Newest Questions
12,769 questions
0
votes
0
answers
14
views
How to train a neural network with a non-differentiable linear program in the loss?
I am currently trying to train a neural network to predict certain input parameters for a linear program based on an input dataset of system measurements related to these parameters. The workflow is ...
0
votes
0
answers
4
views
Python GraphRag retrieval: Is using the embedded user's question when querying the data from graph good or bad?
I'm following a tutorial about using GraphRAG Python library to answer user's questions about actors who acted in the movies that match some specific description. The approach is to embed the user's ...
- 101
0
votes
0
answers
10
views
why aren't my Training and validation losses do not reduce when fine-tuning ViTPose from huggingface
I am trying to fine-tune a transformer/encoder based pose estimation model available here at: https://huggingface.co/docs/transformers/en/model_doc/vitpose
When passing "labels" attribute to ...
- 101
0
votes
0
answers
5
views
Synthetic to real image conversion for Pose Estimation
i have synthetic images of poses and that data is being used for trainig a pose estimatioon model, what i want is that i want to convert it to real images, meanig that the people appear real in it, i ...
0
votes
0
answers
14
views
MATLAB implementation of Manifold-PBO algorithm producing different Monte Carlo simulation convergence results than paper
I'm trying to replicate the results from a paper that uses the Manifold-PBO algorithm in a Monte Carlo simulation, but my implementation is producing significantly different convergence curves ...
0
votes
0
answers
4
views
How to use my neo4j KG in Dify?
I am trying to build a RAG with Dify, and as it is expected to handle hopping questions.
As a POC, I try to use Illama to parse my documents and build a knowledge graph (KG) in my local neo4j ...
- 101
0
votes
0
answers
8
views
Machine Learning model best design setup
I'm looking for the best design based on the following requirements. I'd appreciate an explanation of a possible high level route I can follow.
I need to deploy an Image classifier that takes in an ...
- 101
0
votes
0
answers
9
views
Why is DecisionTree using same feature and same condition twice
When trying to fit scikit-learn DecisionTreeClassifier on my data, I am observing some weird behavior.
x[54] (a boolan feature) ...
- 101
0
votes
0
answers
23
views
Model or LLM that is fast enough to describe an image in detail
The heading might be little weird, but let's get on the point.
I made an chat-bot like application where user can upload video and cant chat/ask anything about the video content, just like we talk to ...
0
votes
0
answers
9
views
Forward chaining proves D but backward chaining does not – is my solution correct?
I came across this logic exercise, and I would like to confirm if my solutions are valid.
Here is the setup:
Fact base: {A}
Rule base:
R1: A → B
R2: B → C
R3: E → D
Goal:
Add one rule such that D can ...
0
votes
0
answers
18
views
Best neural network algorithms/architectures for generating synthetic sequences of tuples of words
I would like to generate sequences of tuples using a neural network algorithm such that the model trains on a dataset of sequences of tuples and generates synthetic sequences of tuples. Each tuple <...
- 101
0
votes
2
answers
61
views
Why can the function that turns the history into one Markov state be any function?
Summary
In David Silver's RL lecture slides, he defines the State $S_t$ formally as a function of the history:
David then goes on to define the Markov state as any state $S_t$ such that the ...
- 1
0
votes
0
answers
8
views
How can I integrate LangSmith for observability in a multi-agent Autogen (AG2) GroupChat setup?
I'm working on a document analysis service using Autogen (AG2). The service has two main agents: a reader and an analyzer. The reader splits the document into chunks and sends them to the analyzer, ...
-1
votes
1
answer
33
views
How to make a DQN that has combinable discrete actions?
I have a DQN network with two possible binary actions, but I want the DQN to be able choose all the possible action combinations give a state. So instead of (1,0) (0,1) I want (0,0) (1,0) (0,1) (1,1).
...
0
votes
0
answers
6
views
Seeking advice for MLP architectures and training strategies for texture reconstruction (mipmap upscaling) focused on residual compression
I am working on a project involving texture compression and reconstruction, specifically targeting mipmap upscaling for game development assets. I'm encountering challenges in achieving the desired ...
0
votes
1
answer
25
views
Disparity between training and evaluation episodes
I am new to reinforcement learning topic. I tried to re-train DDPG control this DDPG control example using exact same configuration (except the max training episode set to 500) and this is the result ...
- 101
1
vote
0
answers
20
views
Why does MARL require full history while single-agent RL (Sutton \& Barto) uses state-based returns?
In single-agent RL (as in Sutton & Barto's framework), the expected return is defined from the current state $s_t$:
$$
G_t = R_t + R_{t+1} + ....
$$
In the Bellman equation, $V(s_t)$ depends only ...
