rassibassi

CFG (context free grammar), which is explained in the book, is used here together with LLMs: https://tree-diffusion.github.io

In this context, RAG isn't what's being discussed. Instead, the reference is to a process similar to monte carlo tree search, such as that used in the AlphaGo algorithm.

Presently, a large language model (LLM) uses the same amount of computing resources for both simple and complex problems, which is seen as a drawback. Imagine if an LLM could adjust its computational effort based on the complexity of the task. During inference, it might then perform a sort of search across the solution space. The "search" mentioned in the article means just that, a method of dynamically managing computational resources at the time of testing, allowing for exploration of the solution space before beginning to "predict the next token."

At OpenAI Noam Brown is working on this, giving AI the ability to "ponder" (or "search"), see his twitter post: https://x.com/polynoamial/status/1676971503261454340

The diskcache docs state:

""" Caching Libraries

    joblib.Memory provides caching functions and works by explicitly saving the inputs and outputs to files. It is designed to work with non-hashable and potentially large input and output data types such as numpy arrays.

What's the difference to using joblibs Memory class similar to this implementation:

https://github.com/stanfordnlp/dspy/blob/main/dsp/modules/ca...

If I understand you correctly, you basically mean some interpolation of knowledge within Z, that was not known to humans before, but can be synthesized by recombination of information within Z.

Yet, that still means, knowledge of O cannot be synthesized.

The upper bound is still the Shannon limit. The experiment does a lot of multiplexing: spatial multi-core fiber, spectral multi channel multiplexing across wavelength, dual polarization.

Each of the multiplexed channels are individually limited by the Shannon limit, and with higher power the fiber's Kerr effect creates interference which creates a sweet spot for the optimal optical launch power.

the novelty here is that the spectral channels are all generated from a single laser source rather than a laser per channel

Not practical yet, the novelty is the frequency comb which allows +200 channels across wavelength with only a single laser, where before one required 200 lasers.

In an experiment like this, only the initial light source is modulated and therefore all channels carry the same data. The equipment for the transmitter and receiver chain is so expensive that university labs can barely afford one of each.

They also multiplex across +200 channels across wavelength (wavelength division multiplexing).

Not sure what the baud rate of a single channel was in their experiment but probably between 32-80Gb which is common for the lab equipment at Universities. The industry is knocking on 100-400Gb where for the actual decoding and signal processing there is massive parallelism applied to reduce the rate even more

Correction: the auto encoder is pre-trained :)

Yes, and another reason for the small model size and the novelty of the underlying paper [1], is that the diffusion model is not acting on the pixel space but rather on a latent space. This means that this 'latent diffusion model' does not only learn the task at hand (image synthesis) but in parallel also a powerful lossy compression model via an outer auto encoder structure. Now, the number of weights (model size) can be reduced drastically as the inner neural network layers act on a lower dimensional latent space rather than a high dimensional pixel space. It's fascinating because it shows that deep learning at its core comes down to compression/decompression (encoding/decoding), with close relation to Shannon's Information Theory (e.g. source coding/channel coding/data processing inequality).

[1] https://arxiv.org/abs/2112.10752

Don't be so hard on yourself, and don't compare yourself to a computer science prodigy. You can teach yourself to become a decent web developer in 1 year and take it from there, your life is far from being wasted in any sense.

Nice, saw you also have a section on fonts! Think the following could make you smile, too. Checkout this font that changes with the users facial expression (shameless plug):

Description: https://danishdesignaward.com/en/arkiver/nominee/adam-lenzin...

Demo: https://facetype3000.herokuapp.com

tribute radio show :)

https://www.nts.live/shows/the-do-you-breakfast-show/episode...

Nice documentary on schiele

https://www.arte.tv/en/videos/073077-000-A/egon-schiele/

Friend of mine created a parameterised font for his master thesis. We hooked it up to a face expression sentiment analyser (pre trained DNN). It works in the browser with your webcam, better to use a laptop or pc :)

http://facetype3000.herokuapp.com/

Had similar intentions half a year ago, it really clicked for me after watching the linear dynamical systems lecture by Boyd. How rank of a matrix, matrix norm and singular values relate was an eye opener. Thereafter it was easy to connect singular values and the stability/robustness of a system intuitively. Great teacher, I can only recommend :)

https://www.youtube.com/playlist?list=PL06960BA52D0DB32B

Nice video lecture series by scot

https://video.ethz.ch/speakers/bernays/2019/7b11b50e-f813-4d...

I agree that the smartphone could potentially be leveraged to do some audio processing in this context, but you have to consider the latency introduced by the forth and back over bluetooth. Not sure if such latency is a pleasent experience for the user.

We are also in Denmark! Feel free to reach out if you're living in Copenhagen :) https://www.hackyourfuture.dk/

Support for 4! Yet, my understanding is that particle filters are superior but computational more demanding. For nonlinear problems, the extended Kalman Filter linearizes the task, whereas particle filters don't and work with many point estimates instead.

