@lampos.net
Associate Professor @ UCL Computer Science | research: AI, NLP, health | dad | 🇺🇳 | ☮ | here I will be focusing more on less scientific matters, go to https://x.com/lampos for research-related content Website: lampos.net
Microsoft Sharepoint... 😡 😡 😡 I do not how they do it, but universities manage to opt-in for the worst software platforms possible. Avoid Sharepoint at all costs, if you can. Then, never use WISEflow either.
One would expect that ML experts would be flagging these things during the peer review process of top-tier ML conferences. But they don't. Because the ML expert sub-cohort is probably a minority in the pool of reviewers.
Example 2: a weather prediction task that encompassed about a year of data from one weather station (do I need to explain to you how idiotic that is / no seasonal modelling / no spatial modelling / no practical value in this model / plus rich weather data sets are available).
Ah, and never mind the many bogus benchmarks (data sets) people use to evaluate models. Example 1: an influenza prediction task that used historical influenza-like illness rates to forecast rates one year ahead or longer (which does not appear to be a plausible biological task).
And of course, many "SOTA" models out there predict garbage at the target forecasting horizon. They are unusable, but still advertised as the best.
This is the same as telling you that in 5 days from now the temperature will be 30C, but evaluating the accuracy of that prediction using all my (easier) predictions from now till the target forecasting horizon (5 days from now).
Current state of science in a nutshell: SOTA time series forecasting models that predict a sequence of outputs up to a time step t+H (t: current time step, H: forecasting horizon) evaluate their forecasting accuracy using the entire output sequence.
The quality of our EMNLP reviews (mostly rubbish) can be summarised by the typo the reviewer found in our paper. According to them "e.g." should by followed by a comma as in "e.g.,". My dear reviewer this is not the case for British English. We are paying the price of corporate "science".
This is yet another great lesson that idiots should never be in positions of power. Yet, no smart person wants to govern. Perhaps, it is about time for a change of perspective on this. People might be forced to become political again when less is left to lose, I suppose.
Interesting interview with a sane explanation for Trump's tariffs. Plus a great commentary about AI+tech lords. Yanis Varoufakis (@yanisvaroufakis.bsky.social) certainly gets economics better than many, but he also understands the AI space better than AI "experts".
www.youtube.com/watch?v=Ax7T...
Talent [?] shows and autotune. Funny!
www.youtube.com/watch?v=Xqrc...
28 days ahead forecasting results for influenza rates for the season 2018/19 in England for all models in our study. DeformTime, the model we proposes, performs best (and quite accurately), whereas some of the other forecasters (e.g. PatchTST, iTransformer, TimeXer) are very similar to a persistence model, i.e. they are performing very poorly.
"DeformTime: capturing variable dependencies with deformable attention for time series forecasting" led by Yuxuan Shu
Just out @tmlrorg.bsky.social (TMLR)
Link: openreview.net/forum?id=M62...
Plus a teaser figure comparing SOTA forecasters (e.g. TimeXer, iTransformer) with a persistence model!
Keir Starmer's statement was deeply confusing. He referred to peace without a deal in place. Yet, he mentioned increased defence budgets and NATO arms on the ground (of Ukraine). I'm not sure Russia will go along with this. My fear is that more intense war will follow.
Perhaps we should thank Trump for showing us how the United States of America is bullying other countries into submission. I have never liked bullies.
www.youtube.com/watch?v=Ch0q...
The land of the free is now directly governed by a billionaire, whereas in the previous "democratic" iteration, it was only an indirect relationship.
Fascism is fed from oligarchy and vice versa. What's new to learn?
Life expectancy losses in the Gaza Strip during the period
October, 2023, to September, 2024
TLDR: Life expectancy was reduced to half in the Gaza Strip area (of Palestine) compared to prewar levels
Paper: www.thelancet.com/journals/lan...
Thank you for your interest in this opportunity. I have now invited successful candidates for an interview and will not be considering further applications (for this particular scholarship).
Nowadays, new PC products tend to be thoroughly tested by the consumers and not the companies that manufacture them.
www.youtube.com/watch?v=Ndmo...
#Nvidia #RTX5090
One of the guys from DOGE who's been let loose in the Treasury payments system wrote a book about using deep learning to predict the stock market. Over the weekend I wrote a short review of it.
magnusross.github.io/posts/sparta...
To time series forecasting people:
Can you please stop evaluating forecasts using the entire output series? It doesn't tell much about the accuracy of a forecast at the actual target forecasting horizon.
And could you please try to understand a task prior to trying to model it?
Thanks!
ChatGPT seems to be a more developed end-user LLM than DeepSeek. However, DeepSeek and its RL/R1 module (DeepThink) is way more fun. The sole existence of DeepSeek that currently contradicts the OpenAI business model is the most fun of all. chat.deepseek.com
Useful links
-
My website: lampos.net
UCL Centre for AI: www.ucl.ac.uk/ai-centre/
/9
These include a bi-annual career mentorship with a G-Research Quantitative Researcher, a place on G-Research's Spring into Quant Finance Programme during their second year, presentation and seminar invitations, an annual dinner, and brand ambassador status.
/8
G-Research will not be involved in the determination of the research topic or the research itself. The successful candidate, however, will have additional benefits for career development (all optional based on the candidate's preferences).
/7
Student applications will be ranked by a committee within the UCL Centre for AI and the top one will be offered the position. Although this is a competitive selection process, the small amount of eligible PhD supervisors means that there is a relatively greater chance of success.
/6
You will then be interviewed (by myself and another faculty member). If you are shortlisted, you will be invited to submit your final application by February 24, 2025.
/5
If you are interested, please send me (v.lampos@ucl.ac.uk) your CV and a short research proposal using this template (www.overleaf.com/read/cdjjmtm...). The deadline for this initial step is January 31, 2025. However, the sooner you submit this material and/or get in touch with me the better.
/4
The topic of the PhD can be about anything quantitative within AI / Machine Learning (ML) with the expectation to produce research outcomes that can be submitted to globally leading AI / ML conferences.
/3
The scholarship is funded by G-Research. PhD supervisors must be academic faculty members of the UCL Centre for Artificial Intelligence (AI).
/2