This paper presents an architectural design of a novel intelligent accelerometer sensor, constructed from a multi- mode fiber optic cable, and studied its performances. Experimental proofs are provided to demonstrate that as a multi-mode fiber deforms, variation in the shape and structure of speckle patterns will provide deterministic information for measuring the deformation parameters. We used a deep learning approach to analyze the shape of speckle patterns. A multilayer feed-forward convolutional neural network (CNN) has been used to utilize and classify images of speckle pattern into distinct pre-known classes of acceleration vectors. A pendulum setup is used for collecting repeatable and predictable sample data. The presented intelligent sensor is compared with a Micro Electromechanical Machine System (MEMS) accelerometer performances. Estimation of magnitude and direction of the acceleration vector in one plane of motion is achieved with high accuracy (over 97%), when the CNN was trained for 150 epochs. The results confirm that this novel accelerometer sensor performs as well as a MEMS accelerometer. With proper manufacturing, this novel fiber accelerometer has the potential to overcome the limitations associated with conventional accelerometer sensors, normally due to their physical characteristic, accuracies or performances. The potential sensor resulting from this research is expected to be simple, compact, and economically feasible. Moreover, the sensing approach can easily be generalized to measure other physical phenomenal including vibration and displacement.