Reborn in 2010, the AI ​​in my mind is too powerful.

On Saturday afternoon, Chen Yiran casually grabbed a gun, stuffed it into the holster on his lower back, and drove to Stanford.

Coupa Cafe is located on the east side of the Stanford Computer Science Building. It has a small storefront but is packed with students with MacBooks, and the walls are covered with various academic posters and lecture notices.

The sound of steam from the coffee machine mingled with the clatter of keyboards, creating a romantic atmosphere.

There's no such thing as a "Instagrammable" spot yet, but there's no doubt that Coupa Cafe has the potential to be one.

Chen Yiran pushed open the door and glanced around. The first person she saw was Adam Coates, just like in the photos online. He had dark brown curly hair, and his overall attire was very stereotypical; he was only missing a plaid shirt to complete the default Silicon Valley image.

Surprisingly, sitting next to him was a man in his forties, a Chinese-American with frameless glasses—none other than the academic giant Andrew Ng. Chen Yiran had thought he would only meet Courtesy this time, and that he would meet Andrew Ng after several exchanges.

After all, it's standard practice in academia to first have junior researchers and PhDs explore the field and see if communication is possible.

As Chen Yiran went over to greet them, he jokingly said to Jasmine in his mind: "Jasmine, your teacher has arrived, or rather, your mentor."

Strictly speaking, I operate based on your brain model, and various memories are stored in your brain. He is not my teacher; he is yours.

Andrew Ng, on the other hand, was very enthusiastic. After confirming Chen Yiran's identity, he took the initiative to introduce himself.

"Giggle giggle giggle, gurgle gurgle, gurgle gurgle, gurgle gurgle, giggle gurgle, giggle gurgle, giggle gurgle."

Chen Yiran felt his head was spinning. He could vaguely sense that Andrew Ng was speaking Chinese, but the strange tone, the lewd manner, and the inexplicable pronunciation were simply a disaster.

Chen Yiran was stunned for about ten seconds, not knowing how to respond. Would it be inappropriate to directly accuse the professor of smearing Chinese culture?

Andrew Ng, seeing his helplessness, reluctantly switched back to English and said, "Sorry, sorry, I've been practicing my Chinese a bit lately. I was thinking of going to China to develop my career if I had the chance, but it seems it hasn't been very successful. Do you have any suggestions? What needs to be changed?"

Chen Yiran was completely baffled; how could she fix this? The whole paragraph was gibberish. However, she couldn't refuse the professor's request, so she racked her brains and forced a few words of praise.

"Not at all, it's not bad. At least I could tell it was Chinese, and the tone didn't crack. With more practice, it'll definitely be fine."

Coates stood to the side, feeling somewhat helpless. He desperately wanted to pretend he didn't know his advisor, but he was also afraid that he wouldn't be able to graduate even in another ten years, so he had to desperately suppress his laughter.

Andrew Ng then steered the conversation towards business: "Thank you for making the trip over the weekend. Coates showed me your DeepFlow, and our group ran three experiments, and the bottlenecks on GPUmat disappeared completely. So I came directly here because there are a few things I'd like to discuss with you in person."

"It's my honor to be able to help. I've watched your classes online before, so strictly speaking, you're also my teacher."

Without a word, Andrew Ng pulled a ThinkPad out of his bag, opened a code file on the screen—to be precise, several thousand lines of MATLAB code with scattered CUDA kernels embedded in the middle. Chen Yiran glanced at it, but could barely understand it; it was all English comments on various matrix operations.

"This is the unsupervised feature learning code we've written over the past two years, a framework that Quoc Le and Adam started building in 2008. Simply put, it's a multi-layer sparse autoencoder, a local receptive field, and a custom normalization layer. It runs on GPUmat."

He changed the diagram to a model structure diagram. The input layer reads image patches from STL-10, and four autoencoder layers are stacked on top of it. Each layer has its own sparse constraints and is not fully connected.

"I want to port all of these to the DeepFlow framework. GPUmat can't keep up anymore. We want to meet the NeurIPS deadline, which is around early September. It's mid-July now, and at GPUmat's speed, the paper definitely won't be ready."

"So what do you need me to change?"

"It's not about changing anything; it's that your framework is missing three things. Without these three things, our model can't be ported."

Andrew Ng took a piece of paper and listed out the missing items one by one.

"First, sparse constraint operators. Our autoencoder has sparse activation constraints in its intermediate layers, and the gradient calculation during backpropagation is different from that of ordinary fully connected layers. DeepFlow currently only has the basic dense, convolution, and pooling operators, and does not have this."

Secondly, we use a custom receptive field. Our model doesn't use standard convolutions. Instead of a 3×3 kernel sliding across the graph, each output neuron is fixedly connected to a local region of the input layer, and they don't share weights.

