This paper presents a spiking neural network (SNN) for digit recognition that uses spike-timing-dependent plasticity (STDP) for unsupervised learning. The network is designed with biologically plausible mechanisms, including conductance-based synapses, STDP with time-dependent weight changes, lateral inhibition, and an adaptive spiking threshold. Unlike most other systems, it does not use a teaching signal or class labels, and it achieves 95% accuracy on the MNIST benchmark, which is better than previous SNN implementations without supervision. The network's performance scales well with the number of neurons and is robust across different learning rules, suggesting its applicability in heterogeneous biological neural networks. The network architecture includes two layers: an input layer with 28×28 neurons and a processing layer with variable numbers of excitatory and inhibitory neurons. The input is encoded as Poisson-distributed spike trains, and the network learns to recognize digits through unsupervised learning. The results show that the network can achieve high accuracy without supervision, and the error analysis reveals that the network can correctly classify most digits, with some common misclassifications. The study also discusses the robustness of the network to different learning rules and the potential applications of the network in neuromorphic hardware for low-power machine learning. The network's design is scalable and flexible, allowing for training without labels and using only a few labels to assign neurons to classes. The study highlights the potential of spiking neural networks for efficient and low-power machine learning applications.This paper presents a spiking neural network (SNN) for digit recognition that uses spike-timing-dependent plasticity (STDP) for unsupervised learning. The network is designed with biologically plausible mechanisms, including conductance-based synapses, STDP with time-dependent weight changes, lateral inhibition, and an adaptive spiking threshold. Unlike most other systems, it does not use a teaching signal or class labels, and it achieves 95% accuracy on the MNIST benchmark, which is better than previous SNN implementations without supervision. The network's performance scales well with the number of neurons and is robust across different learning rules, suggesting its applicability in heterogeneous biological neural networks. The network architecture includes two layers: an input layer with 28×28 neurons and a processing layer with variable numbers of excitatory and inhibitory neurons. The input is encoded as Poisson-distributed spike trains, and the network learns to recognize digits through unsupervised learning. The results show that the network can achieve high accuracy without supervision, and the error analysis reveals that the network can correctly classify most digits, with some common misclassifications. The study also discusses the robustness of the network to different learning rules and the potential applications of the network in neuromorphic hardware for low-power machine learning. The network's design is scalable and flexible, allowing for training without labels and using only a few labels to assign neurons to classes. The study highlights the potential of spiking neural networks for efficient and low-power machine learning applications.