Introduction:

As a developer, have you ever faced the challenges that come with building and deploying software on traditional platforms? I can relate to that. However, when I discovered the Internet Computer protocol, my workload was significantly reduced

About us:

The Internet Computer is a Blockchain that enables all range of people, from developers to entrepreneurs and organisations to deploy secure, autonomous and tamper-proof canisters (commonly referred to as smart contracts)

We provide the following key features:

1. A public blockchain that is globally accessible for running smart contracts at web speed

2. A secure protocol, known as Internet Computer Protocol, which is run by nodes that are operated by Independent node providers in data centres worldwide. This ensures that smart contracts are executed

3. A network of blockchains connected using chain key cryptography, thus allowing for scalability

What makes us different from the other blockchain protocols?

At Internet Computer Protocol, we simplify the development process by reducing the complexities (i.e network configuration, load balancing, firewalls, database maintenance and storage management) that developers face when building and deploying software on traditional platforms.

We also eliminate these risks and distractions that come with developing applications on the traditional platforms.

Therefore, developers can focus on writing code using canisters (or smart contracts) without having to worry about these risks.

This allows them to streamline the development process, reduce time it takes for a product to go to market and thus encourage innovation.

How to get started with developing on our protocol:

Getting started with development on the Internet Computer Protocol is a breeze. If you have experience with Typescript or Javascript, you’re in luck!

Our blockchain platform offers a beginner-friendly and fully remote course that will teach you everything you need to know. Click here to access the course

The course is completely free, self-paced, and can be completed in less than two hours. It is the perfect way to dive into the world of blockchain development and start building your own decentralised applications!

Follow us on twitter to get the latest updates on events and hackathons we’re participating and attending:

Twitter @ICPKushites (https://twitter.com/ICPKushites)

Happy learning!

Meet Our Partner: Kushite ICP was originally published in GDG NAIROBI on Medium, where people are continuing the conversation by highlighting and responding to this story.

From art to state-of-the-art Image Generation

In recent months, there has been a lot of excitement around diffusion models in the field of generative AI. These models have proven to be more powerful than traditional methods like GANs and VAEs! Popular examples include OpenAI’s DALL-E 2 and StabilityAI’s Stable Diffusion.

These models also have impressed with their ability to create high-quality images using a new probabilistic approach and have garnered significant attention for their exceptional performance.

In this article, we delve into the concept of stable diffusion, explore its key components, and demonstrate how to unlock the full potential of image generation using Stable Diffusion in KerasCV.

Let’s explore the unlimited possibilities. But first, let’s understand what they are…

Diffusion Models 101

Diffusion models are generative models designed to produce data similar to the training dataset. These models operate by following a Markov chain of diffusion steps, where they gradually add random noise to the data and then learn to reverse this diffusion process to reconstruct desired data samples from the noise. The training phase involves training the model on a dataset, often consisting of images, by applying denoising algorithms to learn the inverse mapping from the noisy data to the original data.

In simpler terms, diffusion models generate data by first introducing Gaussian noise to the training data and then recovering the original data by undoing the noise. The model learns the intricate relationships between the noisy input and the clean output, allowing it to generate high-quality images that resemble the input data.

By leveraging the learned denoising process, diffusion models can produce diverse and realistic images that closely resemble the data on which they were trained.

Who cares?

To paraphrase Andreessen Horowitz, generative AI, particularly on the text-to-art side, is eating the world. Investors have demonstrated their confidence in this technology by pouring billions of dollars into startups focused on building these systems.

One notable advancement in this field is OpenAI’s Dall-E 2. Introduced in April 2022, Dall-E 2 represents a major milestone as the first mainstream text-to-image model. Microsoft recently partnered with OpenAI for their Image Creator project. Despite its cool capabilities, the model remained closed-source, limiting its accessibility to the wider community.

In a parallel development, Google unveiled its own diffusion-based image generation model called Imagen, just a month after the release of Dall-E 2.

Imagen now part of Vertex AI

It is worth noting that Google has integrated Imagen into Vertex AI, their fully managed AI service. This announcement was made at the Google I/O developer conference this year.

The other one is Stable Diffusion created by StabilityAI. Stable Diffusion stands out as the sole open-source diffusion-based image generation model in this list.

Stable Diffusion! What is it?

It is essentially a text to image generative model. It utilizes text descriptions to generate images and perform image-to-image translations guided by text prompts.

The architecture of Stable Diffusion consists of three main components. First, there is a text encoder that takes the text prompt as input and converts it into a latent vector — computer-readable vector representation.

That prompt vector is then concatenated to a randomly generated noise patch through a process called Conditioning.

The second component is a U-Net diffusion model responsible for generating images. This diffusion model operates in the latent space and focuses on denoising 64x64 latent image patches. By repeatedly refining these patches, the model progressively enhances the generated images.

The third component is a decoder that renders the final 64x64 latent patch into a higher-resolution 512x512 image patch.

And how does the Model recover information?

Probably, this is the question you are now asking. To recover information, the model leverages the training data distribution and uses it to predict the visual details that are most likely to be present in the given input. By utilizing the knowledge learned from the training data, the model generates images that align with the expected visual characteristics. Therefore, it’s possible to train a deep learning model to denoise an input image — and thereby turn it into a higher-resolution version. The process is called super-resolution.

What happens if we Push super-resolution to the limit

You may start asking — what if we just run such a model on pure noise? The model would then “denoise the noise” and start hallucinating a brand-new image. With each iteration, this process gradually transforms a small patch of noise into a increasingly clear and high-resolution artificial picture.

