Aims: Premature ventricular contraction (PVC) is one of the commonly diagnosed arrhythmias according to the established guidelines. Since common approaches towards ECG diagnostics are mostly time-consuming and arrhythmia-type sensitive, deep-learning methods are state-of-the-art for their detection accuracy, which can, in some cases, even surpass qualified medical experts. However, PVCs localization via common deep-learning approaches requires a sufficiently large training set including detailed PVCs annotation. Our model has no such the limitation. It localizes PVCs based on information extracted from the whole signal annotation (containing/not-containing PVCs). Methods: Modified 1D convolutional neural network, commonly used for signal classification, was trained from the whole ECG labels. To enable PVCs localization with no available labels for single PVCs, Multiple Instance Learning (MIL) framework was applied. By processing input ECG via several convolutional and pooling layers, subsampled feature signal of a variable length (bag of instances) is obtained and this information is projected to a single output label by global average pooling. By nature, this feature signal indicates, whether this part of the signal contributed positively or negatively to final prediction. Thus, high feature values correspond to PVC positions (see figure). The maps of PVC likelihood are further processed with thresholding and maxima detection in order to get precise PVC locations. Results: Our method was tested on a publicly available database containing 1590 ECGs, including 672 signals with PVCs. According to visual assessment on 305 test signals, F1-score of the method reaches 0.89. The biggest advantage of our method, as compared to standard methods, is the possibility of application on various types of abnormalities without significant arrhythmia-specific manifestations in ECG. Conclusion: This simple deep-learning method for the localization of PVC achieves a promising precision while being trainable from the whole signal annotations instead of positional labels.