Adrian Hill
LightNews LightNews
banner
adrianhill.de
Adrian Hill
@adrianhill.de
700 followers 600 following 110 posts
PhD student at @bifold.berlin, Machine Learning Group, TU Berlin.
Automatic Differentiation, Explainable AI and #JuliaLang.
Open source person: adrianhill.de/projects
Posts Replies Media Videos
adrianhill.de
„Making things easy breaks systems that use difficulty as signaling“ @zey.bsky.social @neuripsconf.bsky.social
December 3, 2025 at 11:10 PM
adrianhill.de
I'm excited to see whether our idea translates to general MC integration over Jacobians and gradients outside of XAI. Please don't hesitate to talk to us if you have ideas for applications!
December 3, 2025 at 7:36 PM
adrianhill.de
Our proposed SmoothDiff method (see first post) offers a bias-variance tradeoff: By neglecting cross-covariances, both sample efficiency and computational speed are improved over naive Monte Carlo integration.
December 3, 2025 at 7:36 PM
adrianhill.de
To reduce the influence of white noise, we want to apply a Gaussian convolution (in feature space) as a low-pass filter.
Unfortunately, this convolution is computationally infeasible in high dimensions. Naive Monte Carlo approximation results in the popular SmoothGrad method.
December 3, 2025 at 7:36 PM
adrianhill.de
Input feature attributions are a popular tool to explain ML models, increasing their trustworthiness. In this work, we are interested in the gradient of a given output w.r.t. its input features.
Unfortunately, gradients of deep NN resemble white noise, rendering them uninformative:
December 3, 2025 at 7:36 PM
adrianhill.de
Join us today from 4:30 to 7:30 PM @neuripsconf.bsky.social Hall C,D,E #1006 for our poster on SmoothDiff, a novel XAI method leveraging automatic differentiation.
🧵1/6
December 3, 2025 at 7:36 PM
adrianhill.de
Working on a poster in Typst again. But this time from scratch without any package dependency. 💪
Proof of concept for a Typst poster with a two-column layout and the entire code that produces it: ```typst // Poster Template #set page("a0", flipped: true, columns: 2, margin: 3cm, background: rect(fill: blue, width: 100%, height: 100%), ) #set columns(gutter: 2cm) #set par(spacing: 2cm) #set text(font: "Helvetica", size: 90pt) #set rect(width: 100%, inset: 1cm, fill: white) // Content #place(top, scope: "parent", float: true, clearance: 2cm, rect(height: 8cm)[Title], ) #rect(height: 1fr)[Panel 1] #rect(height: 1fr)[Panel 2] #rect(height: 1fr)[Panel 3] #colbreak() #rect(height: 3fr)[Panel 4] #rect(height: 2fr)[Panel 5] #place(bottom, scope: "parent", float: true, clearance: 2cm, rect(height: 5cm)[Footer], ) ```
November 3, 2025 at 8:52 PM
adrianhill.de
Doing XAI research in #JuliaLang
A depressed and jaded Sisyphus pushing a boulder with the Python logo up a mild incline. A happy and fulfilled Sisyphus pushing a boulder with the Julia logo up a steep incline.
September 15, 2025 at 8:40 PM
adrianhill.de
🤙🤙
An Ominum Mini Cargo Bike with a sticker-bombed mudguard.
July 31, 2025 at 2:44 PM
adrianhill.de
The perfect use case for UnicodePlots.jl!
github.com/JuliaPlots/U...
A variety of plots drawn with unicode characters, as rendered by the "UnicodePlots.jl" package. Includes line-, scatter–, contour- and box-plots, histograms and heatmaps among others.
July 28, 2025 at 10:29 AM
adrianhill.de
According to the Typst team, PDF submissions should be allowed, I hope this is correct:
info.arxiv.org/help/submit_...
Screenshot from Twitter of the Typst team writing: "arXiv allows PDFs if you did not create them with LaTeX: https://info.arxiv.org/help/submit_pdf.html"
May 14, 2025 at 5:56 PM
adrianhill.de
I had to endure some teasing from my collaborators for choosing to write our NeurIPS submission in @typst.app – looks like I have one more reason to believe I made the right call... 😅
My BlueSky feed; filled with three mentions of Overleaf being down due the upcoming NeurIPS deadline.