And how does KerasCV comes in

Stable Diffusion has been integrated into Keras, allowing users to generate novel images in a few lines of code while enjoying the powerful performance boosts and benefits that comes with Keras.

So, how can you use it (as an engineer) — Basic

You simply install Keras package and sort out some imports:

pip install tensorflow keras_cv --upgrade --quiet

import time
import keras_cv
from tensorflow import keras
import matplotlib.pyplot as plt

And instantiate the stable diffusion model.

model = keras_cv.models.StableDiffusion(img_width=512, img_height=512)

And get creative with the text prompts. It will take about 4 seconds to give you an output…

images = model.text_to_image("A fisherman on Lake Victoria, Kenya, with a t-shirt written Safaricom.", batch_size=3)


def plot_images(images):
    plt.figure(figsize=(20, 20))
    for i in range(len(images)):
        ax = plt.subplot(1, len(images), i + 1)
        plt.imshow(images[i])
        plt.axis("off")


plot_images(images)

Let’s try a more complex prompt:

images = model.text_to_image(
    "cute magical flying dog, fantasy art, "
    "golden color, high quality, highly detailed, elegant, sharp focus, "
    "concept art, character concepts, digital painting, mystery, adventure",
    batch_size=3,
)
plot_images(images)

Struggling to come up with prompts? You could get some from https://lexica.art/

How to use it as (an engineer )— Advanced

Lets now take a look at advanced usage, both for image generation and inpainting. Here is the Colab notebook (coming soon).

Why would you want to use KerasCV ?

KerasCV provides several compelling reasons to utilize its framework. Firstly, it employs graph mode execution, which enhances performance by leveraging graph optimization and enabling parallelism.

Secondly, KerasCV supports XLA compilation through the option of jit_compile=True, utilizing the powerful XLA compiler provided by TensorFlow. XLA is specifically designed for accelerated linear algebra operations, further boosting performance.

KerasCV also embraces mixed precision computation, a technique that combines 16-bit and 32-bit floating-point types during model training. This approach strikes a balance between speed and memory usage. By performing computations using float16 precision while storing weights in the float32format, KerasCV harnesses the faster float16 kernels available on modern NVIDIA GPUs. This achieves significant performance improvements without sacrificing accuracy.

Lastly, despite these advanced optimizations, KerasCV still maintains a user-friendly interface.

Let’s see the bench marks — how does KerasCV compare to other implementations?

Here is a benchmark comparing the runtime performance of KerasCV implementation against HuggingFace diffusers. It is conducted using a Tesla T4 GPU. The benchmark consisted of generating three images with a step count of 50 for each image.

The benchmark was divided into two parts: warm start and cold start. The cold start considered the time it takes to create and compile the model, which may not be as important in a real-world setting.

The results showed a substantial 30% improvement in execution time on the Tesla T4 GPU. However, the improvement on the V100 GPU was not as significant.

The benchmark is expected to consistently favor the KerasCV implementation across all NVIDIA GPUs in terms of performance.

But that is not all — Other possibilities

Diffusion models can generate videos directly from text prompts. This allows for the creation of dynamic visual content based on narratives, songs, poems, and other textual sources.

Here is demo of video generated by KerasCV and stable diffusion given a song lyric and timestamp.

Closing remarks

KerasCV’s implementation of Stable Diffusion stands at the forefront of the field. With the integration of XLA and mixed precision techniques, it provides the fastest and most efficient Stable Diffusion pipeline currently available.

Whether you’re an artist looking to craft mesmerizing visual masterpieces or a researcher seeking to push the boundaries of generative AI, KerasCV’s Stable Diffusion opens up a world of possibilities.

So, why not take the leap and explore? The journey awaits!

If you found this helpful, click the 👏. Follow this publication for more tutorials on Machine Learning and Cloud Data (Internet of Things).

NB-If you have any Machine Learning concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can

Advancements in Machine Learning and AI

Google I/O Connect Amsterdam 2023 brought together innovators and enthusiasts to explore the latest advancements in different technologies including Machine learning (ML) and artificial intelligence (AI). The event showcased several intriguing announcements and highlighted the innovative developments unveiled by Google. In this article, we will delve into some of the key insights and takeaways from the event.

Advancements in Machine Learning with TensorFlow and Keras

The TensorFlow team showcased their commitment to supporting large models with the introduction of DTensor parallel processing. This advancement enables the training and processing of very large models, opening up new possibilities for complex AI applications. Additionally, JAX2TF was introduced to make it easier to use JAX models within the Tensorflow ecosystem, expanding the range of models available for developers.

Simplifying ML with KerasCV and KerasNLP

Keras, a popular deep learning library, unveiled new APIs called KerasCV and KerasNLP, specifically designed to simplify computer vision and natural language processing tasks.

Laurence Moroney demonstrated how these APIs can be used to build generative AI solutions for images and texts, empowering developers to create sophisticated ML models with ease.

Simple ML for Google Sheets

Do you love catching numbers in spreadsheets? Do you ever wish to use AI to analyze your spreadsheet data? The Simple ML team unveiled an extension for Google Sheets, enabling users to leverage AI technology for analyzing spreadsheet data.

With features such as forecasting, anomaly detection, and the ability to recover missing values, users can leverage AI capabilities directly within their spreadsheets. The extension also provides advanced controls for running ML tasks, including model training, evaluation, and analysis.

