Towards Biologically Plausible Deep Learning (1502.04156v3)

Abstract: Neuroscientists have long criticised deep learning algorithms as incompatible with current knowledge of neurobiology. We explore more biologically plausible versions of deep representation learning, focusing here mostly on unsupervised learning but developing a learning mechanism that could account for supervised, unsupervised and reinforcement learning. The starting point is that the basic learning rule believed to govern synaptic weight updates (Spike-Timing-Dependent Plasticity) arises out of a simple update rule that makes a lot of sense from a machine learning point of view and can be interpreted as gradient descent on some objective function so long as the neuronal dynamics push firing rates towards better values of the objective function (be it supervised, unsupervised, or reward-driven). The second main idea is that this corresponds to a form of the variational EM algorithm, i.e., with approximate rather than exact posteriors, implemented by neural dynamics. Another contribution of this paper is that the gradients required for updating the hidden states in the above variational interpretation can be estimated using an approximation that only requires propagating activations forward and backward, with pairs of layers learning to form a denoising auto-encoder. Finally, we extend the theory about the probabilistic interpretation of auto-encoders to justify improved sampling schemes based on the generative interpretation of denoising auto-encoders, and we validate all these ideas on generative learning tasks.

• The paper reinterprets spike-timing-dependent plasticity as a form of stochastic gradient descent, linking biological learning with machine learning optimization.
• It introduces a variational EM framework to tackle the credit assignment problem without relying on exact weight symmetry in backpropagation.
• The approach advances generative modeling with coordinated forward and backward computations, improving sampling quality and potentially enhancing energy efficiency.

Towards Biologically Plausible Deep Learning: An Expert Overview

The paper entitled "Towards Biologically Plausible Deep Learning" by Bengio et al. addresses fundamental issues in the development of artificial neural networks that are inspired by the brain, yet diverge significantly from biological reality. The authors aim to propose methods and frameworks that align more closely with biological principles while maintaining the efficiency and effectiveness of current deep learning approaches.

Key Contributions

The paper makes several novel contributions:

1. Interpretation of Spike-Timing-Dependent Plasticity (STDP): STDP is a leading theory explaining synaptic changes in biological neurons based on the timing of spikes. The authors propose that STDP can be interpreted as a form of stochastic gradient descent, aligning it with machine learning principles. Specifically, the updates to synaptic weights correlate with a gradient descent on objective functions when neuron firing rates shift towards optimizing these objectives.
2. Variational EM Framework: The paper introduces a rationale for interpreting STDP within a machine learning context, proposing that neural activities and synaptic updates can be viewed as an instance of the variational Expectation-Maximization (EM) algorithm. This framework attempts to address the credit assignment problem without relying on biologically implausible gradient back-propagation.
3. Forward and Backward Computations and Auto-encoders: The authors extend the probabilistic interpretation of denoising auto-encoders to enable sampling from generative models. The proposed approach involves propagating activations forward and backward through layers, aligning closely with biologically feasible learning mechanisms.
4. Solving the Weight Transport Problem: One of the significant challenges in aligning artificial neural networks with biology is the weight transport problem, where feedback connections must use exact symmetric weights of the feedforward paths. This paper presents strategies that eliminate the necessity of exact weight symmetry.
5. Generative Modeling and Sampling Improvements: By leveraging the variational perspective, the authors derive a generative training process that enhances sampling quality through alternating encoding and decoding operations. This approach is posited to yield better samples than traditional directed graphical models.

Implications and Future Directions

The paper's contributions have several implications for the future of AI and deep learning:

• Biological Plausibility and AI: By reducing reliance on back-propagation, the proposed methods offer a path towards AI systems that are more aligned with biological processes. This alignment could inspire new architectures and algorithms that capture the efficiency observed in biological systems.
• Energy Efficiency and Robustness: Aligning AI with biological principles could lead to more energy-efficient and robust models. Biological systems are inherently resilient and can operate under varying conditions, offering insights that could enhance AI systems beyond their current limitations.
• Interdisciplinary Research: The findings underscore the importance of interdisciplinary research, blending insights from neuroscience, biology, and machine learning to drive the next wave of innovations in AI.

Future work is needed to expand these concepts into supervised learning and reinforcement learning contexts. Investigating biological implementations that incorporate spikes and adhere to neural constraints, such as the sign of weights in inhibitory and excitatory neurons, will further bridge the gap between artificial and biological neural networks.

As the field progresses, the exploration of biologically plausible deep learning not only provides a theoretical basis for advancing AI but also enriches our understanding of the brain's learning algorithms. Inviting such cross-pollination of ideas has the potential to transform both AI development and our comprehension of biological intelligence.