As a large volume of digital video data becomes available, along with revolutionary advances in multimedia technologies, demand related to efficiently retrieving and recounting multimedia data has grown. However, the inherent complexity in representing and recognizing multimedia data, especially for large-scale and unconstrained consumer videos, poses significant challenges. In particular, the following challenges are major concerns in the proposed research. One challenge is that consumer-video data (e.g., videos on YouTube) are mostly unstructured; therefore, evidence for a targeted semantic category is often sparsely located across time. To address the issue, a segmental multi-way local feature pooling method by using scene concept analysis is proposed. In particular, the proposed method utilizes scene concepts that are pre-constructed by clustering video segments into categories in an unsupervised manner. Then, a video is represented with multiple feature descriptors with respect to scene concepts. Finally, multiple kernels are constructed from the feature descriptors, and then, are combined into a final kernel that improves the discriminative power for multimedia event detection. Another challenge is that most semantic categories used for multimedia retrieval have inherent within-class diversity that can be dramatic and can raise the question as to whether conventional approaches are still successful and scalable. To consider such huge variability and further improve recounting capabilities, a per-exemplar learning scheme is proposed with a focus on fusing multiple types of heterogeneous features for video retrieval. While the conventional approach for multimedia retrieval involves learning a single classifier per category, the proposed scheme learns multiple detection models, one for each training exemplar. In particular, a local distance function is defined as a linear combination of element distance measured by each features. Then, a weight vector of the local distance function is learned in a discriminative learning method by taking only neighboring samples around an exemplar as training samples. In this way, a retrieval problem is redefined as an association problem, i.e., test samples are retrieved by association-based rules. In addition, the quality of a multimedia-retrieval system is often evaluated by domain-specific performance metrics that serve sophisticated user needs. To address such criteria for evaluating a multimedia-retrieval system, in MFoM learning, novel algorithms were proposed to explicitly optimize two challenging metrics, AP and a weighted sum of the probabilities of false alarms and missed detections at a target error ratio. Most conventional learning schemes attempt to optimize their own learning criteria, as opposed to domain-specific performance measures. By addressing this discrepancy, the proposed learning scheme approximates the given performance measure, which is discrete and makes it difficult to apply conventional optimization schemes, with a continuous and differentiable loss function which can be directly optimized. Then, a GPD algorithm is applied to optimizing this loss function.