[Paper Review] Visualizing Deep Neural Networks for Speech Recognition with Learned Topographic Filter Maps

Andreas Krug, Sebastian Stober|arXiv (Cornell University)|Dec 6, 2019

Neural Networks and Applications6 references4 citations

TL;DR

This paper proposes learning topographic filter maps in deep neural networks for speech recognition by enforcing spatial similarity among neighboring filters during training, enabling more intuitive visualization of neuron activations as topographic maps. The key contribution is demonstrating that phoneme-responsive neurons are distributed across multiple regions rather than localized, limiting interpretability despite improved visual intuition.

ABSTRACT

The uninformative ordering of artificial neurons in Deep Neural Networks complicates visualizing activations in deeper layers. This is one reason why the internal structure of such models is very unintuitive. In neuroscience, activity of real brains can be visualized by highlighting active regions. Inspired by those techniques, we train a convolutional speech recognition model, where filters are arranged in a 2D grid and neighboring filters are similar to each other. We show, how those topographic filter maps visualize artificial neuron activations more intuitively. Moreover, we investigate, whether this causes phoneme-responsive neurons to be grouped in certain regions of the topographic map.

Motivation & Objective

• To address the unintuitive, permuted ordering of neurons in deep neural network layers, which hinders visual interpretation of activations.
• To adapt neuroscience-inspired topographic mapping techniques—commonly used in EEG—to visualize artificial neural network activations more intuitively.
• To investigate whether phoneme-responsive neurons cluster in specific regions of the learned 2D filter map, indicating functional organization.
• To evaluate whether the topographic structure enables more interpretable feature visualization through optimized input generation.

Proposed method

• A 2D grid is defined for each convolutional layer (e.g., 16×16 for 256 filters, 64×32 for 2048 filters) to impose spatial structure on filters.
• A regularization loss is applied to encourage similarity between each filter and its neighbors in a 3×3 neighborhood, weighted by inverse Euclidean distance to the center.
• The constraint minimizes cosine similarity between neighboring filter weights to promote topographic organization while preserving model performance.
• Gradient-adjusted Neuron Activation Profiles (GradNAPs) are used to average activations across inputs of the same phoneme or grapheme group, with baseline normalization.
• Optimal inputs are generated for the most responsive 3×3 regions using activation maximization (Yosinski et al., 2015), both for single filters and joint neighborhoods.
• Smoothed topographic maps are created via non-strided 3×3 average pooling to improve visual interpretability of activation patterns.

Experimental results

Research questions

• RQ1 Do phoneme-responsive neurons in deep speech recognition networks cluster in specific regions of a learned 2D filter map?
• RQ2 Can topographic filter maps improve the visual interpretability of artificial neuron activations compared to standard, unordered layer visualization?
• RQ3 To what extent are distributed representations of phonemes preserved across multiple regions in the topographic map?
• RQ4 Does optimizing inputs for the most responsive region yield interpretable, phoneme-specific patterns?

Key findings

• Phoneme-responsive neurons are not localized to a single region but are distributed across multiple areas of the topographic map, indicating a distributed representation.
• The most responsive region identified via GradNAPs often misses parts of the full distributed representation, as many other regions also show strong activation for the same phoneme.
• Visualizing activations directly on the grid without smoothing results in non-smooth, hard-to-interpret patterns, while average pooling over 3×3 windows produces visually coherent topographic maps.
• Optimizing inputs for individual filters in the most responsive region fails to reveal clear phoneme-typical patterns, suggesting that the representation is not localized to a single filter or small cluster.
• The current topographic regularization strategy does not enforce sufficient representational sparsity, as features remain spread across many regions of the map.
• The study concludes that while topographic maps improve visual intuition, stronger regularization—potentially on activations or with global similarity penalties—is needed to achieve localized, interpretable features.