To select and de-select relevant and irrelevant information from the environment is the core of attentional mechanisms. But does selective attention modulate implicit learning? We tested whether the implicit acquisition of contingencies between features were dependent on the task-relevance of those features. We implemented the contingencies in a novel variant of the contextual cueing paradigm, using non-spatial predictive cues. In Experiment 1 (N=30), the predictive feature for target location was the shape of the distractors (task-relevant). In Experiment 2 (N=30), the color of distractors (task-irrelevant) cued target location. Results showed that participants learned to predict the target location from both the task-relevant and the task-irrelevant feature. For the purpose of further testing the significance of task-relevance, in Experiment 3 (N=30), we provided two predictive cues, shape (task-relevant) and color (task-irrelevant) simultaneously. Task-relevance modulated the learning of predictions from one single feature. We conclude that task relevance does not modulate implicit learning of associations per se, but rather the competition between predictive features via weighting processes. Attentional mechanisms thus do not seem to serve as filters at the encoding stage, but rather to weight the encoded features according to task-relevance.