Dozens of deserted concepts, code unfold throughout numerous repos and lots of of analysis papers you learn over 4 years in the past. The thesis write-up is an try at making sense of this mess — a frightening process in case your PhD journey was something like mine. Fortunately, we now have the best instrument of knowledge synthesis to assist.
I wish to share the ways in which I used AI instruments. These embrace: a analysis assistant for locating citations, combining code from a number of experiments to enhance reproducibility, fact-checking in opposition to cited papers. and producing defence questions based mostly on my work. I will give particular prompts and particulars of the instruments I used. You will notice that we will implement these with a primary subscription from one of many AI corporations.
On the similar time, I hope this information will assist us keep away from works of educational slop. These functions on no account offload psychological bandwidth to an AI system, and in some circumstances they even elevated the period of time I put into the thesis. It is because they don’t seem to be geared toward “writing” the thesis however at serving to me produce a greater piece of labor than I’d have been in a position to alone. On this means, I hope the functions can profit anybody attempting to supply a fancy piece of writing.
1) Quotation finder (analysis assistant)
The subject of my PhD was deep studying for coastal erosion monitoring. It concerned growing fashions and frameworks for robotically detecting shoreline proxies just like the vegetation line (VL) (see Determine 1). I learn many papers, from particular deep studying architectures to these making use of them to common distant sensing issues and even geological research in my nation of curiosity, Eire. By the top, I may write a sentence, be fairly certain it was right and know roughly which assortment of papers might be used to quote it.
Vegetation strains on Bull Island, Eire, from Geological Survey Eire annotation in contrast with a deep studying mannequin on Sentinel-2 imagery. Imply absolute error of 1.9m.
The issue is I learn many of those papers 3-4 years in the past and, with out the hindsight of an exhausted final-year pupil, made no try at summarising them. Making an attempt to dig out one of the best quotation would require rereading them till I discovered the purpose I used to be in search of. As a blessing to my sanity, this use case is a RAG system in disguise — one thing LLMs are good at. Nevertheless, we must be cautious to not incorrectly cite anybody, as it is a lesser-known type of plagiarism.
To assist, I created a Claude Code venture with the directions beneath. On this venture, I included the BibTeX file with all my citations. I additionally included a few of the most necessary papers from the background part, which I believed would come up rather a lot as I wrote the thesis. Then, once I wished to quote one thing, I’d paste the sentence I wrote together with any further analysis papers I believed could be useful.
For instance, as a immediate, I’d have pasted one thing like:
I then, and I can not stress this sufficient, didn’t blindly belief Claude’s output. The final a part of the directions supplied helpful data that allowed me to double-check the instructed citations. This concerned going again to the analysis paper to grasp the related textual content within the context of the remainder of the paper.
That is removed from the method taken by some current papers, the place it’s painfully apparent they’ve simply handed a set of papers and requested AI to supply a abstract. Having marked many pupil papers throughout my PhD, I can assure you that this method will result in errors. The wrestle of manufacturing this abstract can also be necessary to the science itself.
It’s by means of studying, summarising, and discovering the hyperlinks between a number of disjointed fields that you simply start to type your individual concepts. I believe this software solely labored so effectively as a result of I had a great understanding of the analysis space, wrote my very own textual content and requested the AI instrument to quote it utilizing a curated record of papers. In different phrases, utilizing it as a search instrument versus a abstract instrument. We method the subsequent software with an analogous mindset.
2) Code repository consolidation
We transfer on to an apparent software: writing code. Primarily based on my thesis matter, you might count on a number of Python code for constructing datasets, coaching and evaluating fashions, reasonably than working issues in manufacturing. Because it was all written in a tutorial setting, you possibly can additionally think about it was not the cleanest code. To make issues worse, it was unfold throughout many, many repositories. Fortunately, that is one thing coding brokers are perfect for dealing with.
Really, a lot of the code was not written utilizing agentic coding instruments like Claude Code. If I am trustworthy, this was as a result of they weren’t as succesful once I began my PhD. I did, nevertheless, use Copilot in VS Code extensively. That is extra of a coding autocomplete the place it guesses the person line of code you wish to write. Wanting again, I am grateful I wrote most of my code this fashion.