1
vote
0
answers
20
views
Clear instructions on how to convert a model from huggingface to gguf format. The internet will love it!
Since many hours Im not able to convert the pytorch_model-00001-of-00003.bin to GGUF.
I want to convert this LLM https://huggingface.co/LeoLM/leo-hessianai-13b-chat/tree/main
Please help me. All the ...
2
votes
1
answer
36
views
Complete formula to get LLM VRAM usage
I would like to find the GPU size required to run an hypothetical LLM, considering all possible factors, like:
P: Model parameters (total or MoE active parameters)
Q: Quantization bits
C: Context ...
- 121
0
votes
0
answers
9
views
Trained ML model is slowing down python code execution
I’m running a real-time inference loop in Python 3.12.3 with a pre-trained variational autoencoder (VAE) implemented in Tensorflow. Although I load the model once outside my main loop, each call to ...
- 1
0
votes
1
answer
28
views
Almost no fps improvement comparing sbx(sb3+jax) ppo with sb3 ppo
Backgroud
sbx is a jax implementation of stable-baselines3. As claimed here, it can accelate RL training by jit compared to sb3+pytorch.
Question
I tested sbx ppo and sb3 ppo on gym env ...
- 53
2
votes
0
answers
32
views
Fine-tuning ResNet101 stuck at ~50% accuracy while MobileNetV2 reaches ~90% (same data, head, training setup)
I'm fine-tuning two different CNNs for an image classification task:
The first CNN uses a ResNet101 backbone, and the second uses a MobileNetV2 backbone. Both are pre-trained on ImageNet.
I use the ...
- 41
0
votes
0
answers
11
views
Instance Segmentation Performance Issue with Custom Frost-Damaged Grape Dataset (YOLOv11)
I am working on a project involving frost damage detection in grapes in Türkiye. I collected a custom dataset consisting of 972 images. I labeled the data very precisely in Roboflow for classification,...
1
vote
0
answers
10
views
Mathematical definitions of "Box" state and action spaces in Reinforcement Learning
I've made a custom RL environment using the OpenAI Gymnasium library, where I used the gymnasium.spaces.Box class for defining the state and action spaces. Here are ...
1
vote
0
answers
9
views
Classification for the detection of vulnerable codes with poor performance
I'm currently using codeBert-small-v1 to classify code according to CWE, which works perfectly: the model manages to identify the CWE of the code submitted to it (a total of 4 classes with an accuracy ...
- 11
2
votes
1
answer
91
views
Proposal for AGI model
I've been doing a bit of research into formal models for AGI, searching for fertile ground for developing new ideas. One area that didn't seem too thoroughly explored was in designing agents that “...
- 121
1
vote
0
answers
9
views
TDC Fails to Converge in Off-Policy MountainCar-v0 with Tile Coding
I am trying to perform off-policy state-value function evaluation with linear function approximation in the Gymnasium "MountainCar-v0" environment. I set the discount factor $\gamma=0.99$, ...
1
vote
0
answers
10
views
How would PCA change if we center data using the coordinate-wise median instead of the mean?
In standard Principal Component Analysis (PCA), we center the data by subtracting the mean of each coordinate before computing the covariance matrix and solving for the principal components.
I am ...
- 11
0
votes
0
answers
13
views
Convolutional Kernels in CNN learning to find different patterns
Suppose we have an input image of dimensions $w \times h $ and the first hidden layer has dimension $(w-1) \times (h-1) \times 3$. We have $3$ seperate $3 \times 3$ kernels with no padding.
I ...
- 1
3
votes
2
answers
54
views
Understanding Why TD Learning Has Lower Variance Despite Using an Estimated Value
In Temporal Difference (TD) learning, the value function is updated using its own estimate, following the rule:$V (S_t) \leftarrow V (S_t) + \alpha[R_{t+1} + \gamma V (S_{t+1}) − V (S_t)]$. It's often ...
- 51
1
vote
1
answer
46
views
Do neural networks do wishful thinking?
I will give an example of wishful thinking. When you try to prove a theorem you think what would imply that theorem and maybe try to find a lemma that implies it. Maybe neurons try to connect previous ...
- 21
0
votes
0
answers
25
views
Is it feasible and scalable to combine self-replicating automata (after von Neumann) with federated learning and the social web?
Von Neumann’s idea of self-replicating automata describes machines that can reproduce themselves given a blueprint and a suitable environment. I’m exploring a concept that tries to apply this idea to ...
0
votes
0
answers
11
views
How do BPE Tokenizers with `add_prefix_space` handle natural language, such as quotations or poetry, where there won't be a prefix space?
BPE Tokenizers are the standard for modern LLMs. By default, most add_prefix_space, so that John went away is pretokenized to <...
