Despite the success of CNN models on a variety of image classification and segmentation tasks, their extensive computational and storage demands pose considerable challenges for real-world deployment on resource-constrained devices. Quantization is one technique that aims to alleviate these large storage requirements and speed up the inference process by reducing the precision of model parameters to lower-bit representations. In this paper, we introduce a novel post-training quantization method for model weights. Our method finds optimal clipping thresholds and scaling factors along with mathematical guarantees and minimizes quantization noise. Empirical results on real-world datasets demonstrate that our quantization scheme significantly reduces model size and computational requirements while preserving model accuracy.

Measurement of hydrological variables in rain-fed catchments is crucial for any decision support to mitigate flood risk. This paper describes our efforts in installation of a hydrometric sensor network in the Namal watershed area in Mianwali, Pakistan. Since the valley is formerly ungauged, minimal historic on-site data exists. Our network is composed of two devices, one to measure rainfall and another to measure stream level. For sensor placement and interpolation, past satellite data is used to obtain the spatial covariances of the rainfall process. The sensor nodes are placed so as to maximize the mutual information between gauged and ungauged locations. Kriging is applied to obtain the average spatial estimate for rainfall in the watershed. A level sensor on the dam reports the variation in the reservoir water level whereas another set of level sensors is used to evaluate lag times for flows in a major tributary. Results suggest that the data adequately captures the reservoir response to rainfall. Furthermore, we find that stream level measurements in the upper catchment can produce lead times sufficiently long enough to lower the damage caused by flash flooding in the Monsoon period.

Case-based reasoning (CBR) is a nature-inspired machine learning technique. It solves a new problem using the existing similar problems with their solutions stored in central repository known as case-base. It results in continuous growth of the case-base enhancing the problem solving capability of the system but at the same time compromising the performance. First performance challenge is continuous growth of the case-base. Second performance challenge is to handle the performance bottleneck without compromising the relevance of cases with their neighbourhood to solve new problems. Different approaches have been introduced in literature to address this performance challenge. In this work, a knowledge-base maintenance approach using fuzzy clustering has been presented which takes care of the performance bottleneck and does not compromise the problem solving capability of CBR. Performance of the proposed approach has been evaluated on different case-bases and results have been compared with the conventional CBR approach.

Autonomic systems exhibit self-managing behavior using various algorithms. Case-based reasoning is one the techniques that enable the autonomic manager to learn from past experience. Case-base is partitioned into some clusters in order to improve the retrieval efficiency. Deciding an appropriate number of clusters for a case-base is not a trivial problem. This paper proposes a randomized algorithm for determining the number of clusters to be formed of the case-base. Subsequently, a binary search-based case retrieval strategy has been applied to ensure enhanced retrieval time performance. The paper presents two versions of the randomized algorithm. The first version guarantees success but its computational cost is a function of random variable; the other guarantees a deterministic computational cost but success is not guaranteed. The performance of the proposed algorithms has been reported on a simulated case study of the Autonomic Forest Fire Application.