Discovering Pre-Trained Models with Kaggle Models

Pre-trained models are an important part of modern ML workflows, but finding the right pre-trained models for specific ML tasks can be challenging.

To address this, Kaggle launched a dedicated hub that allows users to discover, evaluate, tweak, and test hundreds of open-source models developed by Google and other leading researchers. With over 2,000 pre-trained models organized by task, such as text classification and object detection, researchers and developers now have a valuable resource to accelerate their ML workflows.

PaLM API & MakerSuite

Large Language Models (LLMs) are capturing the imaginations of people around the world. Google introduced the PaLM API, offering developers an easy and safe way to build generative AI applications utilizing their best language models.

PaLM API is complemented by MakerSuite, an intuitive tool that allows developers to quickly prototype ideas by providing features such as prompt engineering, synthetic data generation, and custom model tuning, all backed by robust safety tools.

The PaLM API and MakerSuite together provide access to the latest LLMs hosted in the cloud, empowering developers to leverage state-of-the-art language models.

Scaling LLMs into Google Cloud — Vertex AI

Machine Learning in Google Cloud begins with Vertex AI. Vertex AI provides ML tools, services, workflows and infrastructure all accessible through a simple and unified platform.

Vertex AI expanded its support for Generative AI with two new products: Model Garden and Generative AI Studio.

Model Garden provides access to Google’s latest AI foundational models, including PaLM, and offers a variety of machine learning APIs for tasks such as text, image, and code analysis.

Generative AI Studio simplifies the exploration, prototyping, and customization of generative AI models, enabling developers to create unique applications.

MediaPipe solution for on-Device ML

The MediaPipe team introduced updates to their solution, enhancing its usability across platforms, customization options, and performance.

MediaPipe provides a high-output, low-latency ML pipeline, enabling developers to incorporate on-device ML capabilities into their applications seamlessly.

Duet AI for Google Cloud

There was also Duet AI for Google Cloud, an invaluable and ever-present AI collaborator designed to assist users of all proficiency levels.

Duet AI offers remarkable features including code assistance and chat assistance, which collectively revolutionize the cloud development experience.

With code assistance, Duet AI leverages AI-driven algorithms to provide real-time support to cloud users, including application developers and data engineers. As users type their code, Duet AI offers intelligent recommendations, ensuring accuracy and efficiency throughout the development process.

It goes beyond mere suggestions and even generates complete functions and code blocks, significantly reducing coding effort. Duet AI also actively identifies vulnerabilities and errors within the code, offering insightful suggestions for fixes.

Duet AI’s chat assistance empowers users to seek guidance through simple, natural language queries. This feature serves as a reliable resource for resolving specific development or cloud-related queries in real-time.

Users can engage with chat assistance to receive immediate guidance on diverse topics, ranging from mastering the utilization of specific cloud services and functions to obtaining detailed implementation plans for their cloud projects.

Furthermore, chat assistance offers architectural insights and coding best practices, eliminating the need for extensive research and document searches.

Safe & Responsible AI

The rapid development of AI presents immense opportunities to address complex real-world problems. However, it also raises important questions regarding the responsible development of AI systems that benefit all individuals and communities.

Google recognizes the importance of responsible AI development and advocates for fairness, interpretability, privacy, and security as key considerations.

By upholding these principles, the development and deployment of AI can deliver positive and equitable outcomes while addressing the challenges and questions raised by this transformative technology.

The Google I/O Connect Amsterdam event showcased groundbreaking advancements in cloud and Machine Learning technologies. As we reflect on the highlights, it becomes evident that the future of AI and cloud development is poised for remarkable transformations.

Stay tuned for upcoming articles where I will delve deeper into the technical concepts and provide comprehensive guidance!

If you found this helpful, click the 👏. Follow this publication for more tutorials on Cloud Data (Internet of Things) and AI.

NB-If you have any Machine Learning concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can.

Transfer Learning in Keras

Deep learning has revolutionized the field of artificial intelligence and data science, enabling us to tackle complex problems in various domains. One of the key techniques within deep learning is transfer learning, which allows us to leverage pre-trained models to solve new tasks more efficiently.

But first what is deep learning?

It is basically a branch of machine learning that focuses on training neural networks with multiple layers to extract patterns and representations from data.

Its applications span across a wide range of fields, including robotics, computer vision, natural language processing, image recognition, and more. Deep learning models excel at tasks such as classification, where the goal is to predict a categorical label based on input features.

Deep Learning 101 Concepts

Classification is a supervised machine learning algorithm used to predict categorical labels. It finds applications in various real-world scenarios, such as predicting customer churn, classifying emails as spam or non-spam, and determining whether a bank loan will default or not.

Through classification, we can make accurate predictions and decisions based on input data.

Computer Vision

Computer vision plays a crucial role in the implementation of deep learning classification models. Computer vision techniques are employed to preprocess and extract meaningful features from images.

One Hot Encoding

In classification tasks, it is common to represent categorical labels as one-hot encoded vectors. One hot encoding is a technique used to convert categorical variables into a binary vector representation. Each category is assigned a unique index, and the corresponding element in the one-hot encoded vector is set to 1, while all other elements are set to 0.

This representation allows the neural network to understand the distinct categories and make predictions accordingly.

Neural network

Neural networks are the building blocks of deep learning models, including classifiers. A neural network classifier consists of multiple layers of interconnected neurons. Each neuron performs a computation by taking the weighted sum of its inputs, adding a bias value, and passing the result through an activation function.