I loved this book: https://users.aalto.fi/~ssarkka/pub/cup_book_online_20131111...

and also Thomas Schoen group does great work on Sequential Monte Carlo (SMC), MCMC for sequential data :) http://user.it.uu.se/~thosc112/index.html

They are also building a probabilistic programming language for sequential data! https://github.com/lawmurray/Birch

I received my PhD in December, so I already pursued further education. My masters was in telecommunications, starting my PhD with a background in signal processing/information theory rather than optics/physics. All the master students I supervised either had similar background like me or an optics/physics background. That said, optical communications combines multiple disciplines, which makes it so fascinating, and I think that a PhD is inevitable if you want to pursue a career here. There are so many steps to understand to get from bits to electrical signal to optical signal and all the way back to bits. I don't think a masters degree can cover all of it in depth. Further, a PhD is in particular helpful due to the conferences where all the big players are looking out to hire you, or where one meets senior researchers for potential postdoc positions.

Salary-wise it's difficult for me to answer, as most of the big industry players are in the US or Canada, but I'm in Europe. From hearsay a fresh PhD with reasonable publication list will get around 10k per month in the bay area. However, glassdoor might give you a better idea. Look for companies like: Infinera/ciena/acacia communications/juniper/mellanox/finisar/keysight...

After my PhD, I probably could have pursued a postdoc somewhere in Europe, but I decided to leave academia. I wanted to stay in Copenhagen and therefore had to change my field of work and will be working for a hearing aid company in their signal processing department.

Recent graduate in Optical Communications here :)

1) The fiber is pretty thin, but there are several fibers combined in one transoceanic cable, see here https://en.wikipedia.org/wiki/Submarine_communications_cable. They don't patch them, but splice the fibers. Patching introduces a little loss, so it should be avoided.

2) Not sure how deep they are buried, but I think to remember that the shore end of a submarine cable is better protected than the part in the deep ocean. Due to more ship traffic at the coast etc.

3) Good question. The loss of an optical fiber is roughly 0.2 dB/km. Across the ocean the optical signal must be amplified several times (every 80-100km). Nowadays the amplification is all optical (EDFA or Raman). Before there were electrical regeneration schemes. Check out the history of the field, it's is quite interesting [1].

4) Not sure how much data such a cable can carry, since it depends on how many fibers are deployed within. However, there are multiple interesting things to look into here. In research labs, people are investigating multicore/multimode fibers (space division multiplexing) [2], these fiber have incredible capacity. Personally, I think the most interesting metric is the spectral efficiency, so how much data can be transmitted per second per Herz. Such a metric is independent of multiplexing schemes over space/wavelength/time, and improvements have to come from better devices, signal processing or signal shaping methods [3]. Another mind-blowing area is using the nonlinear Fourier transform for better signaling methods [4].

Feel free to ask more questions :)

and checkout the two biggest conferences for more in depth info.

https://www.ofcconference.org/en-us/home/

https://www.ecocexhibition.com/

[1] https://www.osapublishing.org/oe/abstract.cfm?URI=oe-26-18-2...

[2] Multicore: https://ieeexplore.ieee.org/abstract/document/7341685

Multimode: https://www.osapublishing.org/abstract.cfm?uri=ofc-2018-Th4C...

Multimode+core: https://ieeexplore.ieee.org/document/8535233

[3] https://arxiv.org/abs/1606.04073

[4] https://www.osapublishing.org/optica/abstract.cfm?uri=optica...

Cuddle my girlfriend, we became pretty good at it so far

changed my perspective on research

could you elaborate on that? pro research or against?

I'm about to finish my PhD in the field of optical communications and applied machine learning to increase the data rate. More specifically, I used an autoencoder with an embedded fiber channel model, to learn something called a geometric constellation shape or modulation format. (if you ever heard of QAM then you're on the right track, you can think of it as a modern Morse code for transmitting bits) Thereafter, we took the learned constellations to the lab and conducted an experiment, transmitting the learned constellations through actual fiber. I've learned so much during the whole process

Although, the gains are marginal, I like the method a lot since it combines physics (nonlinear schroedinger equation derived fiber channel model), information theory (optimizing for mutual information) and machine learning. It wouldn't have been possible without the people who published the fiber channel model, my colleagues and in particular the colleagues who could help me in the lab.

The debugging was hell, there are so many dimensions where stuff can go wrong (besides the usual bugs): physical parameters with the wrong unit, the implementation of the fiber model in tensorflow, the machine learning parts with its training process. Plus the things that can go wrong in the lab.

There are still pieces where I'm not 100% sure, and would love to speak to someone with some background in autoencoders.

I've open sourced the autoencoder and the fiber channel model, checkout my github with the same username as here!

http://dtu-dsp.github.io/Robochameleon/htmldoc/html/index.ht...