Today is a big day at LightOn as we unveil a hardware product, the Appliance, the world's first commercially available photonic co-processor for AI and HPC
We have had a few of these optical processing units in our own LightOn Cloud for the past two years and just retired one after more than 800 days working full time.
In the coming days, we'll be making another announcement but I wanted to first share a video we did recently. At LightOn, we don't build photonic computing hardware because it's fancy or cool (even though, it is cool) but because computing hardware is hitting the limits. I know what some say about Moore's law not being dead but the recent focus on Transformers and their attendant scaling laws makes it obvious that in order for more people to have access to these models, we need a new computing paradigm. Indeed not everyone can afford to spend a billion dollars in training these models. As Azeem was recently pointing out in one of his newsletters, this is how bad things will become:
The amazing thing is that we can start to compare the cost of training single AI models with the cost of building the physical fabs that make chips. TSMC’s state-of-the-art 3nm fabwill run to around $20bn when it is completed in two years. A fab like this may be competitive for 5-7 years, which means it’ll need to churn out $7-8m worth of chips every day before it pays back.
And so at LightOn, we think that a combination ofalgorithms and (cool) hardware as the only pathway forward for computing large-scale AI. The video is right here, enjoy!
A curated list of resources on implicit neural representations, inspired by awesome-computer-vision. Work-in-progress.
This list does not aim to be exhaustive, as implicit neural representations are a rapidly evolving & growing research field with hundreds of papers to date.
Instead, this list aims to list papers introducing key concepts & foundations of implicit neural representations across applications. It's a great reading list if you want to get started in this area!
For most papers, there is a short summary of the most important contributions.
Disclosure: I am an author on the following papers:
Implicit Neural Representations (sometimes also referred to coordinate-based representations) are a novel way to parameterize signals of all kinds. Conventional signal representations are usually discrete - for instance, images are discrete grids of pixels, audio signals are discrete samples of amplitudes, and 3D shapes are usually parameterized as grids of voxels, point clouds, or meshes. In contrast, Implicit Neural Representations parameterize a signal as a continuous function that maps the domain of the signal (i.e., a coordinate, such as a pixel coordinate for an image) to whatever is at that coordinate (for an image, an R,G,B color). Of course, these functions are usually not analytically tractable - it is impossible to "write down" the function that parameterizes a natural image as a mathematical formula. Implicit Neural Representations thus approximate that function via a neural network.
Why are they interesting?
Implicit Neural Representations have several benefits: First, they are not coupled to spatial resolution anymore, the way, for instance, an image is coupled to the number of pixels. This is because they are continuous functions! Thus, the memory required to parameterize the signal is independent of spatial resolution, and only scales with the complexity of the underyling signal. Another corollary of this is that implicit representations have "infinite resolution" - they can be sampled at arbitrary spatial resolutions.
This is immediately useful for a number of applications, such as super-resolution, or in parameterizing signals in 3D and higher dimensions, where memory requirements grow intractably fast with spatial resolution.
However, in the future, the key promise of implicit neural representations lie in algorithms that directly operate in the space of these representations. In other words: What's the "convolutional neural network" equivalent of a neural network operating on images represented by implicit representations? Questions like these offer a path towards a class of algorithms that are independent of spatial resolution!..........
The scaling hypothesis motivates the expansion of models past trillions of parameters as a path towards better performance. Recent significant developments, such as GPT-3, have been driven by this conjecture. However, as models scale-up, training them efficiently with backpropagation becomes difficult. Because model, pipeline, and data parallelism distribute parameters and gradients over compute nodes, communication is challenging to orchestrate: this is a bottleneck to further scaling. In this work, we argue that alternative training methods can mitigate these issues, and can inform the design of extreme-scale training hardware. Indeed, using a synaptically asymmetric method with a parallelizable backward pass, such as Direct Feedback Alignement, communication needs are drastically reduced. We present a photonic accelerator for Direct Feedback Alignment, able to compute random projections with trillions of parameters. We demonstrate our system on benchmark tasks, using both fully-connected and graph convolutional networks. Our hardware is the first architecture-agnostic photonic co-processor for training neural networks. This is a significant step towards building scalable hardware, able to go beyond backpropagation, and opening new avenues for deep learning.
Aurélien sent me an email back in October and we are now in December! Time flies.
Dear Igor,
I hope things are well.
