Speech-to-text with Whisper

Whisper is another neat productivity tool that has been translated to a high-performance model that can be run without specialized hardware, even on your phone!

The speed and accuracy is remarkable - it takes only a few minutes to create a transcript of an hour-long seminar. While these capabilities have been around for some time (e.g. subtitle options in Youtube and video conferencing programs), it is great there are now fast, open source tools that can be run locally, without an internet connection or the privacy risks of sending your data to some untrusted server on the cloud.

Some potential applications in research:

• Brainstorming - discussions can be transcribed to a text format that can be more easily reviewed later (e.g. searching for keywords).
  
• Paper drafting - often when writing or typing we fall into the habit of writing long convoluted sentences that need heavy editing to make them more readable and digestible. Dictating parts of a paper might be a better way to come up with clear and concise text.
• Converting recordings of conference / workshop talks into full-length paper drafts or conference proceedings. I am trying this one out on one of the online talks I gave during covid.

The ability to quickly and accurately convert research between different formats (text, audio, visual, different languages, etc.) will ultimately improve the accessibility of research, ensuring that it is open for all to use and build on. Further reading on this important initiative can be found on the arXiv blog.

 

How to run generative text models on your own computer or phone for free

It was only a matter of time - the resource-intensive, proprietary, pay-to-use ChatGPT now has a free-to-run and use competitor in the form of a tweaked version of Facebook's LLaMa model!

llama.cpp is an super-fast implementation of LLaMa in C++ obtained by reducing the numerical precision of the trained model weights. 

I installed it without any trouble, in contrast to other generative models where dealing with the python dependencies and incompatible library versions can be a massive pain.

Similar to ChatGPT, the raw generated output text might look convincing to a non-specialist even though it is spouting nonsense. Thus, careful editing of its output is essential if you want to use it for anything other than your own amusement. 

For example:

Prompt: Quantum neural networks and quantum kernel machines are two promising approaches for machine learning using emerging noisy intermediate-scale quantum computers [1].

Output: Quantum neural networks and quantum kernel machines are two promising approaches for machine learning using emerging noisy intermediate-scale quantum computers [1]. They exploit the inherent parallel nature of existing classical deep network architectures through encoding them into a series of unitary gates, which can then be applied to qubits. However in most cases their expressive power is severely limited by noise that quickly accumulates during gate execution and corrupts information encoded on quantum states [2]. Although methods have been developed for dealing with this issue (either by using error correcting codes or applying a particular type of operations, namely, magic state techniques), they are not fully successful in practice. In other words, current approaches to mitigate errors suffer from an excessive overhead that makes these methods impractical and therefore useless [3]. In this paper we describe novel quantum circuits for encoding NNs architectures onto qubits with a significantly reduced number of gates and without using magic states or error correcting codes. These new constructions are not only more efficient, but also less prone to errors since they employ fewer physical operations (and therefore have lower noise), which could lead them into practical use in future NISQ devices [4].

The pace of machine learning research is incredible.

Hype and anti-hype

Claims of high temperature superconductivity were yesterday published in Nature and presented at the APS March Meeting. Given the history of the group, discussed in detail during a workshop on reproducibility in condensed matter physics, no doubt this should be taken with a pinch of salt.

On arXiv yesterday: Russians tear down claims of QAOA-accelerated factorization algorithms which hit news headlines last December. The comments on Scott Aaronson's blog on the original paper have some amusing (or depressing) background on the group behind this work.

Similarly, a few weeks ago claims of quantum simulation of wormhole dynamics using superconducting processors were heavily criticized.

These examples are all high profile works which have been (and will be) carefully scrutinized. The vast majority of preprints and publications do not attract as much interest. If you're having trouble reproducing a result in a paper, keep in mind that the paper may have errors that went undetected through peer review. The real peer review begins after the paper is published.

Academic working conditions

Comments on a Comment published a few weeks ago in Nature Human Behaviour: Quality research needs good working conditions.

There is "a lack of permanent positions with dedicated research funding, leading to an overreliance on project-based funding with short-term research positions." Unless one lands a tenured position, the longest term of reliable employment one is likely to encounter is during the PhD studies (3+ years). Postdocs depend on periodic renewal of short-term 1 or 2-year contracts and face the prospect of probably needing to move to a different city or country if they need to take up a new position. It's hard.

The relation between turnover and research quality is nuanced. Yes, in certain circumstances pressure can foster creativity and productivity, but the threat of non-renewal is not the only form of pressure that can be applied. Internal project deadlines ("if we don't have a promising result by x, let's try a different line of research"). Performance bonuses. Conference submission deadlines. There are many possibilities.

On the flip side, there are many examples of stagnant research institutions dominated by faculty members nearing retirement with no incentives to undertake quality new research. The problem with research is that it isn't always successful. It is hard for an outsider to distinguish hard but unfruitful work from slacking off.

Unfortunately the terms of employment contracts are typically controlled by upper university management. Thus, even those tenured faculty who would prefer to have a more stable team composition have little power to effect change.

 

Quantum error correction - now published

The paper "Suppressing quantum errors by scaling a surface code logical qubit" by the Google Quantum AI team was published yesterday in Nature. I previously wrote about this work when the preprint was posted to arXiv last year.

