Accurate mobility prediction has become pivotal for a wide range of smart city applications including optimizing electric vehicles (EV) charging management, traffic management, infrastructure planning, etc. However, traditional mobility prediction models face significant challenges including ineffective integration of geographical information, inability to dynamically adapt to changing location popularity, and struggle with long-term dependency. Furthermore, these models are susceptible to catastrophic forgetting, losing previously learned knowledge when exposed to new data.To overcome these challenges, we propose the Lifelong Edge-cloud Adaptive Prediction (LEAP) framework, a fresh approach that integrates lifelong learning into mobility prediction. LEAP improves prediction accuracy and reliability in dynamic real-world environments by fusing a central cloud model for capturing long-term trends with multiple local edge models that process real-time data. LEAP employs Spatially Adaptive LSTM (SA-LSTM) and Temporal Adjustment LSTM (TA-LSTM) to incorporate dynamic spatio-temporal patterns, along with Global Context Operations (GC Ops) to manage long-term dependencies. To prevent catastrophic forgetting, LEAP uses Learning without Forgetting (LwF), enabling on-device continuous learning and adaptation at the edge.Extensive evaluations demonstrate that LEAP surpasses ten state-of-the-art methods, including Scikit-Learn’s LSTM, with average improvements of 2.61x in Recall-5, 2.35x in Recall-10, 2.80x in NDCG-5, and 3.06x in NDCG-10. These results highlight LEAP’s superior effectiveness, accuracy, and adaptability, proving it a worthy choice for dynamic real-world mobility prediction tasks while effectively addressing catastrophic forgetting.

As machine learning moves towards edge deployment, lifelong learning becomes crucial due to evolving data distributions and new tasks. Yet, applying traditional methods to learn from vast, complex IoT data streams poses challenges. These include excessive CPU usage, RAM overflow, prolonged convergence times disrupting device operation, and difficulties in adapting to concept drift. Consequently, models trained on devices struggle to handle frequently changing data, affecting their ability to respond effectively to new inputs.To address these issues, we introduce Poly Instance Lifelong Learning (PILL), an algorithm designed for real-time on-device model training and inference at the edge under lifelong learning settings. PILL is lightweight, operating efficiently on the CPUs of low-power single-board computers (SBCs). It achieves this by partitioning input data into manageable instances, filtering out label noise, and applying early stopping for rapid predictions.PILL was evaluated on three popular low-power SBCs as well as a high-end Windows 10 machine using four datasets of different sizes and features. The results indicate that despite the superior resources of the Windows 10 machine, models trained using PILL on SBCs differ in accuracy by only ±0.05%. Additionally, PILL’s LSTM trains 2.41 - 2.85 times faster than the widely used Scikit-Learn’s LSTM. Additionally, when compared to ten state-of-the-art methods, PILL demonstrated superior performance across key metrics (Precision, Recall, and F1-Score) while minimizing computational overhead, making it an ideal choice for efficient, real-time edge deployment.

Multi-class classifier training using traditional meta-algorithms such as the popular One-vs-One (OvO) method may not always work well under cost-sensitive setups. Also, during inference, OvO becomes computationally challenging for higher class counts K as O(K^2) is its time complexity. In this paper, we present Opt-OvO, an optimized (resource-friendly) version of the One-vs-One algorithm to enable high-performance multi-class ML classifier training and inference directly on microcontroller units (MCUs). Opt-OvO enables billions of tiny IoT devices to self learn/train (offline) after their deployment, using live data from a wide range of IoT use-cases. We demonstrate Opt-OvO by performing live ML model training on 4 popular MCU boards using datasets of varying class counts, sizes, and feature dimensions. The most exciting finding was, on the 3 $ ESP32 chip, Opt-OvO trained a multi-class ML classifier using a dataset of class count 50 and performed unit inference in super real-time of 6.2 ms.

Invariable of the agriculture type (precision, smart, or digital), the monitoring process of factors that increase the crop yield and growth is mostly non-ML, manually structured approaches with practical pain points. In this scenario, to reduce monitoring costs and maintenance efforts, there is a requirement for low-cost semi-autonomous distributed systems that can remotely collect plant data and perform standalone ML-based analytics without depending on cloud servers or the internet. In this work, we provide an embedded ML pipeline, which users can use/follow for end-to-end solution design and implementation for any of their use-cases. To demonstrate the pipeline, we use it to collect image data, train a CNN-based regression algorithm, perform hardware-specific tuning, generate optimized code, and deploy binaries on Sony Spresense setup. The initial testing shows that even the resource-constrained MCU-based Spresense, in real-time (992 ms), high performance (96.2% accuracy, 1.86 cm2RMSE), could analyze a plant in a semi-autonomous environment to predict the leaf area and plant growth.