These weights and biases are initially unknown and randomly initialized. Through the training process, the neural network learns to adjust these weights and biases by analyzing large amounts of labeled data.

Dense Neural Network

This is the simplest neural network for classifying images. It is made of “neurons” arranged in layers. The first layer processes input data and feeds its outputs into other layers. It is called “dense” because each neuron is connected to all the neurons in the previous layer.

You can feed an image into such a network by flattening the RGB values of all of its pixels into a long vector and using it as inputs.

Activation function (Softmax)

Activation functions play a crucial role in neural networks by introducing non-linearity and enabling the model to capture complex patterns. Common activation functions include relu (rectified linear unit) and softmax. The last layer of a classification model typically uses softmax activation with the same number of neurons as the classes to predict the probabilities for each class. Cross-entropy loss is commonly used in classification tasks to compare the predicted probabilities with the actual one-hot encoded labels.

Cross-entropy loss

For classification, cross-entropy is the most commonly used loss function, comparing the one-hot encoded labels (i.e. correct answers) with probabilities predicted by the neural network.

Reducing Loss

To minimize the loss and improve the accuracy of the model, we employ optimization algorithms such as gradient descent. Gradient descent adjusts the weights and biases of the neural network based on the gradients of the loss function. Optimizers like AdamOptimizer with momentum are commonly used due to their efficiency and ability to converge to better solutions. Training on batches of data, known as mini-batching, further enhances the optimization process.

Transfer Learning

Transfer learning is a technique that leverages the knowledge acquired from pre-trained models on one task to improve the performance of a related task. Instead of training a deep learning model from scratch, we can use a pre-trained model as a starting point and fine-tune it for our specific problem. Transfer learning offers several advantages, including reduced training time, improved generalization, and the ability to achieve good results even with limited labeled data.

Implementing Transfer Learning in Keras

Demo:

In this guide, we have built Classification models using the deep learning framework, Keras

By incorporating pre-trained models into our pipeline, we can benefit from their learned representations and expedite the training process.

Embedded Learning Challenge Comes to Kigali, Rwanda

The Internet of things, IoT is transforming how we work, live and relate. Today we have devices at home such as mobile phones, TV, remote control that are either connected to each other or to the internet. There has been an extension of connectivity beyond the mobile phones and laptops to more devices and systems enabling new levels of automation and control.

To develop powerful, connected devices, a range of technologies such as Embedded Systems, Tiny ML and Tensorflow Lite needs to be combined. Together, they can enable an IoT Device to collect, process, and transmit data, and make decisions based on that data even on the device.

Recently, we held an event at FabLab Rwanda that brought together experts in IoT, Tiny ML, embedded systems, and TensorFlow Lite to discuss the latest developments and opportunities in these fields.

Arm Engage and ELC

The event was well attended, with a diverse group of participants from academia, industry, and government. The organizers also made a special effort to encourage women to attend and participate in the event, recognizing the importance of diversity in driving innovation and progress in these fields.

The keynote speakers at the event shared their insights on the current state of IoT and Tiny ML, and discussed the opportunities and challenges facing the industry. There were also several technical sessions and workshops that provided hands-on experience with TensorFlow Lite and other tools for developing Tiny ML applications on embedded systems.

Stephen Ozoigbo

Embedded Systems

Embedded systems are at the heart of IoT and Tiny ML, providing the computational power and connectivity needed to enable these devices to function.

Arm Ecosystem

Arm also provides a wide range of tools and resources for developers, including software development kits (SDKs), operating systems, and other tools that make it easy to develop and deploy IoT applications on Arm-based devices.

Tiny ML & Tensorflow Lite

TensorFlow Lite is a powerful open-source framework that allows developers to run machine learning models on embedded systems, including those used in IoT devices.

Arm’s technology is well-suited for IoT applications and devices due to its low power consumption, high performance, and scalability. The company’s tools and resources also make it easy for developers to create and deploy IoT solutions on Arm-based devices, making it a popular choice among IoT developers.

Microcontroller Units at the Heart of IoT and Devices

Shadrach Munyeshyaka

FabLab Rwanda and Women in Technology

Butoyi Husna

In summary, embedded systems, Tiny ML, TensorFlow and IoT are closely related technologies that allow for the development of powerful, connected devices that can collect, process, and transmit data, and make decisions based on that data. This is driving innovation in a wide range of industries and creating new opportunities for the development of intelligent systems and devices.

Overall, the event was a great success, providing a valuable opportunity for attendees to connect with experts and learn about the latest advances in IoT, Tiny ML, embedded systems, and TensorFlow Lite. The Arm ELC program also provides an important platform for developers to showcase their work and compete for recognition and prizes. The organizers hope that this event will encourage more people, especially women to participate in such events in the future, and help drive the development of new and innovative applications that will make IoT more powerful and useful for everyone.

Embedded Learning Challenge Comes to Kampala, Uganda

The world of technology is constantly evolving and the Internet of Things (IoT) is at the forefront of this change. One of the key technologies driving IoT is Tiny Machine Learning (Tiny ML), which allows machine learning algorithms to run on small, low-power devices, such as those used in IoT networks. The combination of Tiny ML and embedded systems, which are computer systems integrated into other devices or products, is creating new opportunities for intelligent and connected devices.

On 22nd October 2022, we gathered at MoTV Uganda to discuss the latest developments and opportunities in these fields. The event was a great success as it brought together a diverse group of developers, engineers, researchers and industry professionals to share their knowledge, experience and insights.

