In many applications, regulations or best practices often lead to specific requirements in machine learning relating to four key pillars: fairness, privacy, interpretability and greenhouse gas emissions. These all sit in the broader context of sustainability in AI, an emerging practical AI topic. However, although these pillars have been individually addressed by past literature, none of these works have considered all the pillars. There are inherent trade-offs between each of the pillars (for example, utility vs fairness or utility vs privacy), making it even more important to consider them together. This paper outlines a new framework for Sustainable Machine Learning. It proposes FPIG, a general AI pipeline that allows for simultaneous consideration and a better understanding of the tradeoffs between the pillars. Based on the FPIG framework, we propose a meta-learning algorithm to estimate the four key pillars given a dataset summary, model architecture, and hyperparameters before model training. This algorithm allows users to select the optimal model architecture for a given dataset and a set of user requirements on the pillars. We illustrate the trade-offs under the FPIG model on three classical datasets and demonstrate the meta-learning approach with an example of real-world datasets and models with different interpretability, showcasing how it can aid model selection.

We describe the Customer LifeTime Value (CLTV) prediction system deployed at ASOS.com, a global online fashion retailer. CLTV prediction is an important problem in e-commerce where an accurate estimate of future value allows retailers to effectively allocate marketing spend, identify and nurture high value customers and mitigate exposure to losses. The system at ASOS provides daily estimates of the future value of every customer and is one of the cornerstones of the personalised shopping experience. The state of the art in this domain uses large numbers of handcrafted features and ensemble regressors to forecast value, predict churn and evaluate customer loyalty. Recently, domains including language, vision and speech have shown dramatic advances by replacing handcrafted features with features that are learned automatically from data. We detail the system deployed at ASOS and show that learning feature representations is a promising extension to the state of the art in CLTV modelling. We propose a novel way to generate embeddings of customers, which addresses the issue of the ever changing product catalogue and obtain a significant improvement over an exhaustive set of handcrafted features.

The Pulse-Coupled Oscillator (PCO) is a novel protocol inspired by models used in mathematical biology to justify the emergence of synchrony in the natural world. Our paper introduces and demonstrates the efficacy of a new PCO protocol implementation that, by disabling all collision resolution mechanisms for a suitable portion of the node operations, lets the rapid establishment of a common clock and its maintenance. The key idea is to allow signals to be superimposed in time, a feature that is absent in previous implementations, because it is prevented by traditional medium access schemes. We map the PCO protocol into an event-driven asynchronous coloring algorithm, based on the local exchange of information to explain its convergence properties. The event-based description of the PCO protocol sets the stage for our experimental comparison with a competing decentralized network synchronization approach, namely, the Reference Broadcast Protocol (RBS). For comparison, we combined RBS with an asynchronous average consensus protocol, running exactly on the same MicaZ platforms. The experimental results showcase the better scalability of the PCO scheme compared to the competing method based on RBS, proving that the PCO primitive is a reasonable option to consider for wireless sensor network applications.

We study a distributed protocol for time-division multiple access (TDMA) with applications in wireless body area networks. Our scheme is derived from a bio-inspired algorithm known as Pulse-Coupled Oscillator, in which each node in the network update its state based on its knowledge of the neighborhood. The protocol we propose is suitable for negotiation of a shared resource in a distributed manner, and is adaptive to changes in the network size or in the nodes' demands. We discuss the implementation of such scheme from an application perspective, and discuss the impact of the physical layer on its performance.

In recent years, several models introduced in mathematical biology and natural science have been used as the foundation of networking algorithms. These bio-inspired algorithms often solve complex problems by means of simple and local interactions of individuals. In this work, we consider the development of decentralized scheduling in a small network of self-organizing devices that are modeled as pulse-coupled oscillators (PCOs). By appropriately designing the dynamics of the PCO, the network of devices can converge to a desynchronous state where the nodes naturally separate their transmissions in time. Specifically, by following Peskin's PCO model with inhibitory coupling, we first show that round-robin scheduling can be achieved with weak convergence, where the nodes' transmissions are separated by a constant duration, but the differences of their local clocks continue to shift over time. Then, by having each node accept coupling only from the pulses emitted by a subset of neighboring nodes, we show that it is possible to achieve strict desynchronization, where the difference between local clocks remain fixed over time. More interestingly, by having each node maintain two local clocks, we show that it is possible to further achieve proportional fair scheduling, where the time alloted to each node is proportional to their demands. The convergence of these algorithms is studied both analytically and numerically.

We consider the problem of coordinating a group of mobile nodes communicating through a wireless medium. The objective of the network is the alignment of all the nodes towards a common direction through local interactions, without the need for global knowledge such as the network topology or the maximum degree of the network, or even local parameters, such as the number of neighbors. The key feature of our algorithm is that each node state update is done through voting, where the probability of each vote is biased by the state of the node neighbors. We propose two possible physical implementations for our algorithm. The first is based on the explicit exchange of packetized messages, while the second is a cross-layer approach. Our analysis unveils key convergence properties of this simple class of alignment algorithms, via analytical and simulated results.

Pulse coupled oscillators (PCOs) are pulsing devices that pulse individually in a periodic manner but alter their pulsing patterns in response to the pulsing of other nodes. A network of PCOs can produce a number of different dynamics from their pulsing activities, among which the synchrony of pulsing is perhaps the most well known. In this paper, we study the primitive that falls into the class of ldquodesynchronizationrdquo. Specifically, we propose a simple pulse-coupling mechanism that allows each node in the network to converge to a desynchronized state where the nodes will pulse periodically with a constant spacing among each others firing times. We discuss the convergence of the PCO mechanism and propose to apply this primitive to resolve contention in the reservation phase of a reservation-based MAC protocol.

Sensor networks have been used in a wide range of applications. Considerable research effort is currently devoted to design protocols that allow networks of inexpensive sensors to perform reliable remote control and monitoring functions, in spite of the limitations of each device.Our objective is to demonstrate a fully software implementation of the so called Pulse Coupled Oscillator Protocol, proposed in [2, 6] for the decentralized synchronization of radio devices. The PCO protocol is inspired by a model found in mathematical biology [5]. It is decentralized, scalable and simple, as we will show in our demonstration.