I have been following your NuitBlanche blog for quite a few years. It would thus be great for us if you consider a recent paper of ours to appear in your blog, entitled “Diffraction-unlimited imaging based on conventional optical devices”. This paper has been published in Optics Express this year and its link is: https://www.osapublishing.org/oe/abstract.cfm?uri=oe-28-8-11243
This manuscript proposes a new imaging paradigm for objects that are too far away to be illuminated or accessed, which allows them to be resolved beyond the limit of diffraction---which is thus distinct from the microscopy setting. Our concept involves an easy-to-implement acquisition procedure where a spatial light modulator (SLM) is placed some distance from a conventional optical device. After acquisition of a sequence of images for different SLM patterns, the object is reconstructed numerically. The key novelty of our acquisition approach is to ensure that the SLM modulates light before information is lost due to diffraction.
Feel free to let us know what you think, and happy to provide more information/pictures if needed. Thanks a lot for your time and consideration!
We propose a computational paradigm where off-the-shelf optical devices can be used to image objects in a scene well beyond their native optical resolution. By design, our approach is generic, does not require active illumination, and is applicable to several types of optical devices. It only requires the placement of a spatial light modulator some distance from the optical system. In this paper, we first introduce the acquisition strategy together with the reconstruction framework. We then conduct practical experiments with a webcam that confirm that this approach can image objects with substantially enhanced spatial resolution compared to the performance of the native optical device. We finally discuss potential applications, current limitations, and future research directions.
A combination of post-Moore’s law era and the advent of very large ML models require all of us to think up new approaches to computing hardware and AI algorithms at the same time. LightOn is one of the few (20) companies in the world publishing in both AI and hardware venues to engage both communities into thinking how theories and workflows may eventually be transformed by the photonic technology we develop.
This year, thanks to the awesome Machine Learning team at LightOn, we have two accepted papers at NeurIPS, the AI flagship conference, and have five papers in its“Beyond Backpropagation” satellite workshop that will take place on Saturday. This is significant on many levels, not the least being that these papers have been nurtured and spearheaded by two Ph.D. students (Ruben Ohana and Julien Launay) who are doing their thesis as LightOn engineers.
Here is the list of the different papers accepted at NeurIPS this year that involved LightOn members:
Direct Feedback Alignment Scales to Modern Deep Learning Tasks and Architectures, Julien Launay, François Boniface, Iacopo Poli, Florent Krzakala (Presenter: Julien Launay).
Ignorance is Bliss: Adversarial Robustness by Design through Analog Computing and Synaptic Asymmetry, Alessandro Cappelli, Ruben Ohana, Julien Launay, Iacopo Poli, Florent Krzakala (Presenter: Alessandro Cappelli). We had a blog post on this recently.
Align, then Select: Analysing the Learning Dynamics of Feedback Alignment, Maria Refinetti, Stéphane d’Ascoli, Ruben Ohana, Sebastian Goldt paper (Presenter: Ruben Ohana).
How and When does Feedback Alignment Work, Stéphane d’Ascoli, Maria Refinetti, Ruben Ohana, Sebastian Goldt. paper (Presenter: Ruben Ohana)
Ever since we started LightOn, we have been putting some emphasis on having great minds think how new algorithms are possible and how they can be enabled with our photonic chips. We also have a regular meetup where we see how other great minds are devising new algorithms.
Tomorrow, Thursday (October 15th) we are continuing this journey by having Adam Gaier who will talk to us about Weight Agnostic Neural Networks. The virtual meetup will start at:
In the 2000s, after featuring good work on Nuit Blanche, I was usually following through by asking authors where their codes were. This is how the implementation tag was born. Some of the answers were along the lines of: "I didn't make it available because I thought it was not worthy". But what I usually responded was that, in effect, releasing one's code had a compounding effect on the community:
"You may not think it's worthy of release, but somehow, someone somewhere needs your code for reasons you cannot fathom"
As a result, I made a conscious choice of featuring those papers that were actively featuring their implementations. The earliest post with featured implementations was February 28th, 2007 with a blog post featuring three different implementations of reconstruction solver for compressed sensing. Yes, implementations were already available before that, but within the compressive sensing community, it was a point in time with a collective realization that releasing one's code would bring others to reuse one's work and advance the field as a result. At some point, I started making a long list of implementation available but got swamped after a while because it became, most of the time, the default behavior (a good thing).
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email.