May 14, 2025 at 8:48 AM
adrianhill.de
@sirmarcel.bsky.social is still at it
Marcel dipping a t-shirt into a vat of indigo dye.
April 30, 2025 at 7:47 AM
adrianhill.de
Today is a good day, dyed a shirt and made it the front page of HN 🥹
Indigo dyed t-shirt that reads “Off the screen, on the saddle”. Our ICLR blog post on automatic sparse differentiation made it to number 7 on Hacker News.
April 30, 2025 at 7:40 AM
adrianhill.de
If you are at #ICLR2025 and want to chat about automatic sparse differentiation (or just grab a sticker), come see me at poster 471!
A conference poster at ICLR 2025 titled “Fast Jacobians and Hessians by leveraging sparsity”.
April 24, 2025 at 6:55 AM
adrianhill.de
From the Singapore Pavillon at #EXPO2025 to the real thing at #ICLR25!
Image of the Singapore pavilion at the World Expo in Osaka.
April 23, 2025 at 3:08 AM
adrianhill.de
The announcement of the "Julia for Machine Learning" course is my most shared skeet on old-BlueSky. Since then, the course material has been extended to include package development in #JuliaLang!

📖: github.com/adrhill/juli...
April 14, 2025 at 11:44 AM
adrianhill.de
Chaotic good: Buying Go books for their graphic design
April 12, 2025 at 3:36 PM
adrianhill.de
I previously built a more minimalist, low-profile Bluetooth keyboard, but it ended up being too difficult to switch to when I needed to get work done. So this one is pretty much the opposite. #ergomech #splitkeyboards
A second, low-profile split ergonomic keyboard in front of the 3D printed one.
February 13, 2025 at 1:10 PM
adrianhill.de
BTS of the soldering process:
Soldering 72 tiny PCBs to build the keyboard matrix.
February 13, 2025 at 1:10 PM
adrianhill.de
Completed my post preprint project: a split keyboard to help with my wrist pain (pictured with the pain-relieving gel it is trying to replace).
Two halves of a 3D printed ergonomic split keyboard surrounding pain relief gel.
February 13, 2025 at 1:10 PM
adrianhill.de
We demonstrate that even for one-off computations, end-to-end ASD (yellow line) outperforms AD! When the sparsity pattern can be reused, the performance improvements are even more significant (green line).
Performance comparison of AD and ASD Jacobian computations on the Brusselator PDE.
January 30, 2025 at 2:32 PM
adrianhill.de
The bottleneck of previous ASD implementations is the detection of sparsity patterns. We describe a faster implementation of pattern detection based on operator overloading which is compatible with generic code bases.
A part of the derivation of the second-order propagation rules.
January 30, 2025 at 2:32 PM
adrianhill.de
However, many of these matrices exhibit sparsity, a property that can be leveraged to speed up AD by orders of magnitude. We call this automatic sparse differentiation (ASD). While known in the AD community, it remains largely ignored in the ML literature.
Jacobian sparsity pattern of a discretized Brusselator PDE, with colors resulting from greedy column coloring.
January 30, 2025 at 2:32 PM
adrianhill.de
You think Jacobian and Hessian matrices are prohibitively expensive to compute on your problem? Our latest preprint with @gdalle.bsky.social might change your mind!
arxiv.org/abs/2501.17737
🧵1/8
Figure comparing automatic differentiation (AD) and automatic sparse differentiation (ASD). (a) Given a function f, AD backends return a function computing vector-Jacobian products (VJPs). (b) Standard AD computes Jacobians row-by-row by evaluating VJPs with all standard basis vectors. (c) ASD reduces the number of VJP evaluations by first detecting a sparsity pattern of non-zero values, coloring orthogonal rows in the pattern and simultaneously evaluating VJPs of orthogonal rows. The concepts shown in this figure directly translate to forward-mode, which computes Jacobians column-by-column instead of row-by-row.
January 30, 2025 at 2:32 PM