PaperlayerView exact revision on arXiv

Source-linked AI summary

Visualizing Deep Convolutional Neural Networks Using Natural Pre-Images

Aravindh Mahendran, Andrea Vedaldi

arXiv:1512.02017v3cs.CV

TL;DR

Because the design and behavior of image representations remain difficult to understand, the paper studies them through natural-looking pre-images. It unifies inversion, activation maximization, and caricaturization in a regularized energy-minimization framework, finding that CNN layers preserve photographically accurate information while developing different geometric and photometric invariances.

  • Problem

    The design of shallow features and deep CNN representations remains largely empirical, limiting understanding of their behavior.

  • Method

    The paper uses natural pre-images and a unified regularized optimization framework to perform inversion, activation maximization, and caricaturization across image representations.

  • Results

    CNN layers retain photometrically accurate and instance-specific information while progressively building invariance and complexity across depth.

  • Takeaways & Limitations

    Naturalness priors make reconstructed representations more interpretable, and visualizations expose distinct structures and invariances captured by different CNN layers.

  • Takeaways & Limitations

    The formulation assumes the image has null mean, as required by most CNN implementations.

Abstract

from arXiv · show

Image representations, from SIFT and bag of visual words to Convolutional Neural Networks (CNNs) are a crucial component of almost all computer vision systems. However, our understanding of them remains limited. In this paper we study several landmark representations, both shallow and deep, by a number of complementary visualization techniques. These visualizations are based on the concept of "natural pre-image", namely a natural-looking image whose representation has some notable property. We study in particular three such visualizations: inversion, in which the aim is to reconstruct an image from its representation, activation maximization, in which we search for patterns that maximally stimulate a representation component, and caricaturization, in which the visual patterns that a representation detects in an image are exaggerated. We pose these as a regularized energy-minimization framework and demonstrate its generality and effectiveness. In particular, we show that this method can invert representations such as HOG more accurately than recent alternatives while being applicable to CNNs too. Among our findings, we show that several layers in CNNs retain photographically accurate information about the image, with different degrees of geometric and photometric invariance.

1 Introduction

The paper develops natural pre-images as a unified way to visualize and understand shallow and deep image representations. It applies inversion, activation maximization, and caricaturization across representations including HOG and CNNs.

  • Limited understanding of empirically designed image representations motivates visualization methods for shallow features and deep CNNs.
  • Natural pre-images are natural-looking images whose representations have a useful notable property.
  • Inversion reconstructs an image from its representation code, while activation maximization searches for an image that maximally excites one representation component.
  • Caricaturization modifies an image to exaggerate patterns that excite its representation, emphasizing combinations of jointly active components.
  • The unified algorithm optimizes an energy function with gradient descent and backpropagation through the representation architecture.
  • Across HOG and CNNs, the method improves shallow-representation reconstructions and reveals increasing invariance and complexity across CNN layers.

2 Related work

Prior work used pre-images, activation maximization, inverse networks, and transposed CNNs to visualize representations. This paper distinguishes its unified energy-minimization approach from these alternatives while extending pre-image visualization to multiple purposes.

  • Natural-image priors and regularized energy minimization were introduced to make pre-images more meaningful for discriminative representations.
  • Jittering, Gaussian blurring, and pixel clipping were explored as complementary regularizers for sharper or more vivid visualizations.
  • Alternative approaches include genetic programming, learned regressors, patch vocabularies, transposed CNNs, and networks trained jointly with inverses.
  • Earlier studies inverted SIFT, local binary descriptors, HOG, and bag-of-visual-words representations to recover images from encodings.
  • Activation maximization visualizes component responses, while caricaturization exaggerates feature patterns and has also been used for image generation and style-related applications.
  • The pre-image framework can support additional visualization and image-generation tasks, including texture synthesis and style transfer.

3 A method for finding the pre-images of a representation

The method finds a natural-looking image whose representation matches a target code by minimizing a regularized objective. The objective balances representation loss against two image-prior regularizers using a single trade-off constant.

