Spatio-temporal pattern recognisers using spiking neurons and spike-timing-dependent plasticity

It has previously been shown that by using spike-timing-dependent plasticity, neurons can adapt to the beginning of a repeating spatio-temporal firing pattern in their input. In the present work we demonstrate that this mechanism can be extended to train recognisers for longer spatio-temporal input signals. Using a number of neurons that are mutually connected by plastic synapses and subject to a global winner-takes-all mechanism, chains of neurons can form where each neuron is selective to a different segment of a repeating input pattern, and the neurons are feedforwardly connected in such a way that both the correct input segment and the firing of the previous neurons are required in order to activate the next neuron in the chain. This is akin to a simple class of finite state automata. We show that nearest-neighbour spike-timing-dependent plasticity (where only the pre-synaptic spike most recent to a post-synaptic one is considered) leads to ''nearest-neighbour'' chains where connections only form between subsequent states in a chain (similar to classic ''synfire chains''). In contrast, ''all-to-all spike-timing-dependent plasticity'' (where all pre- and post-synaptic spike pairs matter) leads to multiple connections that can span several temporal stages in the chain; these connections respect the temporal order of the neurons. It is also demonstrated that previously learnt individual chains can be ''stitched together'' by repeatedly presenting them in a fixed order. This way longer sequence recognisers can be formed, and potentially also nested structures. Robustness of recognition with respect to speed variations in the input patterns is shown to depend on rise-times of post-synaptic potentials and the membrane noise. It is argued that the memory capacity of the model is high, but could theoretically be increased using sparse codes.

此前已有研究证实，借助脉冲时序依赖可塑性（spike-timing-dependent plasticity），神经元可适配输入中重复出现的时空放电模式的起始片段。本研究证明，该机制可拓展至针对更长时空输入信号的识别器训练。通过利用若干经由可塑性突触相互连接、且受全局胜者全取（winner-takes-all）机制约束的神经元，可形成神经元链：每个神经元对重复输入模式的不同片段具有选择性，且神经元以前馈方式连接，使得激活链中下一个神经元需要同时满足正确的输入片段与前序神经元放电两个条件，这类似于一类简单的有限状态自动机。我们发现，仅考虑与突触后锋电位时间最接近的突触前锋电位的近邻脉冲时序依赖可塑性（nearest-neighbour spike-timing-dependent plasticity），会形成"近邻"神经元链，即连接仅在链中的连续状态之间形成（类似经典的同步发放链（synfire chains））。与之相对，"全连接脉冲时序依赖可塑性"（all-to-all spike-timing-dependent plasticity，即所有突触前-突触后锋电位对均产生影响）则会形成可跨越链中多个时间阶段的多重连接，且这些连接严格遵循神经元的时间顺序。研究同时证实，此前已习得的独立神经元链可通过以固定顺序反复呈现而"拼接"在一起，借此可构建更长的序列识别器，理论上也可形成嵌套结构。针对输入模式的速度变化，识别的鲁棒性取决于突触后电位的上升时间与膜噪声。另有论证指出，该模型的记忆容量较高，但理论上可通过稀疏编码进一步提升。