Mahdi Haghifam
@HaghifamMahdi
Followers
1K
Following
10K
Media
20
Statuses
304
Researcher in ML and Privacy. Distinguished Postdoc at @KhouryCollege. PhD @UofT & @VectorInst. Prev. Research Intern @GoogleDeepMind,@ServiceNowRSRCH
Boston
Joined April 2020
RT @random_walker: I tell students on the first day of class that if you're truly learning it's supposed to feel uncomfortable. The reason….
0
156
0
RT @bremen79: New version of my book on Online Learning!. This is a BIG update, with 43 pages of additional content….
arxiv.org
In this monograph, I introduce the basic concepts of Online Learning through a modern view of Online Convex Optimization. Here, online learning refers to the framework of regret minimization under...
0
100
0
RT @thegautamkamath: Want to see Boaz give a talk on this exciting work in beautiful Milan? . Be sure to register for ALT 2025, where he's….
0
4
0
I’ll be at #NeurIPS2024 this week! Looking forward to presenting my joint work with Thomas Steinke (@shortstein ) and Jon Ullman (@thejonullman). NeurIPS page with video: Link to arxiv:
1
3
35
RT @thejonullman: Join us for Charles River Privacy Day on Nov 22nd!. Featuring talks by @thegautamkamath, @jalajupadhyay, Virginia Smith,….
0
10
0
If you're applying, take advantage of this program. Also, if you're familiar with the process, consider volunteering.
Want feedback on your @UofTCompSci grad school application? . GAAP connects prospective MSc/PhD students with current grad students for application feedback: Please share with those who would benefit/consider volunteering!.
0
0
3
For many undergrads paying application fee is not easy. I really like this suggestion by @SoloGen.
@HaghifamMahdi @newplatonism That's an important factor. I used to have a statement on my website that if for financial reasons you need to know my current opinion about your chance, you can explicitly ask and I do my best to tell you. Some did, and I answered them.
0
0
5
RT @thejonullman: New faculty have been asking me for tips on starting out with PhD students, and I thought I'd share some of them here. B….
0
27
0
RT @FlavioCalmon: Mario was incredibly passionate about math, information theory, and statistics. He was homeschooling his son so he could….
gofundme.com
We lost Dr. Mario Diaz Torres on Saturday August 31, 2024 in Mexico City where Mario an a… James Melbourne needs your support for For the family of Mario Diaz
0
11
0
In particular, we show that our algorithm has an adaptive excess error guarantee given by
1
0
3
We develop private algorithms where excess error scales with the “effective” diameter of the dataset. We define effective diameter using the “quantile radius” which is the radius around the “true” geometric median that contains most data points.
1
0
4
The geometric median of a set of points {x_i} minimizes the sum of L_2 norms to x_i. Given ||x_i|| < R, the excess error of DP-SGD for this problem is sqrt(dimension)*R/epsilon. But what if most of the data points are “close” to each other? Can we do better?
1
0
3
The excess error of DP-SGD in convex optimization scales linearly with the feasible set’s diameter, which is optimal for some losses. In this work, we develop private gradient-based algorithms that reduce the privacy cost to "log(diameter)" for the geometric median task! #NeurIPS
5
6
82
RT @gavinrbrown1: The movie sells this scene as Just Memorization vs Real Intelligence, but it actually shows how cool it is to have memori….
0
4
0
RT @kasperglarsen: Did you also think Empirical Risk Minimization was optimal for agnostic PAC learning? Turns out it is not - but our new….
0
13
0
Happy that our paper with Idan, Karolina, Roi and Dan is selected as a best paper. Congratulations to other winners as well🎉🎉🎉.
14
0
92