The event began with a welcoming session led by Solomon Opio, the head of Tech Community Manager at Innovation Village.

Asa Lugda, who is a community leader, also took the stage and talked about the importance of communities in driving innovation and progress in technology. He emphasized on how the current era presents an opportunity for everyone to aim higher and strive to be the best in what they do.

Arm Engage and Embedded Learning Challenge (ELC) Kamapala

The third session was led by Ngesa Marvin. He discussed the role of IoT in Africa, introduced the (E3)NGAGE Learning Challenge and the importance of the Arm ecosystem in driving innovation in this field.

He discussed IoT use cases and how tiny ML can be used to build new and innovative devices for various industries such as Water, Energy, Gas, Manufacturing, and Asset Tracking.

He also explained that at the heart of these devices there is a company called Arm, a British semiconductor and software design company based in Cambridge, England. He discussed how almost all smartphones and IoT devices run on Arm.

The company provides a blueprint that System-on-a-Chip (SoC) companies use to make devices. Arm is not a chip manufacturer, but their partners use their framework to make chips.

In the first quarter of 2020, Arm’s silicon partners shipped a record 6.7 billion Arm-based chips, which equates to approximately 842 chips shipped per second. In Q1 2022, the partners shipped 7.4 billion Arm-based chips.

He also took the attendees through the various Arm Cortex, including Cortex M and Cortex A. He highlighted how the Cortex M microcontrollers are widely used in embedded systems and are well suited for low-power, cost-sensitive IoT applications.

The session provided valuable insights into the role of IoT in Africa and the importance of the Arm ecosystem in driving innovation in this field. It also highlighted the wide range of applications for Arm Cortex microcontrollers in the IoT industry.

Embedded Systems & Tensorflow Lite

Embedded systems and TensorFlow Lite are two important technologies that are driving the development of IoT and other connected devices. They allow developers to create intelligent and connected devices that can process and analyze data in real-time, leading to new opportunities for automation, control, and monitoring in various industries.

The ELC is a great opportunity for developers, engineers and researchers to gain experience with Arm’s technologies, showcase their skills, and connect with industry experts. It is also a chance for participants to learn about the latest developments in IoT, embedded systems, and machine learning, and to be part of a community of innovators working to shape the future of technology.

Ngesa encouraged the developers to take advantage of the resources and support provided by the Embedded Learning Challenge, such as tutorials and community forums, to learn more about IoT and Machine Learning and to develop their skills.

We then had a lunch and snacks break which allowed attendees to recharge and come back to the sessions with renewed energy and focus.

MicroPython and RP2040

The Fourth session, led by Michael, focused on Micro Python and its role in the development of Tiny ML applications.

He also shared about the current participation of Uganda in Embedded Learning Challenge in comparison to other countries on the continent. He encouraged more Ugandans to be involved in IoT and Machine Learning.

Women Tech Makers Kampala

The Fourth session was led by Janet who talked about the importance of Women in tech and how they should be involved in the development of Tiny ML and IoT. She emphasized on the importance of women in the tech industry, and how they bring diversity and different perspectives to the table, which are crucial for the development of new and innovative technologies.

The ELC event provided a valuable opportunity for attendees to learn about the latest developments and opportunities in IoT, Tiny ML, embedded systems, and TensorFlow Lite. It also highlighted the importance of diversity and inclusion in the tech industry, and how women can play a vital role in driving innovation in this field

Devices meets AI, at the Edge

Imagine smart IoT devices that can see and understand a picture or what is happening in a video at the device level (edge)!

Today we have security cameras everywhere in the streets of Nairobi, but they do not alert us when robbery is taking place. They only sometimes record activities. To take pictures therefore is not the same as to SEE or to Understand!

Imagine devices and sensors installed on strategic traffic locations analyzing traffic information in real time, at the edge and helping route traffic effectively on their own.

Well, that possibility is already a reality! Today we have smart devices at our homes like Amazon Alexa that we can speak to and they will hear, listen and understand what we are telling them, and even act.

On 10th September we gathered the AIoT community ecosystem to engage on the technologies behind these smart devices. We had a quorum at the famous Sir Thomas Moore Building, Strathmore University at around 9 am. We had a welcoming session and then I Introduced the Embedded Learning Challenge (ELC) together with the Arm Ecosystem. But first what is ELC?

Embedded Learning Challenge

Arm and ecosystem partners are supporting digitization initiatives across emerging economies. They are training new generation of African embedded systems engineers, accelerating ecosystem awareness of Arm based MCUs and encouraging the development of local solutions.

The ELC Challenge is aimed at African learners who have been historically underrepresented in STEM. It is designed to bring more learners across the continent into embedded systems and Tiny Machine Learning (TinyML) fields.

The challenge is completely FREE program for all successful applicants

How the Challenge Works?

Members of each cohort gets access to industry oriented self-paced learning curriculum from Arm Education. They then chose a learning path after which successful candidates get access to Arm Virtual Hardware (AVH) and a variety of Arm based microcontrollers. This provides them with hands on experience and practical experimentation.

Learners also gain visibility and are exposed to certification and internship opportunities within the Arm ecosystem. While on their Learning Journey, participants receive free access to Mentors as well as the best Arm ecosystem platforms for technical support.

Arm and Device Ecosystem — Shared oxygen for SoC designers

Today more than 80% of devices runs on Arm Architecture from devkits to commercial IoT devices. In fact, Arm is at the epicenter of the world’s largest compute ecosystem.

