Facebook AI's Flashlight Fast and flexible ML Library.
                     FACEBOOK AI'S 
FLASHLIGHT FAST AND 
FLEXIBLE ML LIBRARY 
IN C++.
Introduction
• Flashlight is a new open source 
machine learning (ML) library 
developed by FAIR to allow teams to 
quickly and easily change deep and ML 
frameworks to better suit their needs. 
It is written entirely in C++.
• Flashlight would make it easier for the 
AI community to tinker with the low-
level code that underpins deep and ML 
frameworks, allowing them to better Source: Twitter
exploit the hardware available and 
push performance limits. 
2
Purpose
• Deep and machine learning systems 
are excellent at what they do, but 
changing their internals has proven 
challenging in the past.
• Since low-level internals may be 
accidentally obfuscated, closed-
source, or hand-tuned for specific 
purposes, finding the correct code to 
modify is time-consuming and error-
prone.
• Recompiling the framework after Source: Facebook Engineering
you've made improvements is both 
time and compute-intensive. 
3
Why C++
• Since modern C++ is a powerful tool 
for doing research in high-performance 
computing environments, we built 
Flashlight from the ground up in that 
language.
• Since modern C++ allows for first-
class parallelism and out-of-the-box 
speed, Flashlight has a very low 
system overhead.
• Flashlight also includes easy-to-use 
interfaces for integrating code from Source: Medium
low-level domain-specific languages 
and libraries. 
4
What is does?
• Flashlight is designed on top of a shallow 
stack of modular and easy-to-use 
abstractions.
• We began with the ArrayFire tensor library, 
which supports dynamic tensor shapes and 
types without the use of compile-time 
specifications or C++ templates.
• With an effective just-in-time compiler, 
ArrayFire also optimizes operations on the Source: ArrayFire
fly. 
5
Deep Learning
• Although standard deep learning primitives 
are implemented using well-optimized kernels 
from hardware-specific vendor libraries, 
writing custom high-performance code can be 
difficult to quickly integrate and iterate on.
• Building new low-level computational 
abstractions is simple with Flashlight.
• With minimal effort, you can cleanly integrate 
CUDA or OpenCL kernels, Halide AOT 
pipelines, or other custom C/C++ code. Source: Levitty
6
Why it Matters?
• As deep and machine learning models 
become more complex, progress is 
dependent on optimization.
• Advanced AI research and models 
necessitate high-performance code that 
makes the most of available hardware.
• Flashlight's fast rebuilds make it easier to 
do research on the library itself, which can 
then be extended to other frameworks 
later.
• Flashlight opens the door to research that 
stretches the boundaries of success by 
making it easier to rapidly iterate on 
custom low-level code. Source: Altabel
7
Why Machine Learning Libraries
• Our machine learning group is fortunate 
in that many solutions are already 
available in the form of libraries and 
packages.
• These machine learning libraries are fast 
and effective, and they've been 
thoroughly tested for a variety of 
scenarios.
• Our learning is fueled by these libraries, 
which make writing code, whether in C+ Source: Medium
+ or Python, much easier and intuitive. 
8
SHARK C++ Library
• Shark is a lightweight modular library 
with extensive support for supervised 
learning algorithms including linear 
regression, neural networks, 
clustering, and k-means.
• It also contains linear algebra and 
numerical optimization capabilities.
• There are important mathematical 
functions or areas to know when 
Source: Medium
doing machine learning tasks. 
9
MLPACK C++
• mlpack is a C++ machine learning 
library that is fast and versatile.
• Its aim is to make cutting-edge machine 
learning algorithms available in a quick 
and extensible manner.
• These algorithms are provided by 
mlpack as simple command-line 
programs, Python bindings, Julia 
bindings, and C++ classes, which can 
then be used in larger-scale machine 
learning solutions. Source: TechLeer
10
K-Means Application
• K-means is essentially an iterative 
method for segmenting data into 
different clusters. 
• We begin by assigning some initial 
centroids, which are completely 
random. 
• After that, we find the nearest centroid 
for each data point. The data point 
would then be assigned to that 
centroid. 
• As a result, each centroid represents a 
distinct class. We'll compute the mean 
of those centroids once we've assigned Source: Towards Data Science
all of the data points to each centroid. 
11