Article, Machine Learning, Philosophy

Notes on Gilbert Simondon’s “On the Mode of Existence of Technical Objects” and Artificial Intelligence

Happy new year ! This is the first post of 2025 and this time it is not a technical article (but it is about philosophy of technology 😄)

Gilbert Simondon (1924-1989). Photo by LeMonde.

This is a short opinion article to share some notes on the book by the French philosopher Gilbert Simondon called “On the Mode of Existence of Technical Objects” (Du mode d’existence des objets techniques) from 1958. Despite his significant contributions, Simondon still (and incredibly) remains relatively unknown, and it seems to me that this is partly due to the delayed translation of his works. I realized recently that his philosophy of technology aligns very well with an actionable understanding of AI/ML. His insights illuminated a lot for me on how we should approach modern technology and what cultural and societal changes are needed to view AI as an evolving entity that can be harmonised with human needs. This perspective offers an alternative to the current cultural polarization between technophilia and technophobia, which often leads to alienation and misoneism. I think that this work from 1958 provides more enlightening and actionable insights than many contemporary discussions of AI and machine learning, which often prioritise media attention over public education. Simondon’s book is very dense and it was very difficult to read (I found it more difficult than Heidegger’s work on philosophy of technology), so in my quest to simplify it, I might be guilty of simplism in some cases.

(more…)

by Christian S. Perone
Machine Learning, Math

The geometry of data: the missing metric tensor and the Stein score [Part II]

Credit: ESA/Webb, NASA & CSA, J. Rigby. / The James Webb Space Telescope captures gravitational lensing, a phenomenon that can be modeled using differential geometry.

Note: This is a continuation of the previous post: Thoughts on Riemannian metrics and its connection with diffusion/score matching [Part I], so if you haven’t read it yet, please consider reading as I won’t be re-introducing in depth the concepts (e.g., the two scores) that I described there already. This article became a bit long, so if you are familiar already with metric tensors and differential geometry you can just skip the first part.

I was planning to write a paper about this topic, but my spare time is not that great so I decided it would be much more fun and educative to write this article in form of a tutorial. If you liked it, please consider citing it:

Cite this article as: Christian S. Perone, "The geometry of data: the missing metric tensor and the Stein score [Part II]," in Terra Incognita, 12/11/2024, https://blog.christianperone.com/2024/11/the-geometry-of-data-part-ii/.

(more…)

by Christian S. Perone
Machine Learning, Programming

Torch Titan distributed training code analysis

I really like to peek into different ML codebases for distributed training and this is a very short post on some things I found interesting in Torch Titan:

Disable and control of Python’s garbage collector (GC): titan codebase disables the Python GC and then manually forces a collection in the beginning of every training step during the training loop. This makes sense, but I’m not sure what are the gains of doing it, I think doing every step can be too much and I’m not sure if taking control of GC would be worth for the gains you get by manually controlling it, especially depending on complexity of other dependencies you use, as this could cause unintended behavior that would be difficult to associate with the GC collection;

Custom GPU memory monitoring: titan has a custom class to monitor GPU memory that is quite nice, it resets peak stats and empty the CUDA caching allocator upon initialization. At every step then they collect the peak stats for both small and large pools by capturing the stats for active, reserved and also failed retries and number of OOMs. It is very common for people to just monitor max GPU usage externally from NVML, however, this ignores the fact that PyTorch uses a caching allocator and that you need to look at the internal memory management mechanism inside PyTorch. If you don’t do that, you will certainly be mislead by what you are getting from NVML;

Custom profiling context manager: they wrote a context manager for profiling, where they measure time it takes to dump the profiling data per rank. Interesting here that there is a barrier at the end, which makes sense, but this is often the pain point of distributed training with PyTorch + NCCL;

Measuring data loading: this is of minor interest, but I liked the idea of not iterating on data loader in the loop statement itself but manually calling next() to get the batches, that makes it easier to measure data loading, that they average at the end for each epoch;

Logging MFU (model FLOPS utilization): they also compute and log MFU, which is quite helpful;

Delete predictions before backward: titan also deletes the model predictions before the backward() call to avoid memory peaks. This can be quite effective since you really don’t need this tensor anymore and you can delete it immediately before the backward pass;

Reduction of NCCL timeout: after the first training step, they reduce the NCCL timeout from the default 10 min to 100 sec. This is nice if you have well behaved replicas code and don’t need to do anything more complex, but 100 sec is a very short timeout that I would be careful using, it might be a good fit for your load but if your replicas drift a bit more, then you will need to keep adding barriers to avoid timeouts that can be incredibly difficult to debug and cause a lot of headaches;

Distributed checkpointing with mid-epoch checkpoint support: this is a very cool implementation, it uses distributed checkpointing from PyTorch. They create some wrappers (e.g. for optimizer) where they implement the Stateful protocol to support checkpointing. They also use the StatefulDataLoader from torchdata to do checkpointing of mid-epoch data loader state;

Misc: there are of course other interesting things, but it is cool to mention that they also implemented a no frills LLaMA model without relying on thousands of different libs (it seems it became fashionable nowadays to keep adding dependencies), so kudos for that to keep it simple.

by Christian S. Perone
Machine Learning, Programming

Memory-mapped CPU tensor between Torch, Numpy, Jax and TensorFlow

This is just a fun experiment to answer the question: how can I share a memory-mapped tensor from PyTorch to Numpy, Jax and TensorFlow in CPU without copy and making sure changes done in memory by torch are reflected on all these shared tensors ?

