Questions tagged [machine-learning]
Questions about computer algorithms that automatically discover patterns in data and make good decisions based on them.
1,245 questions
-3 votes
0 answers
31 views
Can We Design Incentive-Compatible LLM Judge Certification Beyond Shared Information Invariants? [closed]
The TVD-MI (Total Variation Distance–Mutual Information) framework provides a reference-free method for certifying LLM judges—models that evaluate outputs like code correctness or theorem proofs ...
- 3,124
1 vote
2 answers
229 views
Why move memory from LLM to MCP?
I’ve been reading about the Model Context Protocol (MCP) and how it lets LLMs interact with tools like email, file systems, and APIs. One thing I don’t fully get is the idea of moving “memory” from ...
- 11
0 votes
0 answers
37 views
Exact learning model definition
The following is Definition 17 of Kearns and Vazirani's "Introduction to computational learning theory". We say that the representation class $C$ is efficiently exactly learnable from ...
- 2,252
1 vote
0 answers
32 views
Unsupervised Anomaly Detection in Time Series
I have been assigned the following task, I'll try to summarize: TASK: Ubnsupervised Anomaly Detection. INPUT: A single time series, made of 135000 frames, approximately 45000 points. The time series ...
- 11
0 votes
0 answers
26 views
The transitivite property of the DB-SCAN algorithm
I was looking a little closer at the dbscan explanation on wiki (and of course all the medium articles that just copy/paste wiki), and I'm not sure I'm convinced it does what it claims to do on the ...
1 vote
1 answer
104 views
Query regarding random seeds
I am very new to statistics and bioinformatics. For my project, I have been creating a certain number of sets of n patients and splitting them into subsets, say HA and HB, each containing equal number ...
2 votes
0 answers
60 views
The Difference Between a Class and a Hypothesis
I am reading Introduction to Machine Learning by Ethem Alpaydin, and in chapter 2, he mentions 3 things: the class C, the hypothesis class H, and the hypothesis h, which belongs to H. I'll summarize ...
- 21
0 votes
1 answer
65 views
Inpainting of complex images
In want to create an algorithm which can complete the missing part of an image, I gather this can be done with machine learning algorithms. I know nothing about these algorithms but I am willing to ...
- 11
-1 votes
1 answer
92 views
In LLM, the number of weights vs the precision of the wights, which is more important?
I'm considering the edge compute case with LLMs. Recently, I become interested in the quantized MOE models, because the active number of parameters is only a small portion of the total number of ...
1 vote
1 answer
70 views
Large-Scale Image Near-Duplicate Detection for Real Estate Dataset
I want to perform large-scale image similarities detection. For context, I have a large database containing almost 13,000,000 flats. Every time a new flat is added to the database, I need to check ...
- 11
0 votes
0 answers
22 views
How to set up loss function and de-normalized accuracy calculation for training with normalised data
I am training a deep learning regression model where I want to calculate 2-fold accuracy. I am normalising the input data to mean $0$, standard deviation $1$ before passing to the model. Earlier I was ...
1 vote
0 answers
45 views
Mathematically deriving the ultimate activation function
I have been working on finding the ultimate activation function here are the properties I would suspect the function to have: Positive lipshitz continuous, i.e $0 \leq f'(x) \leq 1 \; \forall x\in \...
- 11
1 vote
0 answers
39 views
Error in formula for choosing noise multiplier $\sigma$ in DP-SGD based on Gaussian Mechanism?
While reviewing the literature on differential privacy and its addition to ML, I noticed that I do not end up with the same formula for the noise multiplier $\sigma$ for the differentially-private SGD ...
0 votes
0 answers
64 views
Can prompt injection be used to circumvent the intended use of LLM's?
Let us say that in the not-so-distant future, people might have to convince an LLM that they are worthy of a job. Is it possible to use prompt injection to convince the LLM that you are worthy of a ...
- 165
0 votes
1 answer
50 views
Understanding Clean Chains in Blame Graph Described in the Predictive Marker Algorithm (Lykouris & Vassilvitskii)
Can someone explain to me how clean chains work in this paper? Here's the paper: Competitive caching with machine learned advice by Lykouris & Vassilvitskii I'm trying to implement the predictive ...