Traditionally, original equipment manufacturers (OEMs) send device-specific over-the-air (OTA) packages to ensure the latest firmware, security patches, etc. With millions of IoT devices, even a tiny per-centage of OTA failures will result in tens of thousands of globally affected consumers. The state-of-the-art OTA methods are suited for high-end Android & embedded Linux devices and not for resource-constrained devices (e.g. MCUs, small CPUs) with only a few MB memory. Current OTA methods have been tested only on non-ML use-cases such as remote bugs patching or security flaws, etc. In this paper, we present OTA-TinyML approach that, via HTTPS, loads the C source file of ML models from a webserver into IoT boards. OTA-TinyML does not strain hardware (resource-friendly) as its implementation spans only a few lines of code. It is compatible with a range of ML models (e.g. text, speech, image domains) and MCUs (e.g. Cortex M series, STM32, Xtensa). OTA-TinyML is tested by performing remote fetching of 6 types of ML models, storing them on 4 types of memory units, then loading and executing on 7 popular Arduino IoT boards.

TinyML: Tiny in size, big in impact! In this paper, we present a Tensor Memory Mapping (TMM) method, which can accurately calculate the on-device execution memory consumed by a range of ML and TinyML models during execution on small central processing units (CPUs), microcontroller units (MCUs), and single board computers (SBCs).

TinyML: Tiny in size, big in impact! This paper presents TinyML-CAM pipeline for real-time and memory-efficient image recognition on IoT boards. TinyML-CAM can be used by developers to implement their customized tasks in ≈ 30 minutes, with minimal code configuration. We evaluated TinyML-CAM by using it to create a RandomForestClassifier (RF) based real-time image recognition system, which ran at 80 FPS and consumed only 1 kB of RAM on ESP32.

The majority of IoT edge devices are embedded systems with a tiny microcontroller unit (MCU), which acts as its brain. When users want their edge devices to continuously improve for better edge-analytics results, there is a need to equip their devices with algorithms that can learn/train from the continuously evolving real-world data. Currently, such devices are not capable of executing any machine learning (ML)-based model training tasks due to their resource constraints such as: limited memory (SRAM, Flash, and EEPROM), low operations per second, its inability to perform parallel processing, etc. In this article, we provide ML-MCU, a framework with our novelOptimized-Stochastic Gradient Descent (Opt-SGD)andOptimized One-Versus-One (Opt-OVO)algorithms to enable both binary and multiclass ML classifier training directly on MCUs. Thus,ML-MCUenables billions of MCU-based IoT edge devices to self learn/train (offline) after their deployment, using live data from a wide range of IoT use cases. When evaluating our algorithms on multiple popular MCUs, using various data sets of different sizes and feature dimensions, one of the most exciting findings was, ourOpt-OVOalgorithm trained a multiclass classifier using a data set of class count 50, on a$\$ $3 resource-constrained MCU and also performed onboard unit inference for the same 50 class data in super real time (6.2 ms).

Typically a Neural Networks (NN) is trained on data centers using historic datasets, then a C source file (model as a char array) of the trained model is generated and flashed on IoT devices. This standard process impedes the flexibility of billions of deployed ML-powered devices as they cannot learn unseen/fresh data patterns (static intelligence) and are impossible to adapt to dynamic scenarios. Currently, to address this issue, Online Machine Learning (OL) algorithms are deployed on IoT devices that provide devices the ability to locally re-train themselves -continuously updating the last few NN layers using unseen data patterns encountered after deployment.In OL, catastrophic forgetting is common when NNs are trained using non-stationary data distribution. The majority of recent work in the OL domain embraces the implicit assumption that the distribution of local training data is balanced. But the fact is, the sensor data streams in real-world IoT are severely imbalanced and temporally correlated. This paper introduces Imbal-OL, a resource-friendly technique that can be used as an OL plugin to balance the size of classes in a range of data streams. When Imbal-OL processed stream is used for OL, the models can adapt faster to changes in the stream while parallelly preventing catastrophic forgetting. Experimental evaluation of Imbal-OL using CIFAR datasets over ResNet-18 demonstrates its ability to deal with imperfect data streams, as it manages to produce high-quality models even under challenging learning settings.

The concept of ML model aggregation rather than data aggregation has gained much attention as it boosts pre- diction performance while maintaining stability and preserving privacy. In a non-ideal scenario, there are chances for a base model trained on a single device to make independent but complementary errors. To handle such cases, in this paper, we implement and release the code of 8 robust ML model combining methods that achieves reliable prediction results by combining numerous base models (trained on many devices) to form a central model that effectively limits errors, built-in randomness and uncertainties. We extensively test the model combining performance by performing 15 heterogeneous devices and 3 datasets based experiments that exemplifies how a complicated collective intelligence can be derived from numerous elementary intelligence learned by distributed, ubiquitous IoT devices.

Transmitting updates of high-dimensional models between client IoT devices and the central aggregating server has always been a bottleneck in collaborative learning - especially in uncertain real-world IoT networks where congestion, latency, bandwidth issues are common. In this scenario, gradient quantization is an effective way to reduce bits count when transmitting each model update, but with a trade-off of having an elevated error floor due to higher variance of the stochastic gradients. In this paper, we propose ElastiCL, an elastic quantization strategy that achieves transmission efficiency plus a low error floor by dynamically altering the number of quantization levels during training on distributed IoT devices. Experiments on training ResNet-18, Vanilla CNN shows that ElastiCL can converge in much fewer transmitted bits than fixed quantization level, with little or no compromise on training and test accuracy.