Approximate Bayesian Computation is widely used in systems biology for\ninferring parameters in stochastic gene regulatory network models. Its\nperformance hinges critically on the ability to summarize high-dimensional\nsystem responses such as time series into a few informative, low-dimensional\nsummary statistics. The quality of those statistics acutely impacts the\naccuracy of the inference task. Existing methods to select the best subset out\nof a pool of candidate statistics do not scale well with large pools of several\ntens to hundreds of candidate statistics. Since high quality statistics are\nimperative for good performance, this becomes a serious bottleneck when\nperforming inference on complex and high-dimensional problems. This paper\nproposes a convolutional neural network architecture for automatically learning\ninformative summary statistics of temporal responses. We show that the proposed\nnetwork can effectively circumvent the statistics selection problem of the\npreprocessing step for ABC inference. The proposed approach is demonstrated on\ntwo benchmark problem and one challenging inference problem learning parameters\nin a high-dimensional stochastic genetic oscillator. We also study the impact\nof experimental design on network performance by comparing different data\nrichness and data acquisition strategies.\n