Please check out my Google Scholar page for a complete list of these works, full author lists for all works, and links to PDFs.
Here, I’ll provide a short overview of each work and what I learned while working on it.
Phi-2 and Phi-3 were an absolute blast to work on. I’m ecstatic about both what these models represent for the future of AI and about the reception of these models by the community.
To me, these models represent a shift for LLMs by putting a larger emphasis on high-quality data, especially augmented with targeted synthetic data.
In my opinion, using the term “synthetic generation” doesn’t quite do the data justice. I believe the future of synthetic generation for models entails thinking about generation in a new way - sculpting text samples over many rounds of iteration. This is beyond simply getting the response from a model to a fixed, single-round prompt. Instead, we can push a model to generate high-quality data from scratch about nearly anything.
I enjoyed working on various pieces of these models, including post-training, data generation, and even some parts of pre-training.
Besides the technical details, it was invaluable for me to learn how to be a part of a large team training one model. This kind of project is very different than a group of 2-5 people writing a paper. In many ways, I’ve gone full circle - my engineering background often helped me significantly in these projects.
To build models, we must blend engineering and research. Organizationally, this work is weird. Software engineering is radically different to more traditional engineering, AI research is radically different to more traditional research. Here, we need to successfully do both.
After the “internal rotation”, I transitioned to an internal applied research team known as AI Platform. On this team, I worked on many projects, ranging from direct product work to research.
I was fortunate to be co-author on a couple other works during this time, including Direct Nash Optimization and Adapting LLM Agents Through Communication. These projects were incredibly fun to help with. I would highly recommend reading those works for yourself!
PPO and more broadly RLHF revolutionized LLMs, and after its emergence for LLM post-training, many folks (myself included) were trying to run experiments centered around it. In this work, we attempted to develop a method to make PPO significantly more memory-efficient without hurting final performance. We accomplished this by using multiple sets of LoRA weights to serve as the various mdodels within PPO.
While I’m proud of the idea, I must admit the execution in this work could be significantly better and the idea has much more room for expansion. Ultimately, this idea is not incredibly novel, and this paper leaves some unanswer questions. “Hydra-RLHF” was not a technique that caught on.
This was a valuable project for me to have worked on. I learned lessons from it that still affect my work today and I am glad to have worked on this project.
I’ve always loved network compression as it requires asking interesting questions (and finding interesting answers). In this paper, we developed a method for pruning models based on “neuron uniqueness”.
At the time, there were many methods for pruning models, but they mostly centered around considering all network neurons independently of each other, for example, how much does neuron A affect the final output. We wanted to instead ask the question: how unique is neuron A compared to all the other neurons, and if it is highly unique, is it more important to keep it when pruning?
Beyond this question, we had an additionally interesting finding - simply pruning weights randomly performed almost as well as some other state-of-the-art pruning methods! This was further evidence that pruning methods were looking at the wrong thing.
Overall, I love this work, but in retrospect it could use a lot of love. Better experiments, more solid evidence, and following up on some obvious questions. Still, I’m proud of it overall, and it may be worth re-visiting this concept in the future.
After graduating, I started at Microsoft as a SWE. Quickly, I realized I wanted to get back into research. However, this is much easier said than done without a PhD and I spent a long time trying to break into research without success.
Greg Yang gave me a great opportunity: he allowed me to work with him on this incredible project which was well beyond my depth. In a 6-month “internal internship”, I rotated to MSR.
This project was incredibly challenging and rewarding. I spent a long time understanding the Tensor Programs papers and finally contributed to this paper by running experiments and implementing the Pi-Projected architecture for a few models.
In this paper, we develop a method for exactly computing trainable, infinitely-wide neural networks. This is done by projecting the weights into a lower-rank (and finite) space. Interestingly, in this paradigm, training a network corresponds to appending more rows to these low-rank matrices, meaning the network gets larger over time.