In the fourth quarter of 2020 (Arm FY Q320), Arm reported its silicon partners shipped in the prior quarter a record 6.7 billion Arm-based chips, which equates to ~842 chips shipped per second

And a white paper by The Linley Group stated that, “Having shipped in more than 130 billion chips, the Arm architecture has become as familiar as breathing to many SoC designers,”

The biggest question I have always faced in all the ecosystem and partnership engagements is “what is arm is and where do they play on the Devices value chain?

I explained to the developers that arm basically tell stories to several writers who then publish books.

Arm does not manufacture its own chips. It has no fabrication facilities of its own. Instead, it licenses these rights to other companies, which it calls “partners.” They utilize Arm’s architectural model as a kind of template, building systems that use Arm cores as their central processors.

It’s therefore up to ARM’s partners to actually build and sell the chip. These partners include Raspberry Pi, ST Microelectronics, Siemens, Toshiba, Nvidia, NXP etc.

Arduino RP 2040

Arm recently shipped the community in Nairobi a bunch of Arduino Nano RP2040 Connect. The board brings the new Raspberry Pi RP2040 microcontroller to the Nano form factor. It has a dual core 32-bit Arm Cortex®-M0+, U-blox Nina W102 Bluetooth and WiFi connectivity, onboard accelerometer, gyroscope, RGB LED and microphone.

With the sensors, Learners and Makers can now dive into real-world projects developing robust embedded AI solutions with minimal effort.

After the first session, we had Clinton Oduor take the developers through “Machine learning at the far far edge”. Clinton is an Edge Impulse Expert, and fellow Co-organizer at TinyML Kenya.

AI at the far Edge — Tiny ML

Our smartest devices including cameras are blind, and therefore collectively as a society, we are blind. How do we teach these Cameras to see and understand, just like human beings? How do we enable these devices to name objects, identify people, understand emotions and intentions?

The first step is to get the data (video feeds) in an embedded system. An embedded system is a computing device that is usually small, or tiny, and operates with extremely low power. So much so that some of these devices can run for days, weeks, months, sometimes even years on something like a coin cell battery.

Secondly, a machine learning model, is deployed to run locally on these devices. These ML models then infer interesting patterns about new data that it’s seeing.

Vision begins with eyes, but it truly takes place with the BRAIN.

Tiny machine learning, or TinyML, is therefore the intersection of embedded internet of things (IoT) sensors (devices) and machine learning.

BUILDING with TensorFlow Lite

TinyML has the potential to transform the world. One of the popular tools we can use to enable on-device machine learning (TinyML) is Tensorflow Lite.

It is a mobile library for deploying models on mobile, microcontrollers and other edge devices.

More than 4 billion edge devices have Tensorflow Lite installed on them today

Before you start with Tensorflow Lite, you will need a pre-trained Model. The model is compressed into flat buffer with the TensorFlow Lite Converter and then exported ( .tflite file ) and loaded into memory of a mobile or embedded device.

Once loaded into memory, we can now perform inference using the model on our Microcontroller.

The future of ML is Tiny

The Internet of things device ecosystem field is growing and so is the Machine Learning field. We have more than 250 Billion microcontroller units today, and this number is only going to rise in the future.

While conventional machine learning continues to evolve towards more sophisticated and resource-intensive data centers, tinyML fills a growing need at the other — smaller — end of the spectrum, Edge.

When you take a picture or speak to your device, you want the ML magic to happen there and then. You don’t want to wait for the image or audio to be sent to a data center where it is being processed and sent back again. This takes time and cripples the user experience. You want the ML model to run locally.

Enabling Machine Learning in microcontrollers will open up new opportunities and will become deeply ingrained into our everyday life in the years ahead. There is therefore a real opportunity for anyone who wants to learn about it and get involved.

Diversity and Inclusion in Embedded systems

STEM fields have long struggled with a lack of diversity. After, Clinton we had Esther Mueni dive into a “Diversity and Inclusion in Embedded systems” Session. Esther is the Co-founder of Neverest IoT and Lead for She Code Africa Nairobi.

To have better smarter IoT devices we need to encourage equal participation of women and underrepresented individuals in STEM. We all see devices and coding differently and therefore when need to bring together people from different backgrounds, genders and age groups in developing hardware.

Ecosystem Lab

To support these learners, Arm and Ecosystem partners are setting up Labs focused on engaging, educating, and cultivating local technology ecosystems.

They will incubate startups and prepare training platforms to familiarize developers with Arm tools, resources, and libraries. The lab will also spotlight Arm led sustainability programs and partnerships, while building local stakeholder networks that are developing needed solutions that address regional problems.

The pilot lab has been set up in partnership with the Cortex Hub, a regional technology incubator located in East London, Eastern Cape, South Africa, Africa’s automotive capital.

Lessons learned at Cortex Hub will serve as a blueprint for opening centers elsewhere in Latin America and Africa.

Afterwards, developers separated into groups for a 30 minute ideation session to come up with TinyML idea projects. From the discussions, the teams had a deeper understanding of the event’s theme. The groups engaged their critical thinking in figuring out how to come up with prototypes and later did presentations for the day. The level of creativity in the teams were amazing!

Conclusion

The first ELC event was successful. More than 50 developers attended and the content was great. With a vibrant ecosystem, the number of IoT devices can only rise and therefore more AIoT solutions. The price of these devices will as a result decrease (because, economics! Duh!).