  • Given a representation function Φ and target code Φ0, the method searches for an image x that minimizes a regularized energy.
  • The objective combines representation loss with Rα and TV β regularizers that encode a natural-image prior.
  • The constant C trades off representation-matching loss against the regularizers.
  • Although the formulation contains several parameters, the method is designed so that parameters other than C can usually remain fixed across visualization and representation types.

3.1 Loss functions

Different visualization goals are implemented by changing the representation loss, while masks and priors control which components are emphasized and how natural the reconstruction appears.

  • In inversion, Euclidean representation loss produces an image that resembles a target image from the representation’s viewpoint.
  • A binary mask can restrict inversion to selected representation components.
  • The image is assumed to have null mean, as required by most CNN implementations.
  • Using an indicator vector Φ0 = ei makes minimizing the inner-product loss maximize component [Φ(x)]i.
  • Setting Φ0 to max{Φ(x0), 0} highlights components active for a reference image while ignoring inactive components.

3.2 Regularization

The framework restricts reconstructions toward natural-looking images using pixel, total-variation, and jitter regularization. These priors recover discarded low-level information while controlling saturation, artifacts, and blur.

  • Natural-image priors: Natural-image regularization recovers information discarded by discriminative representations while focusing inversion on images relevant to visualization.The paper uses simple image priors as proxies for the difficult task of modeling the full set of natural images.
  • Pixel intensity: Pixel-intensity regularization bounds image values, preventing arbitrary rescaling from leaving network outputs unchanged and limiting activation growth.A hard constraint caps pixel intensity at B+, while the soft constraint prevents saturation that occurs with hard constraints alone.
  • Total variation: The total-variation regularizer encourages piece-wise constant patches by penalizing image gradients through finite differences.Its normalization uses the typical gradient-norm value V.
  • Total variation: Choosing β > 1 removes TV-induced spikes by distributing intensity changes across regions, trading artifact removal for blurrier reconstructions.The spikes disappear for β = 1.5, 2, but the resulting image is blurrier than for β = 1.
  • Jitter: Jittering averages objectives over random translations and crops, counterbalancing CNN downsampling and generally producing crisper pre-images.It is especially beneficial for activation maximization; the same framework can also support texture synthesis and style transfer.

3.3 Balancing the loss and the regularizers

The loss and regularizer terms are normalized so that reasonable reconstructions give them comparable values, simplifying parameter balancing across representation types.

  • Balancing terms: Normalizing losses and regularizers to values around unity substantially reduces the need to tune parameters separately for different representations.The paper describes this normalization as making a very significant practical difference.

3.4 Optimization

The pre-image objective is optimized with a stochastic adaptive-gradient method whose computational cost is dominated by differentiating the representation. The procedure uses projected updates and fixed iteration schedules.

  • Optimization method: Gradient descent is effective despite the strong nonlinearity of deep representations, including compositions of multiple nonlinear layers.The implementation uses a variant of gradient descent for reconstruction.
  • Optimization method: The algorithm adapts each component’s learning rate using accumulated squared gradients and short-window momentum, then projects reconstructions into the feasible pixel range.It combines AdaGrad-like scaling with AdaDelta-like momentum using ρ = 0.9.
  • Optimizer choice: For inversion, L-BFGS-B showed no significant benefit over stochastic gradient methods, so the paper uses (S)GD-based optimization for simplicity.L-BFGS-B was better for texture generation, indicating that optimizer choice depends on the objective.
  • Iteration schedule: The default schedule runs 300 iterations, with jittered inversion optionally followed by 50 lower-learning-rate iterations that slightly improve reconstruction error.This fine-tuning is not applied to caricaturization or activation maximization.
  • Computational cost: Back-propagating representation derivatives dominates runtime, making visualization faster for shallow representations and slower for deeper ones.The algorithm can use GPU code through the underlying CNN framework; 300–350 iterations is a conservative general choice.

4 Representations

The paper studies dense SIFT, HOG, and several CNNs, reformulating classical descriptors as CNNs to simplify derivative computation. Classical pipelines combine gradient processing, pooling, normalization, and clamping.