We had even greater plans after this paper, from the pilim repo to a few side projects (can we directly do autograd on a pi-projected network instead of growing the rows of the matrices? Can we prune redundant, linearly dependent rows in the pi-projected network?).
However, in retrospect, I had bitten off more than I could chew at that time and did not accomplish many of these efforts. Overall, it was an incredible learning experience for me, and I’m forever grateful to Greg for providing me the room to try and grow.
At UC, I was able to do a Bachelor’s and a Master’s of Science in Computer Engineering. For the thesis, did research on using neural networks for malware detection. I found that I absolutely love research - it is some of the most rewarding work I have ever done.
This, by far, is the paper I’m most excited about from my time at UC.It’s available here.
This journal article summarizes the main highlights of my M.Sc. research. The concept of Malware as Video, the best networks that we were able to build, comparisons to seminal works, and Node-Distance Pruning. Our results out-preform all works that we were able to find while staying lean and extraordinarily practical.
In addition, out of my work, I believe Node-Distance Pruning may be useful more generally. I found it to be a highly effective algorithm and may shed new light on where to go with pruning neural networks.
Overall, the paper is incredibly gratifying. It feels beyond exciting that the culmination of my work out is there, in a format for others to read and hopefully build on. So many iterations and so much work went into this final paper and I am glad it was able to be published.
If you’re interested in my work I highly recommend downloading my M.Sc. thesis, available for free here. It’s the most complete and organized work I’ve written both in terms of my research and in terms of, well, everything I’ve ever done. The other papers I’ve written were cut out of the context of the broader picture and therefore, in my opinion, provide value but leave a lot of unanswered questions. Even if there is only one section that interests you, I’d recommend going straight to it and treating the other sections as possible answers to other questions that come up.
I thoroughly enjoyed the process of assembling the work for, and writing, a thesis!
Chronologically, the work that this paper came from happened last in my time at UC, though it ended up being published sooner due to long review times on the other papers. It is available here.
I love this paper. It’s highly interesting and offers a ton of future work possibilities. Essentially, with saliency, we can see what the networks found to be unique to malware or benign code. This is a fruitful result - what were the commands that were unique to malware or benign code? Can we use these code clusters for a different, faster classification method and cut out the networks altogether? How would saliency differ from Malware as Image methods as compared to Malware as Video?
This short paper, in my mind, raises many interesting questions. It was a great, short project to wrap up my research with at UC as it leaves the door open for future work. There is still so much left to uncover with neural networks and I’m excited to see what the future brings.
ASTESJ invited us to write a paper extending our work done in the NAECON 2017 paper, which was our first neural network attempt.
I’d like to discuss the journal itself. If one searches ASTESJ they will find some accusations of misconduct/unethical publishing, which was worrisome for me especially as someone young in their career. After consideration, we decided to write a paper on the recurrent neural networks we attempted to use for malware detection, which is available here for free.
There are not many papers using recurrent neural networks in this way - our intention was not to get an easy publish, but to contribute some information that would not go into other papers we submit and be otherwise assumed knowledge without justification. I’ve been apprehensive about the journal, however, all my interactions with the journal were professional, the other papers in the journal seem credible, and others vouch for the integrity of the journal. Therefore, my conclusion at the moment is that the journal is simply young and still growing.
In my eyes, the work should speak for itself, regardless of the journal it appears in. This paper was important for establishing that while recurrent neural networks can work for this solution, they are not at all necessary or optimal. I’m happy with how the paper came out!
Work began for my research by replicating the results of others with a simple approach, and the first paper we wrote based on that work was presented at NAECON in 2017. The paper is available here.
There are honestly a multitude of things I would change about that first paper. If I could rewrite it, I would tie in more sources, better present the results of the paper as the figures are lacking, more thoroughly test the networks in various ways, and re-word large portions of it for clarity. At the same time, everyone starts somewhere and I’m glad to have been able to write that conference paper at all. I learned much from this paper and believe that the results of that learning can be seen in the works above.