ELC 1.0 is only the first of many AIoT Chapter events. Let’s do this again in the coming weeks. In the meantime, lets keep BUILDING! What will you BUILD?

…

Still curious? Watch Wei and Pete talk about how TinyML solves challenges with computer vision, audio recognition, speech recognition and natural language processing.

If you found this helpful, click the 👏. Follow this publication for more tutorials on The Internet of Things.

NB-If you have any IoT concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can.

Precision farming in Kenya

Exciting innovations and use cases are coming out of Kenyan farms. This is after Liquid Intelligent Technologies built and deployed a nationwide IoT network infrastructure together with Sigfox, the world’s leading Internet of Things (IoT) services provider.

The infrastructure provides dedicated connectivity for the Internet of Things and enables a two-way communication service through a global radio network.

It is the first global 0G network — a low-power wide area network (LPWAN), which interconnects low-bandwidth, battery powered devices with low bit rates over long ranges.

Challenges

For a long time, Agronomists and farm managers have depended on weather data from meteorological reports and weather apps. The data however usually turn out to be inaccurate and imprecise to their farms.

The mean annual rainfall in some of these areas are also low. This makes Irrigation to be very vital.

During irrigation, farmers have always had a challenge knowing when & how much to irrigate. It has been very difficult to determine if they are over irrigating or under irrigating. Much of the important decisions have been made out of gauze work, but not anymore. This has dramatically changed since launch of the IoT infrastructure.

Enabling IoT in agriculture with Connectivity & Cloud

Weather station

Farmers are leveraging on the Liquid Telecom’s network and cloud infrastructure to deploy smart sensors and devices on their farms. This includes smart weather stations that relay information about Farm Weather conditions.

The digital weather stations enable farm managers to remotely and accurately access real-time information about rainfall, temperature changes, wind conditions, air pressure, and humidity from their farms.

It’s hard to believe that some of the base stations listening to the messages are over 100 Km away from the Farms. These base stations get the messages wirelessly in star topology, and then securely forward them to cloud in a point to point link

To accurately read water consumption data and quickly detect bursts and leaks, farmers are installing smart water meters on their farms. They are also adding retrofits in cases where they don't want to replace entire meter. The
consumption data and alarms are then securely transmitted to the Cloud.

With the data, Farm Managers and agronomists can therefore make accurate decisions about their farms and crops. They are also able to optimize labor, water usage and crop health.

Crop management — soil probes

Capacitance soil probes are also installed to measure soil volumetric water content and temperature. These sensors come in different shape and size. The sensor below for example can read both moisture and temperature simultaneously, with 6 sensors, each distributed every 10cm down to 60cm.

The battery powered devices deliver real-time reading, at customizable intervals, directly from the Feld to the Internet, eliminating the need for manual readings.

The probes give real picture of moisture on all the effective root profile and reflect the exact consumption of plants and climate in the specific context of the plot.

The data enables them to decide the right time to irrigate and what amounts of water to use, thus reducing water wastage and costs linked to pumping.

Effective irrigation helps in improving root development leading to improved crop quality (bud size, stem shelf life) They are also reducing leaching loss thus savings on fertilizer inputs.

Azure

The data from the network can be automatically pushed to Azure. The cloud handles computation and data storage, real-time and batch analytics, machine learning and even visualization.

Custom Platforms & Partners

The data can also be integrated to custom platforms and other cloud.

Impact — The future of farming is connected

• Precise weather data

The platform is able to give the actual weather for the farm, and alerts when the weather data such as temperature or wind goes above a certain threshold.

This helps in knowing the ideal spraying conditions and when a crop will be susceptible to sunburn or blight.

• Soil moisture and effective irrigation

The graphs provide interesting key information such as drainage in the farm and the daily soil water usage etc.

Variable management allowable deficit (MAD) lines are set to create target “Green Zone” typically between 85% and 15% of readily available water (RAW). These lines leaves room for any rain so that any free rain water is not wasted as drainage.

It also gives an indication when soil moisture is approaching stress point.

The charts show the water use of crops planted and the water available in the root zone thus an indication of the liter per acre requirement for the crops. The irrigation schedule can therefore be adjusted according to the recommendations.

The historical patterns are also being analyzed to make better long- term crop management decisions.

• Better water conservation

Knowing the exact rainfall for each crop can help optimize watering, thus preventing overwatering, which can impact not only crop health, but the environment.

Overwatering crops can affect how much oxygen gets to the roots, which prevents them from growing normally. In some plants, overwatering can also cause root rot, which may cause the crop to eventually die.

Underwatering crops usually has the same outcome as overwatering: without proper irrigation, crops may not grow properly or could end up withering up and dying.

Beating the competition to market

• Save costs: Smart farming leads to lower costs on labor, water, and nutrients for crops.
• Save time and be more organized: being able to view water levels and weather conditions remotely saves the time it takes to physically go out to the fields. In addition, by knowing weather patterns, you are able to better plan out what needs done during the day while avoiding rain or other weather
• Easier to make decisions: Everything from pesticides, seeding, irrigation, and labor can be done more accurately with precise data. You can better predict spraying times by tracking historical weather patterns for the exact area and better anticipate disease risks through weather patterns and conditions.
• More efficient crop monitoring with less human error: automatic monitoring will lead to more accurate data in less time and with less labor.

If you found this helpful, click the 👏. Follow this publication for more tutorials on The Internet of Things.