  • Classical representations: DSIFT and HOG are implemented as exact CNN equivalents, enabling derivative computation within standard CNN experimentation frameworks.The equivalents are numerically indistinguishable from the VLFeat reference implementations.
  • Classical representations: SIFT and HOG compute image gradients, bin them by orientation, pool them into cell histograms, group cells into blocks, and normalize the blocks.SIFT uses K = 8 orientations, whereas HOG uses K = 18.
  • Classical representations: Directional filtering projects each gradient onto orientation directions, followed by nonlinear assignment into histogram elements.DSIFT uses bilinear orientation assignment, while the DPM V5 HOG variant uses hard assignment.
  • Classical representations: A modified ReLU with a norm-dependent offset approximates bilinear orientation assignment using standard CNN components.The offset is a∥g∥, with a = cos 2π/K.
  • Classical representations: Pooling and block extraction use linear-filter banks, while block l2 normalization corresponds to a special case of local response normalization.HOG additionally decomposes normalized blocks back into cells before clamping descriptor values.
  • Deep CNN representations: The CNN models include AlexNet, VGG-M, and VGG-VD-16, with VGG-VD-16 using narrow, densely sampled filters and many intermediate convolutional layers.VGG-VD-16 has no normalization layers and is described as slow but powerful.

5 Visualization by inversion

The inversion visualization reconstructs natural-looking images whose representations match a reference code, using feature-space fidelity and image naturalness to assess results. Across HOG, DSIFT, and CNN-related comparisons, reconstruction quality depends strongly on representation design and regularization.

  • Inversion objective: Inversion minimizes the L2 distance between a reconstructed image representation and a reference representation, beginning from random noise so only the code guides reconstruction.Different random initializations can reveal representation invariances because equally good reconstructions may differ in nuisance factors.
  • Evaluation: Reconstruction is evaluated primarily in feature space through normalized reconstruction error, rather than perceptual similarity to the reference image.Human assessment is additionally used to test whether regularization improves visualization interpretability.
  • 5.1.2 Reconstruction quality: Table 2 reports average reconstruction errors for inversion methods applied to HOG and DSIFT, with error bars showing 95% confidence intervals for the mean.HOGb denotes HOG with bilinear orientation assignments.
  • 5.1.2 Reconstruction quality: 60.1% vs. 36.6% reconstruction error: HOGgle is faster but less accurate than the proposed inversion method.The comparison uses average reconstruction error on the first 100 ILSVRC 2012 validation images; GPU implementation provides a ten-fold speed-up.
  • 5.1.3 Representation comparison: 36.6% down to 11.5% reconstruction error: bilinear orientation assignments substantially improve HOG inversion quality.DSIFT achieves quantitatively similar error while producing more detailed images, attributed to HOG’s stronger normalization discarding more visual information.

5.2 Inverting CNNs

CNN representations can be inverted with reconstruction quality comparable to HOG, while regularization trades reconstruction fidelity for interpretability differently across layers. Reconstructions reveal progressively greater invariance and information loss in deeper layers, though some instance-specific details persist.

  • Reconstruction quality: CNNs are not much harder to invert than HOG; with sufficiently large C, reconstruction error remains between 10–20% even for deepest layers.The authors attribute this to relatively benign CNN nonlinearities.
  • Reconstruction quality: Strong regularization substantially harms reconstructions from earlier layers but has little effect on deeper layers, where it improves pre-image interpretability.The recommended C decreases with depth: 300 in earlier layers, 1 in the deepest layers for AlexNet and VGG-M.
  • Reconstruction quality: For C values other than 1, reconstructions from different random initializations have reconstruction-error standard deviations around 0.02 or less.This indicates that most pre-images are similarly good by reconstruction error, including across many deep layers.
  • Reconstruction quality: Classification consistency follows reconstruction quality: shallow layers favor weak regularization, whereas deep layers tolerate stronger regularization and can remain class-consistent despite visually odd inversions.Classification consistency measures whether a reconstruction receives the same CNN class label as its reference image.
  • Inversion across layers: Early CNN layers are nearly exactly invertible, intermediate layers preserve instance details with increasing fuzziness, and fully connected layers discard more geometric and instance-specific information.Even fc8 retains some instance-specific details despite its role as a category predictor.
  • Inversion across architectures: VGG-M produces sharper reconstructions than AlexNet, while VGG-VD achieves more gradual abstraction and preserves fine details better in deep layers.VGG-VD deep reconstructions also tend to emphasize larger, more detailed object occurrences.
  • Ambiguity and invariance: Deep-layer codes can preserve texture statistics while allowing substantial changes in object pose, spatial arrangement, scale, and instance-specific details.For some images, multiple reconstructions contain scrambled or differently arranged parts while remaining similar to the original from the CNN’s viewpoint.
  • Channel subsets: CaffeRef channel subsets separately reveal sensitivity to color versus sharper edges and luminance components.The distinction arises from independently trained filter blocks in the architecture’s early layers.