The peer review file accompanying the paper is an interesting read. The authors mention challenges involved in making superconducting quantum computers that are robust to catastrophic noise induced by cosmic rays - radiation shielding will not be sufficient:

"Given the current error scale of an impact event is essentially unsurvivable, the event rate needs to be at least on the order of the computation time. Reasonable estimates for fault tolerant computations are generally measured in hours (c.f. [5] which proposes 8 hours for Shor’s algorithm), so current event rates should need to improve by around 3000x from current event rates. Additionally, if chip areas grow approximately proportional to number of qubits, given that the event rate is directly proportional to chip area, the event rate will also increase by around 100,000x from this effect, which will need to be overcome as well. The 10x reduction offered by the use of lead shielding and a low-radiation laboratory is nowhere near sufficient to solve this problem at scale."

Solving this problem will require new superconducting qubit and quantum processor designs that can be more robust to cosmic rays, e.g. localizing their disruptive effects to a small fraction of the qubits so that quantum error correction can still be applied.

 The authors also write in their reply to referees that the performance fine-tuning required for the distance-5 code to match the performance of the distance-3 code took place over 6 weeks.

As I wrote before - this is an impressive achievement, but in the race to build a working, fault-tolerant quantum computer it should be seen of the end of the beginning, not the beginning of the end!

Snippets from the topological data analysis workshop

 I had the pleasure of visiting KIAS last week to attend their Workshop on Topological Data Analysis: Mathematics, Physics, and Beyond.

The programme featuring so many pure mathematics-focused talks was daunting at first, but ultimately I learnt a lot more than I would have by presenting the same work at a physics conference. This serves as a nice reminder: When we become experts at a highly specialised topic in our PhD studies (and beyond), it is easy to lose sight of the bigger picture. Changing fields or even attending a wide range of seminars outside our own expertise sparks new ideas.

Some useful tidbits from the talks:

• The Vietoris-Rips complex is not stable with respect to outliers. For example, adding a few points to a middle of a circular point cloud will completely change the form of the 1D persistence diagram (replacing the single high persistence cycle with low-persistence cycles). Stability theorems for persistence diagrams refer to small perturbations of a fixed set of point.
• Similar persistence diagrams do not imply similar data, in fact datasets with an arbitrarily large Hausdorff distance can have the same persistence diagrams.
• Mathematicians like TDA because it involves interesting problems with elegant solutions. But not all these elegant methods end up being useful in practice.
• The chemical and biological sciences benefit from having large datasets publicly available for benchmarking and standardised scoring of new machine learning / data analysis methods including TDA. This allows demonstrations of new heuristic methods to be convincing. 
• TDA can be used to estimate geometric features of data, including their (fractal) dimension, which is useful for understanding the performance of neural networks.
• The generators of 0-dimensional persistence diagrams give the minimal spanning tree for the data. This widely-used "folklore" was only rigorously proved in 2020!
  

Entanglement-enhanced quantum sensing: fact or fiction?

Quantum sensing, the use of quantum systems to perform precision measurements, attracts enormous interest as an potential application of engineered quantum systems being developed. According to this review article, there are three classes of quantum sensing:

1. Sensing based on systems with quantized energy levels such as superconducting qubits, including SQUID magnetometers (already commercialized).

2. Sensing based on quantum coherence or wave-like properties, including noise suppression using squeezing (e.g. in gravitational wave detectors) and macroscopic quantum states of atoms for precision gravimetry and inertial navigation (under development / being commercialized).

3. Entanglement-enhanced sensing to achieve precision beyond what is attainable classically (research in progress). 

It is believed that only entanglement-enhanced sensing makes use of the full power of quantum mechanics (i.e. many-body entangled states intractable for classical computers). 

The growing availability of large controlled quantum systems has led to huge interest in entanglement-enhanced sensing schemes, with many publications in high impact journals, but not everyone is convinced.

Critiques of recent high-profile experiments on entanglement-enhanced sensing published in Nature and Nature Physics have been posted to arXiv: arXiv:2208:14816, arXiv:2301.04396, and (today) arXiv:2302.00733. The first is a particularly interesting read, since it has been updated to include correspondence with the paper authors and Nature Physics editors, who declined to publish it.

I do not work in this field. I do not have the expertise to judge whether the criticism is valid or not. But it seems to me that the comments come from a knowledgeable expert, are written in a scientific style, and are of a reviewable standard. Moreover, the claims in the critiqued articles (unprecedented sensitivity at measuring some quantity) are quantitative, and can thus be unambiguously proved or disproved. Thus, it should be concerning that while one of the articles claiming entanglement-enhanced sensitivity has already been cited 30 times according to Google Scholar, the criticism seems to be ignored - not cited, not responded to, not even upvoted on scirate.

Several years ago there was a similar controversy in photonics, with many researchers racing to be the first to claim to demonstrate lasing in a variety of exotic materials. In response to this, Nature Photonics introduced a "laser checklist" to ensure that all submissions reporting claims of lasing provide a standardized set of measurements and experimental details which can be scrutinized and easily compared between different platforms and research groups. Perhaps something similar can be done for entanglement-enhanced sensing papers?