Deploying and Training Deep Learning Models on iOS and Android: Core ML, NNAPI, and TensorFlow Lite

Recently an open‑source project called MNIST‑CoreML‑Training demonstrated that an iOS device can train a LeNet‑style CNN on the MNIST dataset without any external ML framework, achieving over 98% accuracy; training on an iPhone 11 took about 248 seconds compared with 158 seconds on a MacBook Pro using TensorFlow 2.0.

The article then compares iOS’s Core ML with Android’s NNAPI, describing how mobile devices typically train models on GPUs or TPUs before compressing them for deployment, and noting earlier Android projects that used OpenCV and Caffe for tasks such as license‑plate and gender recognition.

Core ML is presented as a model‑conversion tool that enables on‑device training, while NNAPI is an Android C‑API that provides low‑level support for ML frameworks like TensorFlow Lite and Caffe 2, allowing developers to run models via higher‑level libraries rather than calling NNAPI directly.

TensorFlow Lite, released by Google in 2017, is highlighted as the most widely used mobile deep‑learning framework; it can run on CPU, GPU, or directly through NNAPI. The workflow includes selecting or designing a model, converting it to the TensorFlow Lite flat‑buffer format, performing inference with the TensorFlow Lite interpreter, and optionally optimizing the model for size and speed.

Code examples are provided:

import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

Java inference example:

try (Interpreter interpreter = new Interpreter(tensorflow_lite_model_file)) {
    interpreter.run(input, output);
}

To add TensorFlow Lite to an Android project, include the AAR in build.gradle :

dependencies {
    implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly'
}

To reduce APK size, filter ABIs to only the needed ones (e.g., armeabi‑v7a and arm64‑v8a ) in the Gradle configuration:

android {
    defaultConfig {
        ndk {
            abiFilters 'armeabi-v7a', 'arm64-v8a'
        }
    }
}

Finally, the article mentions that developers can build a custom TensorFlow Lite AAR locally and integrate it into their apps, completing the end‑to‑end pipeline for mobile AI development.