1. First goal is mentioned as, "prove that you can learn japanese without using your mother tongue". But when I click on start you show a Monkey with some Japanese text. In my mind I see the picture of a monkey and think of the word in my mother tongue, Telugu(Kothi), then I check the Japanese script. So I am not strictly learning Japanese without using my mother tongue.
2. Since Japanese script is also new to me it becomes harder. As I have to learn the alphabet first and then words and then sentences.
3. To really try this experiment out instead of the Japanese text the picture should be associated with the Japanese word and Japanese sentences. If you realize, thats how we learn language as a kid. We are shown images and told to repeat words and then sentences.
4. Deep learning tries to do something similar. Here what you are trying to do is encode the word in an image as a translation layer and then we jump across languages. Deep learning also learns embeddings and then uses those embeddings to translate. In this case, the embedding is an image, but computers can use lesser info than all the pixels to embed.
5. This embedding approach has limitations in deep learning as we can see from the current translation engines. But by using sound and image maybe our brain embeds information better and hence we learn languages better.