NB-If you have any IoT concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can.

Supply Chain transformation with Liquid Telecom’s 0G Network, Azure & Maps

Smart Logistics solution results

• Enabled development of Smart Logistics Solutions in just months.
• Improved security for warehouses and fleet vehicles
• Helped to monitor transport condition throughout the entire supply chain
• Democratized data access and use across the organization.

Challenges

Solution — Connected Logistics

Organizations can now capture and track relevant events that occur in the logistics space and time thanks to Liquid connectivity and cloud solutions. Vehicles are being equipped with IoT devices that regularly send telemetry data such as location information, distance traveled, and even engine running state to Azure Cloud via Sigfox.

With the help of Azure and maps, Geofencing can be done on the stored information and notifications sent whenever a car leaves the correct authorized geographic location.

Examples of solutions in this space includes fleet management, asset tracking, mobility, and smart city applications.

IoT starts with Liquid Telecom 0G Connectivity

SIGFOX is a leading provider of dedicated connectivity for the Internet of Things. They provide a two-way communication service through a global radio network.

It is the first global 0G network — a low-power wide area network (LPWAN), which interconnects low-bandwidth, battery powered devices with low bit rates over long ranges.

Liquid Telecom, a leading communications and solutions provider in 13 African Countries, operates the network in Kenya

Azure IoT Hub

Azure IoT Hub enables secure bi-directional communication between devices and Azure (via Sigfox network). It authenticates the connected devices and help set up individual identities and credentials for each. It also help retain the confidentiality of both cloud-to-device and device-to-cloud messages.

Azure IoT Hub can also be integrated with other Azure services such as Event Grid to build complete end-to-end IoT solutions.

Event Grid

The Event Grid will listen to data sources such as Azure IoT Hub and Trigger serverless function within Azure.

Azure functions

SIGFOX does not understand or transform the payload being transmitted through the network. Manufacturers pack messages, payload in different ways/data formats and it is only them who knows how it is structured. Here is guide on the warfab v1 devkit payload structure.

To manage messages that will be coming from Sigfox webhook integration, go toIoTCIntegration/index.js file of the Function App and replace code in the trytry block with the following

Configure sigfox backend to trigger HTTP function URL

The configuration of callbacks is done in the Sigfox backend device type page.
The callbacks are triggered when a new device message is received or when a device communication loss has been detected.

Ensure that you can see messages sent by your device in your Sigfox backend, in Device > Messages.

Azure Maps

Location is a powerful component of providing insights. IoT data can then be taken to Azure Maps to provide geographic context to the telematics.

Azure Maps also provide Geofence API that can be used to in Azure function to determine whether a vehicle has moved outside the geofence area.

Blob storage

The data can also be stored in Azure Blob storage.

Custom Platforms

Liquid connectivity and cloud provides a simple and secure way to integrate IoT data to other custom platforms. Developers can leverage on Callbacks and Rest APIs to send IoT data into their IT infrastructure.

Custom Platforms

Returnable Asset Tracking

Warehouse

Until recently,

If you found this helpful, click the 👏. Follow this publication for more tutorials on The Internet of Things.

NB-If you have any IoT concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can.

Industry 4.0 — Your Future’s Factory is Smarter than Today’s

Are you ready to embrace Liquid Telecom’s 0G ( Sigfox ) and Artificial Intelligence?

Smart factory solution results

• Helped to cut down on unnecessary maintenance trips.
• Helped to collect consumption data effortlessly
• Helped to monitor vital parameter, leaks and rapture in pipes
• Helped to streamline operations, cut costs, and gain new insights.
• Enabled development of Smart Factory Solutions in just months.
• Democratized data access and use across the organization.

Challenges

Solution — Industry 4.0

Evolution of Industry with IoT beyond Connectivity

Sigfox

SIGFOX is a leading provider of dedicated connectivity for the Internet of Things. They provide a two-way communication service through a global radio network.

It is the first global 0G network — a low-power wide area network (LPWAN), which interconnects low-bandwidth, battery powered devices with low bit rates over long ranges.

Liquid Telecom, a leading communications and solutions provider in 13 African Countries, operates the network in Kenya

Azure functions

SIGFOX does not understand or transform the payload being transmitted through the network. Manufacturers pack messages, payload in different ways/data formats and it is only them who knows how it is structured. Here is guide on the warfab v1 devkit payload structure.

To manage messages that will be coming from Sigfox webhook integration, go toIoTCIntegration/index.js file of the Function App and replace code in the trytry block with the following

Cosmos DB

The outcome of Azure functions is then binded into Cosmos DB. Using Azure Functions and Azure Cosmos DB, you can create and deploy event-driven serverless apps with low-latency access to rich data for a global user base.

Jupiter Notebook

Once the data is in Cosmos DB, built-in Jupyter Notebooks can be used to explore it. In fact you can do data exploration, data cleaning, data transformations, numerical simulations, statistical modeling, data visualization, and machine learning within the build Notebooks.

Power BI

The data is then taken to Power BI for transformation and rich visuals. It enables you to retrieve data from various data sources including Cosmos DB and create dashboards and reports.

Power BI Desktop is connected to Azure Cosmos DB account with the Azure Cosmos DB connector for Power BI.

If you found this helpful, click the 👏. Follow this publication for more tutorials on The Internet of Things.

NB-If you have any IoT concepts you’d like explained in a post, any question, comment, or suggestions, please let me know. I’ll get back to as many as I can.