Machine learning heavily relies on the ability to learn/approximate real functions. State variables, the perceptions, internal states, etc, of an agent are often represented as real numbers; grounded on them, the agent has to predict something, or act in some way. In this view, this outcome is a nonlinear function of the inputs. It is thus a very common task to fit a nonlinear function to observations, namely solving a regression problem. Among other approaches, the LARS is very appealing, for its nice theoretical properties, and actual efficiency to compute the whole $l_1$ regularization path of a supervised learning problem, along with the sparsity. In this paper, we consider the kernelized version of the LARS. In this setting, kernel functions generally have some parameters that have to be tuned. In this paper, we propose a new algorithm, the Equi-Correlation Network (ECON), which originality is that while computing the regularization path, ECON automatically tunes kernel hyper-parameters; thus, this opens the way to working with infinitely many kernel functions, from which, the most interesting are selected. Interestingly, our algorithm is still computationaly efficient, and provide state-of-the-art results on standard benchmarks, while lessening the hand-tuning burden.