How Auto-Encoders Could Provide Credit Assignment in Deep Networks via Target Propagation (1407.7906v3)

Abstract: We propose to exploit {\em reconstruction} as a layer-local training signal for deep learning. Reconstructions can be propagated in a form of target propagation playing a role similar to back-propagation but helping to reduce the reliance on derivatives in order to perform credit assignment across many levels of possibly strong non-linearities (which is difficult for back-propagation). A regularized auto-encoder tends produce a reconstruction that is a more likely version of its input, i.e., a small move in the direction of higher likelihood. By generalizing gradients, target propagation may also allow to train deep networks with discrete hidden units. If the auto-encoder takes both a representation of input and target (or of any side information) in input, then its reconstruction of input representation provides a target towards a representation that is more likely, conditioned on all the side information. A deep auto-encoder decoding path generalizes gradient propagation in a learned way that can could thus handle not just infinitesimal changes but larger, discrete changes, hopefully allowing credit assignment through a long chain of non-linear operations. In addition to each layer being a good auto-encoder, the encoder also learns to please the upper layers by transforming the data into a space where it is easier to model by them, flattening manifolds and disentangling factors. The motivations and theoretical justifications for this approach are laid down in this paper, along with conjectures that will have to be verified either mathematically or experimentally, including a hypothesis stating that such auto-encoder mediated target propagation could play in brains the role of credit assignment through many non-linear, noisy and discrete transformations.

• The paper proposes using auto-encoders to facilitate credit assignment in deep networks through target propagation, presenting a novel alternative to standard back-propagation.
• The methodology employs layer-local auto-encoders trained to reconstruct inputs, using this learned reconstruction signal to propagate targets and update weights across layers, even with discrete hidden units.
• This approach offers potential benefits for training highly non-linear or discrete deep networks and speculates on providing a biologically plausible model where reconstruction acts as a local learning signal.

Exploiting Auto-Encoders for Credit Assignment in Deep Learning via Target Propagation

The paper "How Auto-Encoders Could Provide Credit Assignment in Deep Networks via Target Propagation" presents a novel approach for training deep networks by leveraging auto-encoders to facilitate target propagation, offering an alternative to the classic back-propagation method. Back-propagation has been central to deep learning, yet it faces challenges with gradient vanishing or explosion in deep or recurrent networks. This paper proposes target propagation using layer-wise reconstruction as a local training signal to overcome these difficulties, potentially providing a biologically plausible mechanism for learning.

Proposed Methodology

The central idea is to employ auto-encoders to perform credit assignment through target propagation, minimizing reliance on back-propagated gradients. The auto-encoders, trained to perform layer-local reconstructions, propagate targets similar to how gradients guide updates across layers. This learned reconstruction, through target propagation, allows training deep networks with discrete hidden units, handling both infinitesimal and larger discrete changes effectively.

For each layer in a deep network, the authors suggest implementing a local training criterion that matches the distributions generated by top-down and bottom-up approaches para-metrically. This involves maximizing the likelihood of a reconstruction of observed inputs, which serves as a proxy for $\frac{\partial \log P(h_l)}{\partial h_l}$ across layers.

Strong Numerical Results and Bold Claims

The authors assert, through theoretical conjectures, that auto-encoder-mediated target propagation could better model long chains of non-linear operations in deep networks, potentially even offering insights into biological brain function regarding credit assignment through many non-linear and noisy transformations. The reduced dependency on back-propagation requirements could extend deep networks to cases involving discrete or binary node activations, enhancing the capability of deep learning models to represent complex data distributions. The paper hypothesizes that such mechanisms could naturally extend to biologically plausible models of brain function.

Implications and Theoretical Speculations

The implications of this research are substantial both practically and theoretically. Practically, it suggests potential improvements in deep network training efficiency, especially for networks with high non-linearities or discrete components. Theoretically, it speculates on a biologically inspired learning paradigm whereby the reconstruction acts as a local gradient signal for learning. This aligns with the idea that feedback connections in neural circuits convey a form of training signal, driven here through reconstruction, which represents high-probability configurations more typical of actual neural circuit operations.

Future Developments in AI

Future research could probe the practical viability of target propagation in real-world applications, scrutinizing robustness when scaling and adapting across diverse data modalities. Moreover, exploring systematic links to biology, this paper lays a foundation for investigating neuron-based computation models and synaptic learning akin to the mechanisms proposed, potentially unveiling further insights into artificial intelligence inspired by biological systems.

Overall, this paper provokes substantial inquiry into leveraging reconstruction as a potent signal for learning representation, hinting at transformative repercussions in both artificial and biological intelligence paradigms.