Modern machine learning techniques focus on extremely deep and multi-pathed networks, resulting in large memory and computational requirements. This thesis explores techniques for designing efficient convolutional networks including pixel shuffling, depthwise convolutions, and various activation fucntions. These techniques are then applied to two image processing domains: single-image super-resolution and image compression. The super-resolution model, TinyPSSR, is one-third the size of the next smallest model in literature while performing similar to or better than other larger models on representative test sets. The efficient deep image compression model is significantly smaller than any other model in literature and performs similarly in both computational cost and reconstruction quality to the JPEG standard.

TinyML är ett snabb växande tvärvetenskapligt område i maskininlärning. Den fokuserar på att möjliggöra maskininlärnings algoritmer på inbyggda enheter (mikrokontroller) som arbetar vid lågt effektområde. Syftet med denna studie är att analysera hur bra TinyML kan är lösa typiska ML-uppgifter. Studien hade fyra forskningsfrågor som svarades genom att undersöka olika litteraturstudier och implementera testmodell både på laptop och på inbyggda enheter (Arduino nano 33). Implementationen började med att skapa maskininlärningsmodell i form av sinusfunktion genom att skapa ett 3- lagers, fullt anslutet neuralt nätverk som kan förutsäga sinusfunktionens utdata, på detta sätt används modellen som en regressionsanalys. Idéen är att träna modellen som accepterar värden mellan 0 och 2π och sedan matar ut ett värde mellan -1 och 1. Därefter konverteras modellen till en Tensorflow Lite för att kunna distribuera den på Arduino nano 33. Resultatet visade att TinyML är bra lösning för att lösa ML-uppgifter eftersom det lyckades överföra ML-algoritmen till mikrokontrollen Arduino nano 33. TinyML kunde hantera och bearbeta data utan behov till internetanslutning vilket gav möjlighet för utvecklare att programmera på ett effektivt och lämpligt sätt. TinyML verkar ha en ljus framtid och många vetenskapliga studier påpekar att maskininlärningens största fotavtryck i framtiden kan vara genom TinyML. / TinyML is a rapidly growing interdisciplinary field in machine learning. They focus on enabling machine learning algorithms on built-in devices (microcontrollers) that work at low power ranges. The purpose of this study is to analyze how well Tiny-ML can solve typical ML tasks. The study had four research questions that were answered by examining different literature studies and implementing test model both on laptop and on built-in devices (Arduino nano 33). The implementation began with creating a machine learning model in the form of a sine function by creating a 3-layer, fully connected neural network that can predict the output of the sine function, in this way the model is used as a regression analysis. The idea is to train the model that accepts values between 0 and 2π and then outputs a value between -1 and 1. Then the model is converted to a Tensorflow Lite to be able to distribute it on the Arduino nano 33. The results showed that TinyML is a good solution for solving ML data, as they managed to transfer the ML algorithm to the microcontroller Arduino nano 33. They could handle and process data without the need for an Internet connection, which allowed developers to program, anywhere and anytime any. TinyML seems to have a bright future and many scientific studies point out that the biggest footprint of machine learning in the future may be through TinyML.

As an essential and indispensable contributor to pollinating the world's crops and plants, the honey bee is key to the sustainability of humans' and our ecosystems' continued survival. Following in the footsteps of the companies TietoEvry and Beelabs project, this report also works towards monitoring bees during their daily activities. This project aims to investigate the feasibility of using wireless, battery-driven devices inside beehives to detect the sound of bees using machine learning for edge devices. Beelab has focused on measurements in and around the beehive regarding weight, temperature, barometric pressure and humidity. Sound analysis is still in its infancy with few finished working alternatives; therefore, this project will focus on the sound attribute by implementing machine learning and classification algorithms and applying it to a prototype—the progress is thoroughly documented in this report. The device records a snippet of sound and prepares to send it over a wireless transmission medium. By streamlining the code and optimizing the hardware, the device runs continuously for a month using a small, cheap battery.

Lite-Agro studies applications of TinyML in pear (Pyrus communis) tree disease identification and explores hardware implementations with an ESP32 microcontroller. The study works with the DiaMOS Pear Dataset to learn through image analysis whether the leaf is healthy or not, and classifies it according to curl, healthy, spot or slug categories. The system is designed as a low cost and light-duty computing detection edge solution that compares models such as InceptionV3, XceptionV3, EfficientNetB0, and MobileNetV2. This work also researches integration with federated learning frameworks and provides an introduction to federated averaging algorithms.