Going line by line gave me a great understanding of the code. You may clearly nonetheless perceive the code you write with brokers, however paper deadlines and inherent laziness would probably have allowed me to easily belief the output. Just lately, I’ve tried to do a couple of initiatives this fashion, and I in a short time find yourself with an awesome mass of advanced code. Now, I did discover agentic coding helpful in direction of the top of my PhD.
One chapter of my thesis is geared toward producing picture segmentation fashions that classify each pixel in a picture as both land or ocean. As this was the exploratory section of the thesis, I experimented with many approaches. You may see a few of these in Determine 2. I skilled fashions throughout 4 completely different datasets and utilized completely different picture segmentation architectures, information augmentation, and fine-tuning approaches. The outcomes that made it previous the untitled.ipynb section had been unfold throughout 4 tutorial papers, all with an accompanying code repo.
Coastal picture segmentation on Landsat and Sentinel-2 imagery, evaluating NDWI thresholding, a U-Internet skilled from scratch and a fine-tuned U-Internet.
The thesis was a chance to consolidate all this work. For reproducibility, I wished to create one repository that might be used to retrain any of the fashions. I did this utilizing Claude Code in VS Code, following these steps:
-
I requested the agent to learn all of the repos containing the modelling code and summarise the completely different experiments.
-
I requested the agent to provide you with a plan to consolidate all the fashions and output the plan together with the record of adjustable hyperparameters (i.e. modelling selections I may change). These hyperparameters might be modified utilizing flags when working the code by way of the command line.
-
I made corrections to the plan. This largely concerned correcting the record of hyperparameters.
-
Let it do its factor.
The end result was a single repository that might retrain any mannequin from this chapter of my thesis. This saved me a number of time that might in any other case have been devoted to the tedious process of copying and debugging code. Nevertheless, it was solely efficient as a result of I had such a great understanding of the code the consolidation was based mostly on. I may learn by means of the repo and understood if the agent made a mistake or an surprising change. As a ultimate verify, I retrained the important thing fashions utilizing the brand new repo and in contrast their analysis outcomes to those within the unique papers.
In the identical means {that a} good understanding of the analysis space was vital for the primary software, a great understanding of the code was vital for the second. If I had began from zero and requested the agent to write down all of the code for the in depth record of experiments, there’s a excessive probability that one thing would go unsuitable. Even when there have been no errors, I’d not be capable of belief that my understanding of the methodology matched what the code really did — a clear case of poor tutorial apply.
These first two functions lowered the time I spent on the ultimate thesis. Sadly, all that further time went proper again into this subsequent one. In brief, I used an AI instrument to extensively fact-check the thesis. Other than discovering many errors, this course of taught me a lot extra in regards to the subject I had spent the final 4 years researching.
3) Truth checker for citations
A thesis is a fancy piece of labor, in my case containing roughly 200 pages and over 200 citations. A bit demotivating, however solely 2-3 persons are ever going to learn it, and much more demotivating, solely you’re actually going to care about it. This makes discovering errors exhausting. It requires you to meticulously comb by means of your individual writing, rechecking each quotation, determine and metric, and even then you’re additionally virtually assured to overlook one thing. To assist, I used AI for lots of proofreading-related duties.
This contains utilizing it as a extra detailed spell checker that may take into account the context of phrases, in addition to a instrument to cross-reference all my metrics derived from bigger tables. Nevertheless, maybe probably the most helpful method was to check the data I cited in opposition to the unique papers I claimed to get that data from. Attaining this was a multi-step course of utilizing a Claude Code venture.
Step 1: rename all cited papers with their BibTeX quotation key
Assuming you’re working with LaTeX, by the top you’ll have a bibtex.tex file containing lots of of references just like the one you see beneath. The primary string, sun2023coastline, is the quotation key. All through a doc, you’ll cite work utilizing these keys. For instance, a passage from my thesis that cited this paper:
Lastly, optical aerial images gives a helpful companion for optical satellite tv for pc imagery. It’s the supply that’s historically used to digitise the shoreline~cite{sun2023coastline}.
To organise your papers, it’s good to rename them utilizing this quotation key. Fortunately, all through my PhD, I named the papers utilizing the important thing adopted by a touch after which a part of the paper’s title, like this:
sun2023coastline – Shoreline extraction utilizing distant sensing.pdf
This can make sense in the case of step 3 of this course of. As we are going to see, it permits you to programmatically cut up your papers based mostly on the quotation keys in a bit of textual content.
