This paper proposes a simple but highly efficient expansion-based model for continual learning. The recent feature transformation, masking and factorization-based methods are efficient, but they grow the model only over the global or shared parameter. Therefore, these approaches do not fully utilize the previously learned information because the same task-specific parameter forgets the earlier knowledge. Thus, these approaches show limited transfer learning ability. Moreover, most of these models have constant parameter growth for all tasks, irrespective of the task complexity. Our work proposes a simple filter and channel expansion-based method that grows the model over the previous task parameters and not just over the global parameter. Therefore, it fully utilizes all the previously learned information without forgetting, which results in better knowledge transfer. The growth rate in our proposed model is a function of task complexity; therefore for a simple task, the model has a smaller parameter growth while for complex tasks, the model requires more parameters to adapt to the current task. Recent expansion-based models show promising results for task incremental learning (TIL). However, for class incremental learning (CIL), prediction of task id is a crucial challenge; hence, their results degrade rapidly as the number of tasks increase. In this work, we propose a robust task prediction method that leverages entropy weighted data augmentations and the model’s gradient using pseudo labels. We evaluate our model on various datasets and architectures in the TIL, CIL and generative continual learning settings. The proposed approach shows state-of-the-art results in all these settings. Our extensive ablation studies show the efficacy of the proposed components.

The human brain has billions of neurons. However, we perform tasks using only a few concurrently active neurons. Moreover, an activated neuron inhibits the activity of its neighbors. Selfless Learning exploits these neurobiological principles to solve the problem of catastrophic forgetting in continual learning. In this paper, we ask a basic question: can the selfless learning idea be used to improve the accuracy of deep convolutional networks on a single classification task? To achieve this goal, we introduce two regularizers and formulate a curriculum learning-esque strategy to effectively enforce these regularizers on a network. This has resulted in significant gains over vanilla cross-entropy training. Moreover, we have shown that our method can be used in conjunction with other popular learning paradigms like curriculum learning.

The Micro, Small, and Medium Enterprises (MSME) sector is one of the fastest growing industrial sectors all over the world. Clusters of small and medium enterprises have been very successful worldwide in promoting their growth. The contributions of MSME clusters to Indian economy in terms of increasing manufacturing capacities, increasing exports, employment generation etc. have been widely studied. The success of these clusters is largely due to the individual small firms of the cluster closely collaborating and sharing resources. Sharing is the theme of our paper, where we focus on understanding how productivity and efficiency of an SME cluster in developing region can be improved by furthering collaboration among its units. The hypotheses of our study were that (a) there exists limited sharing and collaboration through informal networks among SME cluster units in India, (b) these clusters could be more productive by improving the collaboration among cluster actors, and (c) the sharing and cooperation within the cluster can be improved with formal networks among the cluster entities. To validate our hypotheses, we undertook a field study of the SME jeans cluster in Bellary, Karnataka. In this paper, we have put forth our findings from the study, followed by a proposal of an affordable ICT infrastructure solution to increase collaboration and hence productivity of the enterprises in Indian SME clusters.

Indian business clusters have contributed immensely to the country's industrial output, poverty alleviation and employment generation. However, with recent globalization these clusters can loose out to international competitors if they do not continuously innovate and take advantage of the new opportunities that are available through economic liberalization. In this paper, we discuss how information and communication technologies (ICT) can help in improving the productivity and growth of these clusters.

We present the on-demand link vector (OLIVE) protocol, a routing protocol for ad hoc networks based on link-state information that is free of routing loops and supports destination-based packet forwarding. Routers exchange routing information reactively for each destination in the form of complete paths, and each node creates a labeled source graph based on the paths advertised by its neighbors. A node originates a broadcast route request (RREQ) to obtain a route for a destination for which a complete path does not exist in its source graph. When the original path breaks, a node can select an alternative path based on information reported by neighbors, and a node can send a unicast RREQ to verify that the route is still active. A node that cannot find any alternate path to a destination sends route errors reliably to those neighbors that were using it as next hop to the destination. Using simulation experiments in ns2, OLIVE is shown to outperform dynamic source routing, ad hoc on-demand distance vector, optimized link-state routing protocol, and topology broadcast based on reverse-path forwarding, in terms of control overhead, throughput, and average network delay, while maintaining loop-free routing with no need for source routes.

We present the on-demand link vector (OLIVE) protocol, a routing protocol for ad-hoc networks based on link-state information that is free of routing loops and supports destination-based packet forwarding. Routers exchange routing information reactively for each destination in the form of complete paths, and each node creates a labeled source graph based on the paths advertised by its neighbors. A node originates a broadcast route request to obtain a route for a destination for which a complete path does not exist in its source graph. When the original path breaks, a node can select an alternative path based on information reported by neighbors, and a node can send a unicast route request to verify that the route is still active. A node that cannot find any alternate path to a destination sends route errors reliably to those neighbors that were using it as next hop to the destination. Using simulation experiments in ns2, OLIVE is shown to outperform DSR, AODV, OLSR and TBRPF, in terms of control overhead, throughput, and average network delay, while maintaining loop-free routing with no need for source routes.

We present a node-centric approach to hybrid routing for ad hoc networks in which normal nodes are distinguished from special nodes, called netmarks, hosting popular network services or functioning as points of attachment to the Internet. With node-centric hybrid routing, netmarks force other common nodes to maintain routing information for them by advertising their routing information as in table-driven routing protocols. Routes between peer nodes are set up on-demand. A node-centric routing solution is presented based on partial link state information. Simulation results using ns2 show that this approach performs much better than on-demand routing protocols.

Traditional routing protocols based on link-state information form a network topology through the exchange of link-state information by flooding or by reporting partial topology information and compute shortest routes to each reachable destination using a path-selection algorithm like Dijkstra's algorithm or the Bellman-Ford algorithm. However, in an on-demand link-state routing protocol, no one node needs to know the paths to every other node in the network. Accordingly, when a node chooses a next hop for a given destination, it must be true that the next hop has reported a path to the same destination; otherwise, packets sent through that node would be dropped. We present a new path-selection algorithm that unlike traditional shortest path algorithms, computes shortest paths with the above on-demand routing constraint.

The on-demand routing protocols that have been proposed to date use either path information (e.g., DSR) or distance information (e.g., AODV). We present SOAR, an on-demand link-state protocol based on partial link-state information in which a wireless router communicates to its neighbors the link states of only those links in its source tree that belong to the paths it chooses to advertise for reaching destinations with which it has active flows, SOAR does not require periodic link-state advertisements when there are no link connectivity changes in the network. Simulation studies for several scenarios of node mobility and traffic flows reveal that SOAR performs more efficiently than DSR, which is one of the best performing on-demand routing approaches based on path information.