learned bottleneck representation directly (plot the 3 bottleneck activations, colored by true label) to see whether the anomalies are even separable in that latent space, which would tell you whether the problem is representation or thresholding; and consider a feature-weighted reconstruction error, so the two informative features aren't averaged down by uninformative ones. None of these are exotic — they're the actual engineering work "just add an autoencoder" skips over. The actual decision framework Before reaching for an autoencoder over a simpler reconstruction-based method like PCA, three questions are worth answering first: Does your data have genuinely nonlinear relationships between features, or does it just feel like it should? "Complex-sounding data" and "data with nonlinear structure a linear method can't capture" are not the same thing — verify the second one specifically before assuming it justifies the extra model complexity. Do you have enough normal-only training data for a neural network to actually learn that structure? PCA's linear solution is nearly data-efficient by construction; a neural network's nonlinear solution generally isn't. A few hundred samples might be plenty for PCA and not nearly enough for a network to find real signal instead of noise. Have you actually looked at the reconstruction error distribution, for either method, before trusting either one's threshold? A histogram like the ones above takes one line of code and tells you immediately whether you're dealing with a clean separation problem (where the choice of method barely matters) or a genuine overlap problem (where neither method's threshold will save you without more fundamental changes). Where this comparison falls short Synthetic data, deliberately constructed — both experiments use data I generated specifically to test a hypothesis, not real sensor or fraud data. The qualitative lesson (default architectures don't automatically deliver their theoretical advantage) is more likely to generalize than the exact numbers. One architecture, one training run. I didn't search over autoencoder depth, width, or training duration — which is precisely the point (this article tests the "just use an autoencoder" default, not the ceiling of what a well-tuned one can do), but it means these results describe the default, not the best case. A single random seed for the anomaly generation. Different synthetic anomaly constructions could shift these specific numbers; the direction — no autoencoder advantage materializing by default — is the more robust part of the finding. Conclusion "Autoencoders can model nonlinear relationships that PCA can't" is true as a statement about representational capacity. It is not the same claim as "an autoencoder will outperform PCA on your anomaly detection task by default" — and conflating the two is where the practical disappointment comes from. Realizing a neural network's theoretical advantage over a simpler linear method takes real tuning effort, real data volume, and real architecture decisions; none of that comes for free just from choosing the more powerful model class. Before reaching for the autoencoder, it's worth asking the same question that applies to every "fancier method" decision: does the improvement show up when you actually measure it, or only when you assume it should?