Park Woojun, Kim Jongmu, Jeong Inhoi, Lee Kyoung J
Physics, Korea University, Seoul, 02841, Korea.
Mechanical Engineering, Korea University, Seoul, 02841, Korea.
J Comput Neurosci. 2025 Mar;53(1):163-179. doi: 10.1007/s10827-025-00896-4. Epub 2025 Feb 1.
The brain's ability to learn and distinguish rapid sequences of events is essential for timing-dependent tasks, such as those in sports and music. However, the mechanisms underlying this ability remain an active area of research. Here, we present a Pavlovian-conditioned spiking neural network model that may help elucidate these mechanisms. Using "three-factor learning rule," we conditioned an initially random spiking neural network to discriminate a specific spatiotemporal stimulus - a sequence of two or three pulses delivered within ms to two or three distinct neuronal subpopulations - from other pulse sequences differing by only a few milliseconds. Through conditioning, a feedforward structure emerges that encodes the target pattern's temporal information into specific topographic arrangements of stimulated subpopulations. In the readout phase, discrimination of different inputs is achieved by evaluating the shape and peak-shift characteristics of the spike density functions (SDFs) of input-triggered population bursts. The network's dynamic range - defined by the duration over which pulse sequences are processed accurately - is limited to around 10 ms, as determined by the duration of the input-triggered population burst. However, by introducing axonal conduction delays, we show that the network can generate "superbursts," producing a more complex and extended SDF lasting up to 30 ms, and potentially much longer. This extension effectively broadens the network's dynamic range for processing temporal sequences. We propose that such conditioning mechanisms may provide insight into the brain's ability to perceive and interpret complex spatiotemporal sensory information encountered in real-world contexts.
大脑学习和区分快速事件序列的能力对于依赖时间的任务至关重要，比如体育和音乐中的任务。然而，这种能力背后的机制仍是一个活跃的研究领域。在此，我们提出一种巴甫洛夫条件化的脉冲神经网络模型，它可能有助于阐明这些机制。使用“三因素学习规则”，我们让一个最初随机的脉冲神经网络进行条件训练，以区分一种特定的时空刺激——在几毫秒内传递给两三个不同神经元亚群的两个或三个脉冲序列——与其他仅相差几毫秒的脉冲序列。通过条件训练，出现了一种前馈结构，它将目标模式的时间信息编码到受刺激亚群的特定拓扑排列中。在读出阶段，通过评估输入触发的群体爆发的脉冲密度函数（SDF）的形状和峰值偏移特征来实现对不同输入的区分。该网络的动态范围——由准确处理脉冲序列的持续时间定义——限制在大约10毫秒左右，这是由输入触发的群体爆发的持续时间决定的。然而，通过引入轴突传导延迟，我们表明该网络可以产生“超级爆发”，产生一个更复杂、持续时间长达30毫秒甚至可能更长的SDF。这种扩展有效地拓宽了网络处理时间序列的动态范围。我们提出，这种条件训练机制可能有助于深入了解大脑感知和解释现实世界中遇到的复杂时空感官信息的能力。
