Search

Running machine learning algorithms for image classification locally on small, cheap, and low-power microcontroller units (MCUs) has advantages in terms of bandwidth, inference time, energy, reliability, and privacy for different applications. Therefore, TinyML focuses on deploying neural networks on MCUs with random access memory sizes between 2 KB and 512 KB and read-only memory storage capacities between 32 KB and 2 MB. Models designed for high-end devices are usually ported to MCUs using model scaling factors provided by the model architecture’s designers. However, our analysis shows that this naive approach of substantially scaling down convolutional neural networks (CNNs) for image classification using such default scaling factors results in suboptimal performance. Consequently, in this paper we present a systematic strategy for efficiently scaling down CNN model architectures to run on MCUs. Moreover, we present our CNN Analyzer, a dashboard-based tool for determining optimal CNN model architecture scaling factors for the downscaling strategy by gaining layer-wise insights into the model architecture scaling factors that drive model size, peak memory, and inference time. Using our strategy, we were able to introduce additional new model architecture scaling factors for MobileNet v1, MobileNet v2, MobileNet v3, and ShuffleNet v2 and to optimize these model architectures. Our best model variation outperforms the MobileNet v1 version provided in the MLPerf Tiny Benchmark on the Visual Wake Words image classification task, reducing the model size by 20.5% while increasing the accuracy by 4.0%.

Our paper compares the correctness, efficiency, and maintainability of human-generated and AI-generated program code. For that, we analyzed the computational resources of AI- and human-generated program code using metrics such as time and space complexity as well as runtime and memory usage. Additionally, we evaluated the maintainability using metrics such as lines of code, cyclomatic complexity, Halstead complexity and maintainability index. For our experiments, we had generative AIs produce program code in Java, Python, and C++ that solves problems defined on the competition coding website leetcode.com. We selected six LeetCode problems of varying difficulty, resulting in 18 program codes generated by each generative AI. GitHub Copilot, powered by Codex (GPT-3.0), performed best, solving 9 of the 18 problems (50.0%), whereas CodeWhisperer did not solve a single problem. BingAI Chat (GPT-4.0) generated correct program code for seven problems (38.9%), ChatGPT (GPT-3.5) and Code Llama (Llama 2) for four problems (22.2%) and StarCoder and InstructCodeT5+ for only one problem (5.6%). Surprisingly, although ChatGPT generated only four correct program codes, it was the only generative AI capable of providing a correct solution to a coding problem of difficulty level hard. In summary, 26 AI-generated codes (20.6%) solve the respective problem. For 11 AI-generated incorrect codes (8.7%), only minimal modifications to the program code are necessary to solve the problem, which results in time savings between 8.9% and even 71.3% in comparison to programming the program code from scratch.

The automated transcription of mathematical formulas represents a complex challenge that is of great importance for digital processing and comprehensibility of mathematical content. Consequently, our goal was to analyze state-of-the-art approaches for the transcription of printed mathematical formulas on images into spoken English text. We focused on two approaches: (1) The combination of mathematical expression recognition (MER) models and natural language processing (NLP) models to convert formula images first into LaTeX code and then into text, and (2) the direct conversion of formula images into text using vision-language (VL) models. Since no dataset with printed mathematical formulas and corresponding English transcriptions existed, we created a new dataset, Formula2Text, for fine-tuning and evaluating our systems. Our best system for (1) combines the MER model LaTeX-OCR and the NLP model BART-Base, achieving a translation error rate of 36.14% compared with our reference transcriptions. In the task of converting LaTeX code to text, BART-Base, T5-Base, and FLAN-T5-Base even outperformed ChatGPT, GPT-3.5 Turbo, and GPT-4. For (2), the best VL model, TrOCR, achieves a translation error rate of 42.09%. This demonstrates that VL models, predominantly employed for classical image captioning tasks, possess significant potential for the transcription of mathematical formulas in images.

French is a strategically and economically important language in the regions where the African language Twi is spoken. However, only a very small proportion of Twi speakers in Ghana speak French. The development of a Twi–French parallel corpus and corresponding machine translation applications would provide various advantages, including stimulating trade and job creation, supporting the Ghanaian diaspora in French-speaking nations, assisting French-speaking tourists and immigrants seeking medical care in Ghana, and facilitating numerous downstream natural language processing tasks. Since there are hardly any machine translation systems or parallel corpora between Twi and French that cover a modern and versatile vocabulary, our goal was to extend a modern Twi–English corpus with French and develop machine translation systems between Twi and French: Consequently, in this paper, we present our Twi–French corpus of 10,708 parallel sentences. Furthermore, we describe our machine translation experiments with this corpus. We investigated direct machine translation and cascading systems that use English as a pivot language. Our best Twi–French system is a direct state-of-the-art transformer-based machine translation system that achieves a BLEU score of 0.76. Our best French–Twi system, which is a cascading system that uses English as a pivot language, results in a BLEU score of 0.81. Both systems are fine tuned with our corpus, and our French–Twi system even slightly outperforms Google Translate on our test set by 7% relative.