The incoming IoT big data era requires efficient and resource-saving mining of large sets of distributed data. To this end, a viable approach could be to attempt to resort to the synergic exploitation of two emerging paradigms, namely, Deep Neural Networks (DNNs) with early exits and Fog Computing. From an IoT perspective, the most appealing feature of DNNs is its hierarchical layer-wise organization that is exploited for feature learning and/or pattern analysis/classification, potentially matching the hierarchical organization of the underlying Fog-supported execution platform. However, both the training and inference phases of current DNNs are too cumbersome to be executed by IoT devices, and their remote and centralized implementation induces large communication delays and is bandwidth-wasting.
Motivated by these considerations, the two-year DeepFog project aims at developing the integration of the emerging paradigm of the Conditional DNNs (CDNNs) with early exits on the novel Fog Computing (FC) platform that is hierarchically organized in multiple networked tiers. For this purpose, the exploitation of the powerful and recent paradigm of Federated Learning (FL) is also utilized for the distributed and asynchronous Fog-supported training of the underlying CDNN under resource and fairness constraints.
Overall, main goals of the DeepFog project are to:
- formalize the main building blocks and functionalities of the proposed DeepFog technological platform, by designing the most suitable algorithms for the CDNN training;
- design and validate through software simulations a novel asynchronous fast-adaptive FL protocol for the training of CDNNs for Fog-supported IoT stream applications;
- design and validate through extensive numerical tests the performance of the overall resulting optimized DeepFog technological platform, and compare its energy-vs.-complexity-vs.-accuracy performance under some real-world datasets for IoT-oriented real-time big-data applications.
5) EXPECTED INNOVATIVE PROJECT OUTPUTS AND THEIR POTENTIAL TECHNO-SOCIAL IMPACTS
The DeepFog technological platform is expected to quickly foster the development of Fog-supported Future Internet Applications [11], that are considered a world-wide huge business opportunity (Fig.3). In particular, it is foreseen that both DL applications and Fog related outputs could move from the research cycle to prototype cycle in the next years, since in between five years we will expect six times the software revenue with respect today (Fig.3a).
The main services that are expected to be fostered by the resulting IoT-CDNN-FC converged platform promoted by the DeepFog paradigm are summarized in Table 2, together with some possible supported applications. Specifically, it is expected that the IoT-CDNN-FC convergence highlights three benefits: i) the demand for real-time analytics instead of batch processing at remote Cloud data centers; ii) allowing the acquisition and joint mining of data generated by spatially scattered IoT devices under both real-time and fairness requirements; and, iii) the use of resource-limited IoT devices supporting the planned services of Table 2. All these benefits well match with the native features of both CDNN and FC paradigms, namely, their local processing capability and their inherently distributed and scalable nature.
Expected Technological Impact: The proposed DeepFog project is expected to have a significant positive impact on the development of future low-latency and energy-aware intelligent applications done feasible by the convergence of FC and DL paradigms. This is expected to have a relevant technological impact on many novel families of technologies, such as Smart Wireless Sensor Networks (SWSNs), Smart Objects of IoT and Intelligent Embedded Systems. Overall, it is expected that the proposed DeepFog project allows to enter a new phase where real-world problems emerging from complex real-time IoT applications are addressed.
Expected Social-Economical Impact: It is expected that intelligent FC platforms are able to provide more robust, energy-efficient, lower latency and higher QoS applications and, for this reason, are expected to increase the trust people have in machine-related and real-time critical applications. Furthermore, since the novel DeepFog platform relies on FL protocols that does not need to share local data, it is expected that it solves data security and related privacy issues. Overall, it is expected that the proposed DeepFog paradigm will provide a feasible technological platform for the actual support and real-world implementation of a large overall spectrum of Fog-supported DL-aided IoT-oriented stream applications.
6) PROJECT PLAN AND PROJECT TARGETS
The two-year research activity of the overall DeepFog workprogramme is organized into two Workpackages, coordinated Prof. M. Scarpiniti and by Prof. E. Baccarelli, respectively. As detailed in Table III, each WP comprises two tasks.
The DeepFog project engages 4 researchers (3 promoters supported by 1 PhD student) and will be coordinated by Prof. Enzo Baccarelli. The multiple scientific sectors involved by the DeepFog project (ING-INF/03 and ING-IND/31) confirm, indeed, the inter-disciplinary nature of the DeepFog project. Table IV details the role/task covered by each participant, together with the planned workprogramme over the TWO-year life cycle of the project, and related cross-interactions.
It follows a detailed description of tasks.
T1.1-Optimization of layers placement and dynamic energy-saving computing-networking resource allocation - Equipping a layer of a baseline DNN with a local classifier is not for free, because there are three main issues to be accounted for: i) it adds a per-input computational complexity; ii) the energy consumed by a CDNN for processing each input data is nearly proportional to the overall number of numerical operations carried out by the layers, as reported by results of the field trials in [25]; and iii) in some cases, the increment in the overall accuracy of the CDNN gained by the placement of an additional local classifier may be offset by the corresponding increment of the per-input consumed energy [24].
Hence, motivated by these considerations, task T1.1 focuses on the companion topics of the optimized setting of a suitable number of local classifiers and their optimized placement over the layer stack of the considered baseline DNN. As all the problems of optimal placements involving discrete variables (number of local classifiers and their allowed locations), even the optimal placement of the local classifier is an NP-hard discrete combinatorial problem, whose globally optimal solution resists closed-form evaluation and exhibits an exponential-type solving complexity. Therefore, main goal of T1.1 is to develop a greedy-type solving approach that is locally optimal and exhibits a computational complexity which scales in a linear way with the depth of the considered CDNN.