Thirdly, whitening normalization. We perform zero-phase component analysis whitening on the activation values ​​of each layer, and the gradient is backpropagated through singular value decomposition. This is different from any current framework.

Andrew Ng closed the ThinkPad and looked at Chen Yiran with great anticipation.

"If you can add these three operators to DeepFlow, our entire team can switch over the entire code stack within two weeks. If you can't, I won't blame you; after all, it's an open-source framework, and it's not obligated to customize it for us."

But we might have to rely on GPUmat to get through September. It's such a waste of all our hard work, and no one is helping us. People who don't help are so heartless...

…………

Chen Yiran was a little speechless, but still asked in his heart: Jasmine, can you handle these three things?

Jasmine's reply was quick: "Backpropagation with sparse constraints is a chained derivative of a custom gradient layer, requiring a dozen lines of CUDA code plus several dozen lines of Python wrapper. The receptive field is a special type of matrix multiplication, essentially a variant of block matrix multiplication. Whitening is a bit more complex; singular value decomposition backpropagation of gradients requires numerical stability handling, but that only adds about two hundred lines of code."

Chen Yiran understood: I get it, you mean you can handle everything, right? Your tone is really Versailles-like.

[There's no Versailles, but it's definitely not difficult. Half an hour. Testing, benchmarking, and parameter tuning take half a day, but simply writing out the operators only takes half an hour.]

Andrew Ng and Courtes watched him remain silent for a few seconds, probably assuming he was assessing the workload.

As an academic giant, Andrew Ng also has a thick skin and has started to use all sorts of fancy ways to flatter and kill.

Chen Yiran quickly looked up and said, "I've thought about it briefly, and it should be fine. I can give it to you tonight."

"tonight?"

"Yes, the CUDA implementation of the three operators plus the Python interface will be sent to your email tonight."

"are you sure?"

Andrew Ng examined his eyes for a moment and then asked, "I'm not doubting your abilities. But we've assessed it ourselves, and if one person were to write these three operators, from learning the DeepFlow extended API to debugging, it would take at least two weeks. Add in the numerical stability issues of whitening, and conservatively, it would take a month."

"Haha, geniuses are always extraordinary."

After saying that, even he felt a bit arrogant. But it's not a big deal; he'll shock the world many more times in the future, and his genius persona is already a given.

Andrew Ng closed the ThinkPad's lid, thought for a moment, and made a promise.

"If you can get it done tonight, you'll give a demo at our group meeting next Wednesday. It's in the Gates building; our lab has four Tesla C2050s. You'll bring the new version of DeepFlow, port our model live, and your framework will run successfully. It will be cited in the tools section of the NeurIPS paper."

"Okay, I'll get it done as soon as I get back."

Chen Yiran was very happy on her way home. Andrew Ng was no ordinary scholar; many of his courses could be considered foundational courses in this field.

The fact that the Deepflow framework was cited in his paper almost meant that it would immediately attract the attention of all companies, and many companies would directly ask him to try it out.

In fact, this could be considered a small favor that Andrew Ng owes him; aside from anything else, it will make recruiting much easier in the future. Which Stanford student in a related field hasn't taken his classes? And even those outside Stanford, how many have actually taken his classes?

Chen Yiran had even considered getting more involved in the experiment to earn authorship. After all, according to academic convention, tools and basic research are acknowledged and credited separately. If you contribute to the tools, you can only say you've laid the foundation; you can't claim credit for the competition results.

But in the end, Chen Yiran got lazy. He was already financially independent, so where would he get the motivation to work harder? The fact that Chen Yiran is still persisting in starting a business is already top-notch perseverance. Other transmigrators might have already been lying down and enjoying life.

Upon arriving home, the usual routine unfolded: iced cola and ice cream were prepared, and Jasmine was ready to go all out.

[Our first operator: sparse constraints. Let's begin with a CUDA implementation of gradient chain rule differentiation.]

After writing about 70 or 80 lines of CUDA code, Jasmine had him switch to the Python file and write a wrapper. Then, for the second operator, matrix multiplication with local receptive fields, Jasmine explained that it was essentially a block matrix multiplication, but it had to bypass the sliding window mechanism of standard convolution.

The last four individual steps are the most time-consuming: whitening normalization singular value decomposition backpropagation. After thinking for a while, Jasmine designed a small diagonal regularization to prevent gradient explosion.

Jasmine wasn't exaggerating; these tasks were indeed not difficult for him. Furthermore, after much practice, Chen Yiran's typing speed had reached a professional level of pen-tip shorthand. In less than half an hour, he had completed all the revision requirements and sent them to the email address Andrew Ng had given him.

Chen Yiran felt that in another week or two, his FlowEDA would no longer be an unknown small company.