While the COVID-19 pandemic shows signs of ending, the number of infected people still suddenly increases at times, and the situation continues to require vigilance.

The world situation is also changing dramatically. I hope I can write travel articles again like before, but I don’t know when that will be.

Deep Learning

The honorific translation app was originally created as a project to study deep learning, but for a long time I couldn’t find a use case for it.

In last year's article, I wrote that I would use Seq2seq for speed improvement, but the processing of Mecab and the Dart compiler was faster than expected, and even with over 300,000 words in the honorific conversion dictionary, it finishes instantly, so it wasn’t necessary. Moreover, since it’s already fast enough, I haven’t modified it yet, but there’s still room for further speed improvements.

Also, I believe that using deep learning to convert written text into honorific language is not practical with current technology. While it’s possible to create honorific sentences, people have their own preferences for honorific language.

For example, “as the teacher said” can be replaced with “as the teacher graciously said,” but depending on the distance between the user and the teacher, “said” might be just right in some cases. To determine this, we would need to input the distance with the teacher in advance, but I don’t think anyone would use such a troublesome app.

After researching various uses, I found that using BERT for conjunctive particles (particles that come after nouns, such as “no,” “ga,” “wa”) and correcting kanji conversion errors in nouns was practical, so I implemented these. However, at present, this is only for iOS, and from an accuracy standpoint, it doesn’t make automatic corrections but only suggests candidates. I studied BERT by reading this book.

With BERT, it’s common to use a model you’ve pre-trained yourself or apply fine-tuning to pre-trained models published by others. There’s also another approach where you use models that have been pre-trained on a scale nearly impossible for individuals, accessing them via API, such as GPT-3 and ChatGPT published by OpenAI. This is a mechanism that “generates text in a predetermined format based on pre-written text,” but the text is incredibly natural and often indistinguishable from human-written content.

I plan to try this approach next year. However, since deep learning is a rapidly evolving field, other methods might emerge before I get to try it.

App Research

The 2021 portion ended at the end of June, same as last year, and I installed about 13,000 apps, also similar to last year. Originally, I planned to complete it by the end of March. The delay was caused by iOS bugs that occur when installing apps in large quantities. For some reason, when the number of apps exceeded about 2,000, I couldn’t restore from backup no matter how many times I tried.

Also, when it reached about 3,000 apps, the device would suddenly shut down and never boot up again, which happened frequently. Last year, I managed to fit just over 8,000 apps, so the situation has worsened.

Apparently, iOS 15 is the cause. Due to iOS specifications, you can’t downgrade the version, so I need to buy a new device with version 14 or earlier, but there are already apps that require a minimum version of 15. Until I find an appropriate solution, I can’t proceed, resulting in the installation work stagnating for about three months.

In the end, I bought a used iPhone with iOS 13, which seemed to have the most stable operation, and installed apps that couldn’t be installed due to minimum version requirements on the iOS 15 device. Although it took unexpected effort, if I want to launch multiple businesses, I believe it’s more efficient in terms of both time and cost to investigate the common elements of successful apps on the App Store through this work and understand the trends of what’s popular and what’s not.

Orchestra

• Mother’s day Wind Orchestra

A game music orchestra after a long time.

They apparently practiced for two years for a one-time performance. The content showed that they had been gradually developing ideas over two years, and it was simply wonderful.

Movies

• Suzume no Tojimari (Suzume)

It was as good as “Your Name,” hard to say which is better.

Many popular movies have innovations for spreading through word of mouth, which is educational.

Ring Fit Adventure

I started at the end of 2019, and finished my third playthrough at the end of August. It took 2 years and 8 months.

I somehow adjusted it so that I could get all the titles at the same time as completing all worlds.

It was tough until I finished the first playthrough. I could barely manage 15 minutes twice a week, but from the second playthrough, I didn’t find it burdensome anymore. That said, there were months when I was too busy to do it at all, so it took time.

As of the end of December, I’ve started over from the beginning while keeping my exercise records, and have essentially finished my fourth playthrough.

In the past, I contracted a fitness gym for about a year, but in the end, I only properly attended for about the first three months, which was a huge waste, so I think Ring Fit Adventure is great.

I feel there’s potential in gamification.