Generation of Internal Representation by α-Transformation

NeurIPS 1993

TL;DR

This paper introduces an "α-transformation" method to dynamically adjust the internal representation of neural networks by manipulating an entropy function. This allows for creating more interpretable "kernel networks" by minimizing entropy and improving generalization performance by increasing it, offering a flexible way to tailor network architectures to specific tasks.

What Problem Does This Paper Address?

Neural networks are powerful because they can learn to create internal representations of data. However, there's been limited research on what kind of internal representations are best for different tasks, or how to automatically generate them. The authors identify a need for methods that can transform these internal representations to achieve specific goals, such as making the network's decision-making process more understandable (interpretability) or improving its ability to generalize to unseen data.

Methods

The core of the paper is the "entropy method" for transforming internal representations. They define an "internal entropy" based on the activity of the hidden units in a neural network. This entropy can be manipulated by changing a parameter, denoted as 'α', in the entropy function.

  • α-Transformation: This is the novel technique proposed. By adjusting 'α', the network's internal representation can be shifted towards:
  • Minimum Entropy: This state activates only a single hidden unit per input pattern. This leads to "kernel networks" with fewer, more interpretable hidden units (kernel hidden units). The authors propose "kernel hidden unit analysis" to identify these units.
  • Maximum Entropy: This state activates all hidden units equally.
  • Intermediate Representations: These lie between the minimum and maximum entropy states, aiming to balance interpretability and generalization.
  • Entropy Minimization/Maximization: The α-transformation is integrated into the back-propagation algorithm. The update rule for input-to-hidden connections includes a term that either minimizes or maximizes the internal entropy, controlled by 'α', alongside the standard error minimization.

Datasets

The paper describes experiments using:

  • Autoencoder: Used to demonstrate the creation of kernel networks. The specific dataset is not detailed, but the task involved reproducing five similar alphabet letters (B, C, D, E, F, G).
  • Frequency Identification Problem: A benchmark task where networks must identify three different frequencies of sine waves. The input data consisted of 64 samples from sine waves, and the output was one of three classes corresponding to the frequencies.

Key Insights and Ideas

  • Entropy as a Control Knob: The paper elegantly frames internal representation as a spectrum controlled by entropy. Minimizing entropy leads to sparse, interpretable representations, while increasing it can improve generalization.
  • Kernel Networks for Interpretability: By minimizing entropy, the authors show that it's possible to extract a small set of "kernel hidden units" that are crucial for the network's function. This significantly reduces network complexity and makes understanding the network's behavior much easier.
  • Generalization Trade-off: The authors hypothesize and demonstrate that while minimum entropy representations are interpretable, they might not generalize well. Conversely, intermediate entropy levels are proposed as a way to achieve better generalization by distributing information across more hidden units.
  • Unified Approach: The α-transformation provides a single framework to address both interpretability and generalization, allowing researchers to tune the network's internal workings based on the specific problem requirements.

Key Results

  • Autoencoder: The entropy method, with kernel hidden unit analysis, successfully reduced a network with 35 hidden units to a "kernel network" with only 3 hidden units, while maintaining performance. The roles of these kernel hidden units could be explicitly determined.
  • Frequency Identification: Experiments showed that increasing entropy (from 0.623 to 0.778) led to a significant decrease in Hamming distance (from 6.296 to 1.111), indicating improved generalization. In some cases, perfect generalization (zero Hamming distance) was achieved. Similarly, SSE (Sum of Squared Errors) also decreased with increased entropy.

Limitations and Future Work

  • Interpretability vs. Generalization Balance: The paper suggests that finding the optimal 'α' for the best balance between interpretability and generalization is crucial. Explicitly determining the relationship between kernel and intermediate representations is highlighted as a future direction to achieve both goals simultaneously.
  • Parameter Tuning: While the method is presented, the paper doesn't delve deeply into systematic methods for finding the optimal 'α' for any given problem, which would be a natural area for future research.
  • Specific Datasets: The details of the autoencoder dataset are sparse, making it difficult to fully assess the impact of the method on that specific task without further information.

Summary updated

BibTeX
@inproceedings{1993generation,
  title = {{Generation of Internal Representation by α-Transformation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1993},
}
Figure