TL;DR
This paper proposes that the inherent randomness (stochasticity) in how neurons release neurotransmitters, combined with learning, can make neural networks more energy-efficient and better at lifelong learning. By learning to control this randomness, networks can focus their "effort" on the most important connections, leading to sparser, more efficient processing and improved memory retention.
What Problem Does This Paper Address?
The brain is remarkably energy-efficient, yet neuronal communication is inherently noisy. Specifically, when a neuron "fires," there's a high chance it won't actually trigger a response in the next neuron (a "synaptic failure"). While this seems counterintuitive for efficient information processing, some theories suggest it might be beneficial. Previous computational models explored this at the level of single synapses, but this paper tackles a bigger question: how does this presynaptic stochasticity affect the entire network's energy efficiency and its ability to learn continuously (lifelong learning), especially when facing the "stability-plasticity dilemma" (the challenge of learning new things without forgetting old ones)?
Methods
The authors introduce a model of artificial neural networks (ANNs) that incorporates presynaptic stochasticity. This means each synapse has a probability ($p_i$) of releasing neurotransmitters, modeled as a Bernoulli random variable. Crucially, this release probability ($p_i$) is not fixed but is learned alongside the synaptic strength ($m_i$).
The learning rule has two key components:
- Synaptic Strength Update: Standard gradient descent is used to update the expected synaptic strength.
- Release Probability Update: The release probability ($p_i$) is updated based on the gradient of the loss function. Synapses with larger gradients (indicating higher importance for the task) are encouraged to have higher release probabilities. This creates a link between a synapse's importance and its stochasticity.
To address the stability-plasticity dilemma, they introduce presynaptic consolidation:
- Learning Rate Modulation: The learning rate for a synapse is reduced if its release probability is high, effectively making important synapses more stable.
- Freezing Release Probabilities: Once a release probability surpasses a certain threshold, it's "frozen," further stabilizing important connections.
Datasets
The primary dataset used is MNIST, a benchmark for handwritten digit recognition. They also use variations:
- Split MNIST: A lifelong learning task where the model learns to classify digits in pairs (e.g., 0s and 1s, then 2s and 3s, etc.).
- Permuted MNIST: Each task involves a random permutation of MNIST pixels, creating tasks of equal difficulty.
- Permuted & Split Fashion MNIST: Similar tasks but using the Fashion MNIST dataset, which has more complex image features.
- Continuous Permuted MNIST: A variation that introduces a smooth transition between tasks, testing robustness to implicit task switch signals.
Key Insights and Ideas
- Stochasticity as an Importance Signal: Learned release probabilities can act as a direct indicator of a synapse's importance for the task. High release probability means the synapse is crucial.
- Energy Efficiency through Sparsity: By learning to control release probabilities, networks can become sparser. They allocate more "energy" (represented by synaptic strength and release probability) to important synapses and less to less important ones, leading to overall energy savings.
- Bridging Synaptic and Network Function: The paper connects low-level synaptic properties (stochasticity) to high-level network functions (energy efficiency, lifelong learning).
- Presynaptic Consolidation for Memory: Stabilizing synapses with high release probabilities (i.e., important ones) significantly improves memory retention in lifelong learning scenarios.
Key Results
- Energy Efficiency: The proposed model with plastic, stochastic release is significantly more energy-efficient than deterministic synapse models or stochastic models with fixed release probabilities. It achieves better mutual information per unit of energy.
- Sparsity: After learning, networks develop a sparse distribution of release probabilities, with a few synapses having high probabilities and most having low ones. This sparsity is task-dependent.
- Synapse Importance: Lesion experiments show that synapses with high release probabilities are indeed critical for network function, far more so than randomly selected synapses.
- Lifelong Learning: The presynaptic consolidation mechanism significantly improves memory preservation (accuracy on earlier tasks) and overall performance in lifelong learning tasks compared to models without consolidation and even state-of-the-art algorithms like EWC and BGD. For example, on Split MNIST, their model achieved ~82.9% average accuracy compared to ~77.7% for no consolidation.
Limitations and Future Work
- Model Simplification: The model uses a simplified Bernoulli distribution for synaptic release, whereas biological release is more complex (e.g., multiple release sites). However, they argue this simplification captures the essential event of neurotransmitter release failure.
- Metaplasticity Parameters: The model has several metaplasticity parameters (e.g., $p_{up}, p_{down}, g_{lim}, p_{freeze}$) that require tuning. While they show robustness across tasks, optimal tuning might still be a challenge.
- Biological Plausibility: While inspired by biology, the exact mechanisms of credit assignment and plasticity are still areas of active research. The paper notes their model is compatible with various biologically plausible learning rules.
- Future Work: The authors suggest exploring how this mechanism might apply to other network architectures and learning rules, and further investigating the link between presynaptic properties and synaptic stability in more detail. They also highlight the potential for this framework to inform the design of more efficient and robust artificial learning systems.
Summary updated