Step 2: convert all papers to .txt information
This step is optionally available. The Claude Code venture context window is proscribed, and I discovered that some massive papers may take up most of it. The best way I bought round this was writing a Python script that extracted all of the textual content from the PDFs and saved them as .txt information. So as an alternative of
sun2023coastline.pdf
we may now add
sun2023coastline.txt.
Doing so allowed me to add extra papers for any given immediate.
Step 3: divide papers into sections and group cited papers accordingly
For a simpler reality verify, I discovered it higher to separate up the thesis into manageable chunks. For instance, the background chapter had 4 subsections. I cut up every of those into its personal textual content file. I then used a Python script to seek out all of the quotation keys in a given chunk of textual content and create a brand new folder containing all of the related papers. This makes the method of organising a fact-checking run smoother.
Step 4: run the actual fact checker
As seen in Determine 3, I then uploaded all of the textual content information to the Claude Code venture together with the bibtex.txt file containing all of the references. I then pasted the related chunk of textual content and requested the agent to fact-check it. When organising the venture, I additionally gave it an in depth set of directions. The complete immediate will be discovered on the finish of this text. It gives context about my thesis, a proof of the paper naming conference and detailed directions on how you can divide errors based mostly on their severity and what data to output when an error is discovered.
Utilizing a Claude Code venture to fact-check a PhD thesis in opposition to its cited papers and BibTeX file.
I went part by part by means of the thesis, with every step importing the brand new set of related papers and eradicating the previous ones. This was vital because it lowered the quantity of context for every fact-checking iteration, serving to me keep away from hitting a utilization restrict mid-prompt. Extra importantly, it broke the method up into manageable steps.
The bottleneck was me studying the errors, evaluating them to the cited textual content and confirming they had been real. In some circumstances, I had to return to reevaluate my understanding of a subject. This may shortly develop into overwhelming had I carried out it for the complete thesis directly.
Ultimately, a lot of the errors had been nitpicky finer particulars, one thing an knowledgeable within the subject (i.e. your examiner) would nonetheless decide up on and that, collectively, may end in a weaker thesis. Did it discover all of the errors? In all probability not. Did it discover many who I’d by no means have noticed? Undoubtedly sure. Correcting them will hopefully make the subsequent stage of the PhD journey somewhat smoother.
4) Defence query generator
This subsequent software is one thing I plan to do in a couple of months, and it’s impressed by how I’m utilizing AI proper now — getting ready for interviews. Sadly, submitting my thesis additionally coincides with me rejoining the job market, and interviewing is just not one thing I’ve carried out in over 4 years. To cowl the probably generic questions, I comply with this course of:
-
Add my CV.
-
Add the job description and different associated firm data.
-
Ask the chatbot to supply an inventory of 10-20 interview questions, with some context about the kind of interview (offering the precise transient if accessible).
-
I then begin an iterative spherical of questions with the immediate beneath.
-
With every query, I exploit the dictation performance to document my reply.
-
After studying the suggestions, I’ll typically reanswer the query if there are good factors.
No affords but, so it stays to be seen how helpful this technique is. Even so, I plan to do one thing comparable for my thesis defence. On this case, I’ll add the complete thesis and a few background about my examiners. This ought to be good apply verbalising the arguments that I’ve in textual content. With about two months between thesis submission and my defence, that is extra a means of reminding me what I wrote than predicting the precise questions that will probably be requested.
If it wasn’t already, it’s throughout the defence that it’ll develop into obvious how you’ve got used AI. In the identical means that it’s troublesome to totally perceive a fancy piece of code you did not write, it will likely be troublesome to memorise the textual content an AI instrument has outputted. I do know if I had leaned on it closely to do my writing, I’d probably fumble a query a couple of area of interest matter or reasoning behind some methodological alternative. Extra ideologically, I believe, with this method, you’re doing your self a disservice as a scientist.
Writing is pondering. It’s true that the quotation finder, reality checker and even code consolidator wouldn’t have labored as effectively if I had not struggled by means of the writing and modifying of the thesis beforehand. Extra importantly, it’s by means of this course of that I got here up with a few of my finest concepts. Revisiting previous work sparked new evaluation, visualisations and outcomes from the identical uncooked information, and it’s only from this foundation that AI may make enhancements.
···
I hope you loved this text. If you wish to hear extra from me, try my YouTube channel, and if you wish to give me a job, discover me on LinkedIn 😉
···
