Chapter 59 Choosing a Graduation Thesis Topic

On Monday, after finishing his morning classes and having lunch, Zhou Yun didn't even take a break before heading straight to Deng Yonghua's office.

There was no other reason: he wanted to graduate early!

Such superficial classes in the morning were simply a waste of time. Sleeping would at least allow him to recharge his energy, whereas attending these pointless lectures was a dual torture of both mind and body.

"Professor, I want to graduate early." Zhou Yun looked at Deng Yonghua with an unusual determination in his eyes.

The latter was not surprised by Zhou Yun's thought, after all, the idea of early graduation was something he had brought up in the first place.

"Have you decided?"

"Decided!"

"Good, I'll help you look into the credit situation. Although it doesn't strictly follow the rules, rules are dead while people are alive. I can't speak for others, but in your case, there shouldn't be any problem from the school's end.

However, since you want to graduate early, you definitely have to write a thesis. You can think about it: if you want to fast-track it, you can skip the proposal defense and graduate alongside this year's third-year master's students; if you're not in a hurry, you can do the proposal defense first and write the thesis to graduate with Qiu Yan and the others next year. The choice is yours."

After a brief moment of thought, Zhou Yun made his decision on the spot: "I want to graduate with this year's third-year senior students. I honestly don't want to attend even one more of these classes!"

Deng Yonghua nodded: "Since you've made up your mind, I will certainly support you." As he spoke, he took out a few sheets of paper: "However, the necessary procedures still must be followed. This is the proposal defense form; fill it out first so that processing the paperwork later will be smoother."

"Then do I still need to go to regular classes?"

"No need. During this period, just focus entirely on your graduation thesis. However, we are heading to Vancouver for a conference in December, so it's best to finish it within this timeframe. Once we return from the conference, it'll probably be time for the final defense. Do you think one month is too tight?"

"Not at all! I will definitely finish it before the conference."

"By the way, even for a graduation thesis, don't think about cutting corners. If you want to graduate early, the requirements will only be stricter. Moreover, a graduation thesis is something that will stay with you for the rest of your life, so you must take it seriously!"

Deng Yonghua reminded him very solemnly, fearing that Zhou Yun might slap together a mediocre thesis just to graduate quickly.

"Alright, I understand."

Holding the sheets of paper, he returned to the laboratory.

Zhou Yun sat at his desk, pondering the topic of his graduation thesis while tapping his fingers rhythmically on the tabletop.

Anyone who has ever written a thesis knows that creating a graduation thesis from scratch in a month is practically impossible.

Therefore, if he wanted to accomplish this, basing it on his existing work seemed like a viable approach.

The work he was currently doing was limited, leaving him with only two options to choose from: model compression and multimodal large language models.

A multimodal large language model was definitely out of the question; running experiments alone would consume the entire month, let alone writing a graduation thesis.

Model compression? Actually, that wasn't very easy to do either, as running experiments would also take up a considerable amount of time.

That left him with only one path—pursuing theory.

By pursuing theory, he didn't mean pure abstract mathematics, but rather research into model foundational architectures that only required small-scale validation experiments. For instance, the paper introducing the Transformer, *Attention Is All You Need*, actually didn't feature a massive amount of experimental work.

Could he directly propose an entirely new technical model architecture?

To be honest, this was a question he had pondered for a long time—so long that it could be traced back to his past life.

Although Zhou Yun wasn't this smart in his past life, the very first time he encountered the Transformer, he had wondered whether he could research an even more powerful foundational architecture.

Of course, in his past life, that was purely a fantasy.

Therefore, whenever he had free time, he would set aside moments to ponder this question, which was why Qiu Yan and the others frequently saw Zhou Yun sitting alone at his desk in a daze, sometimes for hours on end.

Yet even though he felt he was now smart enough, after nearly half a year of contemplation, he still hadn't managed to conceive any entirely new architecture, ultimately remaining bound within the framework of the Transformer.

Still, he wasn't discouraged; after all, if something like that were successfully researched, calling it a legacy that would echo through history would not be an overstatement.

Furthermore, he was still young.

Since a new architecture was out of the question, he had no choice but to start from the two fields he was already familiar with.

His fingers paused as he suddenly thought of an exceptionally good thesis topic. If he could pull it off, the contribution would undoubtedly be monumental.

That topic was—Optimal Transport Theory in Multimodal Fusion.

The core of multimodal learning lies in how to align the feature spaces of different modalities (such as vision and language). Current methods typically rely on cross-attention mechanisms, or even simpler dot-product or cosine similarities.

Optimal Transport (OT), on the other hand, is a mathematical theory dedicated to finding the most efficient way to move mass or probability from one configuration to another while minimizing a given cost.

His idea was to model the alignment problem as an OT problem. Of course, this OT problem would undoubtedly be extremely difficult, because the dimensionality of each feature is exceptionally high, and there are a vast number of multimodal features.

Thus, he naturally associated this with the human brain. When the human brain processes different sensory information (visual, auditory, tactile), there seems to be a unified "stream of consciousness" at the foundational level.

After information from various modalities pours in, it gets mapped, exchanged, fused, and understood within this stream—much like eating watermelon while smelling durian, which produces a form of synesthesia.

What he needed to do was find a mathematical way to represent this so-called "stream of consciousness," and the closest theory he could think of was OT theory.

As for why information fusion is even necessary, it is actually quite easy to understand. To use an analogy, suppose you need to explain what an "apple" is to an alien from another galaxy who completely lacks Earth culture.

You cannot simply show it a picture of an apple, or just give it an apple to eat.

Because to the alien, this information is fragmented and incomprehensible. Therefore, we need to transform each piece of information into a sensation—such as the apple being sweet, sour, green, or red...

When humans learn an unfamiliar thing, the process is quite similar. Take Shennong tasting the hundred herbs as an example: he first understood the various properties of each medicinal herb before finally defining it.

Similarly, to make AI understand the world, this step is also necessary. You must first let the AI understand the basic properties of an apple, which is a step many large language models are already performing.

The next step is information fusion, which means telling the AI about a more advanced, unified semantic space, conveying to it that although the modalities of the information differ, they are all describing the exact same thing. This is the fusion step.

This is also the principle by which modern multimodal large language models are able to process different modalities of information simultaneously.

However, when incorporating different modal information into this unified semantic space, current methods still suffer from numerous shortcomings.

This problem is precisely a stumbling block on the road to truly achieving AGI.