Presynaptic stochasticity improves energy efficiency and helps alleviate the stability-plasticity dilemma

Simon Schug, Frederik Benzing, Angelika Steger

TL;DR

This paper proposes that the inherent randomness (stochasticity) in how neurons communicate, specifically at the synapse, isn't just noise but a crucial mechanism for energy efficiency and better learning. By modeling networks where synapses can "fail" to transmit signals, the authors show this leads to sparser, more efficient energy allocation and helps solve the long-standing "stability-plasticity dilemma" in lifelong learning.

What Problem Does This Paper Address?

The brain is incredibly energy-efficient, yet synaptic transmission, a fundamental process, is notoriously unreliable – action potentials often arrive at a synapse without triggering a signal. This "synaptic failure" seems counterintuitive for efficient information processing. While some models suggest this stochasticity saves energy at the single-synapse level, it's unclear how this scales to complex, goal-driven neural networks. Furthermore, neural networks face the "stability-plasticity dilemma": how to learn new information (plasticity) without overwriting or forgetting previously learned knowledge (stability). This paper investigates how presynaptic stochasticity, combined with plasticity, impacts both energy efficiency and the stability-plasticity dilemma in artificial neural networks.

Methods

The authors introduce a model of artificial neural networks (ANNs) that explicitly incorporates presynaptic stochasticity. Instead of assuming a deterministic signal transmission, each synapse has a "release probability" ($p_i$) that determines whether a signal is transmitted. This probability is learned alongside the synaptic strength ($m_i$), which represents the efficacy of the synapse when it does transmit.

Key methodological aspects include:

  • Stochastic Synapse Model: Synaptic transmission is modeled as a Bernoulli process, where $r_i \sim \text{Bernoulli}(p_i)$ determines if a neurotransmitter is released. The effective synaptic strength is then $w_i = r_i m_i$.
  • Learning Rule: A standard gradient descent approach is used to update synaptic strengths ($m_i$). Crucially, the release probabilities ($p_i$) are also updated, with an increase for synapses that receive larger gradient updates (indicating importance) and a decrease for smaller ones. This links synaptic importance to release probability.
  • Energy Measurement: Energy consumption is approximated by the expected synaptic strength, reflecting the metabolic cost of maintaining synaptic efficacy.
  • Information Measurement: Mutual information is used to quantify how much task-relevant information is transmitted by the network's output.
  • Lifelong Learning Mechanism: To address the stability-plasticity dilemma, a "presynaptic consolidation" mechanism is introduced. This involves modulating the learning rate based on release probability (lower learning rate for high release probability synapses) and freezing release probabilities above a certain threshold. This effectively stabilizes synapses deemed important.

Datasets

The paper uses the following datasets:

  • MNIST: A standard dataset of handwritten digits, used for evaluating energy efficiency and information processing.
  • Split MNIST: A lifelong learning benchmark where the network learns to classify digit parity sequentially across different subsets of digits.
  • Permuted MNIST: Another lifelong learning task where input pixels are randomly permuted for each task, creating distinct but equally difficult tasks.
  • Fashion MNIST: Used for testing the lifelong learning capabilities on a more complex dataset.
  • Split and Permuted Fashion MNIST: Variations of the Fashion MNIST dataset used for lifelong learning experiments.

Key Insights and Ideas

  • Stochasticity as an Efficiency Driver: Synaptic failures, rather than being detrimental, can be leveraged by the network to allocate energy more efficiently. By making synapses with low importance less likely to transmit signals, the network conserves energy.
  • Release Probability as Synaptic Importance: The learned release probability ($p_i$) emerges as a direct proxy for a synapse's importance to the network's task. Synapses that are critical for processing information tend to have higher release probabilities.
  • Fisher Information Link: The paper provides a theoretical link between release probability and Fisher information, a measure of a parameter's sensitivity. Synapses with higher Fisher information (more critical for network function) are more likely to develop higher release probabilities.
  • Consolidation for Stability: By selectively stabilizing synapses with high release probabilities (and thus high importance), the network can better retain learned information over time, effectively mitigating the stability-plasticity dilemma. This is akin to "consolidating" important memories.
  • Sparsity at Multiple Levels: The model promotes sparsity not only at the synaptic level (few synapses are highly active) but also at the neuronal level (some neurons become disproportionately important).

Key Results

  • Energy Efficiency: The model with plastic, stochastic release achieved significantly more energy-efficient solutions compared to deterministic synapses or stochastic synapses with fixed release probabilities. Figure 1a shows a better trade-off between mutual information and energy.
  • Synaptic Importance Encoding: Figure 2a demonstrates that after learning, networks develop a sparse distribution of release probabilities, with a small subset of synapses having high probabilities. Lesion experiments (Figure 2b) confirm that synapses with high release probabilities are crucial for task performance.
  • Lifelong Learning Improvement: The presynaptic consolidation mechanism significantly improved memory retention (Figure 4b) and overall performance across multiple tasks (Figure 4c) compared to models without consolidation. It outperformed standard machine learning algorithms like Bayesian Gradient Descent and Elastic Weight Consolidation in some lifelong learning scenarios.
  • Robustness: The model's performance was found to be robust to variations in metaplasticity parameters and transferable across different datasets and task types.

Limitations and Future Work

  • Model Simplification: The model uses a simplified representation of synaptic transmission (Bernoulli) and energy consumption. Real biological synapses are more complex.
  • Biological Plausibility of Learning Rule: While the learning rule for release probability is motivated by gradient updates, its direct biological counterpart is not fully established. The paper notes that alternative credit assignment mechanisms exist and are compatible with the model.
  • Future Work: The authors suggest exploring more complex synaptic models and investigating the role of presynaptic stochasticity in other computational tasks and biological contexts. They also highlight the potential for this mechanism to inform the design of more robust and efficient artificial learning systems.

Summary updated

Abstract

When an action potential arrives at a synapse there is a large probability that no neurotransmitter is released. Surprisingly, simple computational models suggest that these synaptic failures enable information processing at lower metabolic costs. However, these models only consider information transmission at single synapses ignoring the remainder of the neural network as well as its overall computational goal. Here, we investigate how synaptic failures affect the energy efficiency of models of entire neural networks that solve a goal-driven task. We find that presynaptic stochasticity and plasticity improve energy efficiency and show that the network allocates most energy to a sparse subset of important synapses. We demonstrate that stabilising these synapses helps to alleviate the stability-plasticity dilemma, thus connecting a presynaptic notion of importance to a computational role in lifelong learning. Overall, our findings present a set of hypotheses for how presynaptic plasticity and stochasticity contribute to sparsity, energy efficiency and improved trade-offs in the stability-plasticity dilemma.
BibTeX
@inproceedings{schug2021presynaptic,
  author = {Simon Schug and Frederik Benzing and Angelika Steger},
  title = {{Presynaptic stochasticity improves energy efficiency and helps alleviate the stability-plasticity dilemma}},
  booktitle = {eLife},
  year = {2021},
  doi = {10.7554/elife.69884},
}
Figure