- 63
-1
votes
0
answers
19
views
can you recommend an image filename(caption?) generating ai?
2 to 6 words output and filenames should be unique and relevant to the image content.
0
votes
1
answer
37
views
Who argued that we're entering a 4th era of science with machine learning?
I remember reading a reference to a recent paper that argued that science today is in its 4th stage (paradigm?), the era of modelling with machine learning. The 3rd was that of Newton, Kepler, et al.
...
- 525
0
votes
0
answers
10
views
Is there other ways than using negative log-likelihood or KL-divergence to compute a loss function?
I've read that the two common ways to express a loss function in ML problems was to start either from the likelihood, then use the negative log likelihood to find a good expression of the loss, or to ...
0
votes
0
answers
17
views
How can AI be leveraged to improve fraud detection in real-time financial transactions?
I'm exploring how AI can help financial institutions enhance their fraud detection systems, particularly in real-time transactions. What are the best approaches, tools, or models for building more ...
0
votes
1
answer
52
views
How do tools like V0.dev provide the project's code context to the AI while minimizing its input tokens?
Under the hood, tools like V0.dev use OpenAI API to generate code. How does V0 know which file context to send to the AI when user makes prompts like "make xyz change/change the button color/add ...
- 9
0
votes
0
answers
30
views
Why would the tokenizer for encoder-decoder model for machine translation use bos_token_id == eos_token_id? How does it know when a sequence ends?
I see on this PyTorch model Helsinki-NLP/opus-mt-fr-en (HuggingFace), which is an encoder-decoder model for machine translation:
...
- 3,353
0
votes
0
answers
15
views
Torch gradient estimates disagreeing with analytic and perturbation approximated gradients
I'm faced with a problem where as the title says I'm having trouble with the torch package's built in automatic differentiation algorithms (or my usage?). I think it was meant to be used on mini-...
-1
votes
0
answers
42
views
What are some notable connections between quantum mechanics and machine learning?
I have recently started to learn about machine learning, and I feel like there is a lot of connection to quantum mechanics in it. For example, this idea of the predictin through layers of neurons look ...
0
votes
1
answer
25
views
Different approaches to probability, connection to measure theory and machine learning
I am trying to understand the probability as used in machine learning. So far I understand there is multipile approaches to probability. The two I know of are bayseian and frequntist approach. So far ...
0
votes
2
answers
37
views
Can you train YOLOv11 using the Coral USB Accelerator?
I have a dataset that I want to train a YOLOv11 model on. The problem is, my laptop is too slow and using Google Colab is hard due to the inactivity limit and the training interrupting.
My idea was to ...
0
votes
0
answers
29
views
How to cluster data which has unique distribution for all classes?
I wanted to use MLP to classify my data. But it didn't show good result. When I saw its scatterplot I realized that my data has unique distribution for all classes which makes learning impossible, ...
2
votes
1
answer
43
views
Can the output of a language model be identical to it's training data if finetuned with reference documents also present on the training data?
E.g. Finetuning a language model using text from Wikipedia articles (without modifications) when the language model has Wikipedia data in its training dataset will cause the model to reproduce the ...
- 377
0
votes
0
answers
14
views
PPO: How to exploit action equivalences in continuous approximation of large discrete (and constrained) action spaces
I face a reinforcement learning problem where the action space is large and constrained (integer points in an n-dimensional polyhedron that depends on the state).
To train the RL agent (PPO) I make ...
0
votes
3
answers
76
views
What are some practical use cases where generative AI has saved you time or boosted creativity?
I’ve been testing out different generative AI tools recently, and I’m wondering what kinds of real, everyday use cases people here have found most useful. Not just flashy demos — I mean the tools that ...
0
votes
0
answers
17
views
Intuition behind Load-Balancing Loss in the paper OUTRAGEOUSLY LARGE NEURAL NETWORKS: THE SPARSELY-GATED MIXTURE-OF-EXPERTS LAYER
I'm trying to implement the paper "OUTRAGEOUSLY LARGE NEURAL NETWORKS: THE SPARSELY-GATED MIXTURE-OF-EXPERTS LAYER"
But got stuck while implementing the Load-Balancing Loss.
Could someone ...
- 1
2
votes
1
answer
669
views
How can the exact same model give different confusion matrices for the test dataset and the entire dataset?
I have recently implemented a simple artificial neural network with 1 hidden layer. I split my data using train_test_split and I end up with the following confusion matrix in my test set.
...
- 21
2
votes
1
answer
37
views
Can Self Attention capture rate of change of token?
From what I understand, the self-attention mechanism captures the dependency of a given token on various other tokens in a sequence. Inspired by nature, where natural laws are often expressed in terms ...