6 Visualization by activation maximization

Activation maximization visualizes representation components by optimizing images that stimulate them, revealing differences between classical HOG templates and CNN layers. The resulting pre-images expose component selectivity, architectural effects, and representation invariances.

  • Classical representations: Activation maximization visualizes HOG templates by maximizing detector scores, unlike HOGgle, which reconstructs an approximate pre-image of a rectified template.Because HOG templates contain negative components and are not generally valid HOG descriptors, activation maximization can reflect both positive and negative template effects.
  • Classical representations: Activation maximization reconstructs finer details than HOGgle, while HOGgle produces stronger, straighter edges.The difference may reflect HOGgle’s natural-image dictionary versus the more generic smoothness prior used here.
  • CNN representations: CNN activation-maximization patterns become more complex with depth, progressing from colored edges and blobs in conv1 to intricate patterns in deeper convolutional layers.In VGG-VD, complexity builds more gradually than in VGG-M and AlexNet, with deeper layers producing especially complex stimuli.
  • CNN representations: Class-specific fc8 pre-images reveal target-object parts but are fragmented and scrambled, indicating invariance to occlusions and pose changes.VGG-M produces sharper reconstructions than AlexNet, while VGG-VD-16 emphasizes larger objects and has more washed-out colors.
  • CNN representations: Different random initializations produce geometric and style variations, showing that a component can represent multiple visual variants such as distinct vehicle wheels.The experiment probes invariances by generating four pre-images for the same representation components.

7 Visualization by caricaturization

Caricaturization starts from an image and exaggerates the patterns active in its representation, emphasizing combinations of components rather than individual activations. Across CNN layers, it exposes a progression from color and simple structures to object parts and mixed class-related patterns.

  • Method: Caricaturization exaggerates patterns active in a reference image and emphasizes combinations of multiple representation components.Unlike activation maximization, optimization starts from the reference image rather than random initialization.
  • VGG-M layer behavior: In VGG-M, early layers mainly saturate color, while conv2 and conv4 respond to long linear and round structures, respectively.The layer-specific caricatures show increasingly structured visual selectivity.
  • VGG-M layer behavior: Deeper VGG-M layers emphasize fox heads, and fully connected layers generate mixtures of fox heads, including hallucinated heads in the background.Selected VGG-M and VGG-VD layers produce similar caricaturization results on additional test images.

8 Summary

The paper unifies inversion, activation maximization, and caricaturization to probe classical and CNN representations. Its visualizations show that deep CNNs preserve detailed and instance-specific information while encoding different invariances and increasingly complex structures.

  • Unified framework: The paper experiments with inversion, activation maximization, and caricaturization in a unified framework for comparing classical representations and CNNs.The framework is intended to support visualization-based analysis across representation types.
  • CNN findings: Deep CNNs preserve photometrically accurate information and instance-specific object information even in very deep layers.The summary identifies these as among the most important findings from the CNN visualizations.
  • CNN findings: Intermediate convolutional layers capture local pose invariances, while fully connected layers capture larger variations in object layouts.These findings distinguish the types of invariance associated with different CNN stages.
  • CNN findings: Individual CNN components encode complex patterns that are usually not semantically obvious, while layers progress from lines and curves to object parts.The visualizations provide a layer-wise view of how image structures are represented.
  • Implications: The visualization methods can serve as diagnostic tools for testing whether proposed deep-layer activation patterns are semantically meaningful.Inversion and activation-maximization variants are suggested for validating such patterns through visualization.
Loading 1512.02017v3…