One approach is shown below:

import torch
import tensorflow as tf
import numpy as np
import jax.numpy as jnp
import jax.dlpack

# Create the tensor and persist
t = torch.randn(10, dtype=torch.float32)
t.numpy().tofile("tensor.pt")

# Memory-map the file with PyTorch
t_mapped = torch.from_file("tensor.pt", shared=True, size=10, dtype=torch.float32)

# Memory-map it with numpy, the same tensor
n_mapped = np.memmap("tensor.pt", dtype='float32', mode='r+', shape=(10))

# Convert it to Jax, will reuse the same buffer
j_mapped = jnp.asarray(n_mapped)

# Convert it to dlpack capsule and load in TensorFlow
dlcapsule = jax.dlpack.to_dlpack(j_mapped)
tf_mapped = tf.experimental.dlpack.from_dlpack(dlcapsule)

Now the fun part begins, I will change the tensor in PyTorch and we will check what happens in the Numpy, Jax and TensorFlow tensors:

>>> t_mapped.fill_(42.0) # Changing only PyTorch tensorA
tensor([42., 42., 42., 42., 42., 42., 42., 42., 42., 42.])

>>> n_mapped # Numpy Array
memmap([42., 42., 42., 42., 42., 42., 42., 42., 42., 42.], dtype=float32)

>>> j_mapped # Jax Array
Array([42., 42., 42., 42., 42., 42., 42., 42., 42., 42.], dtype=float32)

>>> tf_mapped # TensorFlow Tensor
<tf.Tensor: shape=(10,), dtype=float32, numpy=array([42., 42., 42., 42., 42., 42., 42., 42., 42., 42.], dtype=float32)>

As you can see from above, changes in the torch tensor reflected back into Numpy, Jax and TensorFlow, that’s the magic of memmap().

by Christian S. Perone
Machine Learning, Philosophy

Generalisation, Kant’s schematism and Borges’ Funes el memorioso – Part I

Introduction

Portrait of Immanuel Kant by Johann Gottlieb Becker, 1768.

One of the most interesting, but also obscure and difficult parts of Kant’s critique is schematism. Every time I reflect on generalisation in Machine Learning and how concepts should be grounded, it always leads to the same central problem of schematism. Friedrich H. Jacobi said that schematism was “the most wonderful and most mysterious of all unfathomable mysteries and wonders …” [1], and Schopenhauer also said that it was “famous for its profound darkness, because nobody has yet been able to make sense of it” [1].

It is very rewarding, however, to realize that it is impossible to read Kant without relating much of his revolutionary philosophy to the difficult problems we are facing (and had always been) in AI, especially regarding generalisation. The first edition of the Critique of Pure Reason (CPR) was published more than 240 years ago, therefore historical context is often required to understand Kant’s writing, and to make things worse there is a lot of debate and lack of consensus among Kant’s scholars, however, even with these difficulties, it is still one of the most relevant and worth reading works of philosophy today.

(more…)

by Christian S. Perone
Machine Learning, Math

Thoughts on Riemannian metrics and its connection with diffusion/score matching [Part I]

Different gaussian curvature surfaces. Image by Nicoguaro.

We are so used to Euclidean geometry that we often overlook the significance of curved geometries and the methods for measuring things that don’t reside on orthonormal bases. Just as understanding physics and the curvature of spacetime requires Riemannian geometry, I believe a profound comprehension of Machine Learning (ML) and data is also not possible without it. There is an increasing body of research that integrates differential geometry into ML. Unfortunately, the term “geometric deep learning” has predominantly become associated with graphs. However, modern geometry offers much more than just graph-related applications in ML.

I was reading the excellent article from Sander Dieleman about different perspectives on diffusion, so I thought it would be cool to try to contribute a bit with a new perspective.

(more…)

by Christian S. Perone
Machine Learning

Large language model data pipelines and Common Crawl (WARC/WAT/WET)

Erik Desmazieres’s “La Bibliothèque de Babel”. 1997.

We have been training language models (LMs) for years, but finding valuable resources about the data pipelines commonly used to build the datasets for training these models is paradoxically challenging. It may be because we often take it for granted that these datasets exist (or at least existed? As replicating them is becoming increasingly difficult). However, one must consider the numerous decisions involved in creating such pipelines, as it can significantly impact the final model’s quality, as seen recently in the struggle of models aiming to replicate LLaMA (LLaMA: Open and Efficient Foundation Language Models). It might be tempting to think that now, with large models that can scale well, data is becoming more critical than modeling, since model architectures are not radically changing much. However, data has always been critical.

This article provides a short introduction to the pipeline used to create the data to train LLaMA, but it allows for many variations and I will add details about other similar pipelines when relevant, such as RefinedWeb (The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only) and The Pile (The Pile: An 800GB Dataset of Diverse Text for Language Modeling). This article is mainly based on the pipeline described in CCNet (CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data) and LLaMA’s paper, both from Meta. CCNet was developed focusing on the data source that is often the largest one, but also the most challenging in terms of quality: Common Crawl.

The big picture

The entire pipeline of CCNet (plus some minor modifications made by LLaMA’s paper) can be seen below. It has the following stages: data source, deduplication, language, filtering, and the “is-reference” filtering which was added in LLaMA. I will go through each one of them in the sections below.

Visual overview of the CCNet pipeline with some modifications done in LLaMA. Click to enlarge.

Let’s dive into it !

(more…)

by Christian S. Perone